Audiogravity
All documents2

Changelog — Audiogravity

On this page87

All notable changes to Audiogravity (core, ui, license server, installers and this landing) are documented here. Format based on Keep a Changelog; the project follows Semantic Versioning.


[Unreleased]

Added

  • [site][ui][lic] The editions, the EULA, the third-party notices, the release notes and the changelog open on audiogravity.app. Each of these Markdown documents, at the root of the site's repository, is rendered by the manual's generator into a page of the site — docs/legal/ for the editions, the EULA and the third-party notices, docs/releases/ for the release notes and the changelog —, with its parts in the outline. The landing's footer, announce bar and version, the app's Proprietary License link and sign-in page, the license server's release-notes address and the manual's front page opened GitHub's view of the file; they now open these pages. The Manual workflow regenerates a page when its document changes.
  • [site] The landing quotes its first press. Under the hero's four figures, one line quotes the opinion sempre-audio.at, an Austrian hi-fi magazine, gave of Audiogravity 1.0 on 2026-11-10, translated from German and saying so, and links the article. The magazine ran the press release and added that opinion; nobody there listened to it, so the line never calls it a review, and a test holds it to that.
  • [site] A blog post's pictures open full screen, and the press release's post shows its nine screenshots. A picture of a post opens over the page, as large as the screen allows, with its caption and its place among the post's pictures; its arrows, the keyboard's or a swipe go from one to the next, and Escape, the cross or a touch beside the picture close it. A browser too old for the window opens the picture alone, at full size. The press release's post shows the nine screenshots of the app that come with it, in their order and with their captions, where it offered them in a ZIP only; they open in the same window, a series of their own numbered 1 to 9. The manual and the guides keep their pictures in the page for now.
  • [site] A blog post on the Naim NAC 52. Naim's top pre-amplifier from 1990 to 2002: its sound at the heart of a Naim system, from a French review, translated and attributed, and its standing in Hi-Fi World, Hi-Fi News and Stereophile; how it is built; its power supplies, the 52 PS then the SuperCap; what followed it, from the NAC 82 to the NAC 252 and NAC 552, by Naim's product history; what to pair it with. The NAC 82's photograph shows a NAC 82, as published by Hi-Fi News.
  • [core][ui][manual] Qobuz's editorial playlists by category and by genre, with their own artwork. Playlists › Editorial showed Qobuz's 6,716 selections in one pile. A strip now offers the categories Qobuz files them under, as its own app does — Hi-Res, New Releases, Focus, Moods, Artists, Listening with…, Label Stories, Qobuz Digs, Events & Media, Hi-Fi Audio Partners, Test your speakers, Essential Discography, Top playlists —, with All first; a second strip offers the thirteen top-level genres, with All genres first; the two combine, and the heading names both — Hi-Res · Jazz holds 89. The categories come from Qobuz, in its order and under its English names, so one it adds appears once the core restarts; the genres are those of the Genres shelf, without their sub-genres, which almost never hold a selection (11 of 16 sampled held none). Each selection shows the landscape artwork Qobuz made for it, drawn 2:1 like HIGHRESAUDIO's, where the grid showed the cover of the playlist's first album; the account's own playlists keep that square cover. New: GET /library/qobuz-playlist-tags; GET /library/qobuz-playlists takes tag and genre. The manual describes the two strips.

Changed

  • [ops] A release keeps the footer's version and the documents' pages in step. changelog_rollover.py finds the footer's version by the changelog page's link as well as by the former GitHub one — anchored on that one alone, the version would have stopped moving without a word —, and release.sh prepare commits docs/legal and docs/releases with the regenerated manual.
  • [site] The EULA's definitions are a list. The five definitions ran together as one paragraph; the words are unchanged.
  • [site] The blog's Quad 33 and 303 post opens on the sound, and adds Haute Fidélité's review. Haute Fidélité (issue 275, June–July 2025) compared a restored pair of 1973 with the pair of 2024 on Rogers LS3/5a, and gave the new one its Meilleur achat: its side-by-side account, its hearing of the new pair and of its phono stage, the LS3/5a as a partner, the design team, the technical details and the French price join the post, attributed and translated. How the pairs sound now comes first, history after. The team replaces "engineers led by Peter Comeau", which no readable source supported: Quad's own history gives Jan Ertner as its electronics designer and Peter Comeau as head of acoustic design.
  • [site] The blog's Quad 33 and 303 post adds the black pair of 2026. The 90 Collection, announced on 13 November for Quad's 90th year, dresses the 33 and the 303 in satin black, with green lights where the grey pair has orange: the same circuits and the same price, £1,199 each, by The Ear, Hifi Pig and StereoIndex, from late November 2026, and not a limited edition. Two of Quad's images of the black pair illustrate it, and the post notes that every review it quotes heard the grey one.

[1.0.1] - 2026-11-08

Added

  • [core][ui][manual] The library's Refresh says when it is over, and whether the music folder had changed. Once a Refresh of the local library has read the music folder and the list is back, a message says so: Library refreshed, or Library up to date when the music folder has not changed since it was last read. The turning icon was easy to miss on a phone, and nothing said the end. What the box compares is the player's own database stamp, before and after the reading: measured on both boxes, it moves when an album is added or removed or a file modified, and never when nothing changed. A refusal keeps its own message, a failed reading or reload shows in the list, and no message comes for a Refresh left for another source or from an artist's page that was left, nor for the streaming services, whose Refresh reloads in a second. The API's POST and GET /library/rescan gain changed. The manual says what each message means.
  • [core][ui][site][manual] A USB drive plugged in or pulled out is told, as a notification and in the app. A drive plugged into the box went unannounced, and nothing mounted the music library's drive until something read its folder: on the x86 test box, a minute and a half later, by a shell. The box now hears the drive arrive and says what became of it. The library's drive is mounted on the spot, and the message says it is ready — your music library is available, or the library is elsewhere and the drive can be chosen in Config → MPD → Guided —, or that it could not be mounted, with the mount's own reason. Another drive is said to be not mounted, with how to make it the library, or that it cannot be used, its format named, when the box does not mount it — only ext4 and exFAT are. A drive pulled out is told too — for the library's drive while MPD reads it, that the music is unavailable until it is back. The same words go to the devices subscribed to notifications and, as a toast, to every app open; the box listens only while one of them can be told, and at rest that costs nothing — no signal in 90 s, measured. A drive pulled out while in use leaves a mount of a device that is gone, which kept the drive plugged back in from mounting (measured on the x86 test box): it is taken away first, once nothing holds it, and the message says when something still does. The library is not read again when its drive comes back: it is read when it is chosen, or on a Refresh. The dashboard stream gains a usb_drive event; the manual's notifications, EDITIONS.md and the landing list it.
  • [site] The landing shows the VU meters, and the player on a phone held sideways. A screen joins the hero's carousel after the radio: the player with its two VU meters in place of the cover, with its caption and three labels — a needle per channel, the peak lamp, the Cover | VU meter switch. Under the VU meters' card, a full-width card, Held sideways, shows the player on a phone turned on its side, the same with its meters, and the showcase; three side by side, one under the other below 900 px. Each capture follows the page's theme, as the others do.
  • [site] The press release of Audiogravity 1.0, in English and in French. Each is a PDF of four pages — the release, a fact sheet, nine screenshots —, beside a ZIP of the nine screenshots at full size, all three in press/. A blog post presents them, Audiogravity 1.0: the press release, and the blog's front page now says it carries news of Audiogravity too. The landing's footer links the two PDFs on one line, Press release EN / FR.
  • [manual] What to do when the page does not open, and how to find the box's name again. Troubleshooting gains The page does not open — "refused to connect": what Chrome, Safari and Firefox say, told apart from took too long to respond; three commands typed on the box — ip route get 1.1.1.1 for its address, after src; pgrep -a avahi for its name; pgrep -af audiogravity-ui for how the interface is served, its port right after --port and --ssl for HTTPS —, the address to open in each case, sudo systemctl restart ag-ui-server for an interface that is stopped and sudo systemctl status ag-ui-server for why it stays so. Installation gains Find your box's name, for whoever no longer has the installer's last screen: the router's list of devices, the app's System tab and its support report, pgrep -a avahi on the box, and the names a box nobody renamed answers to, raspberrypi.local and dietpi.local; every page that writes <name>.local links to it.
  • [core][ui][manual] A track of the queue moves by its handle, and plays at a tap. Each row of the box's own queue has a handle at its left: held and slid up or down, the track follows the finger, the rows it passes slide aside, and the list scrolls by itself while the finger stays near the top or the bottom of what shows of it — above the mini-player on a phone. Dropped at the top of the list, the track plays next; the music goes on through a move, the track playing included (measured on a private MPD 0.24.5). The track takes its place at once, before the box answers, and a refused move puts it back with the box's reason. The box is told the track it lands beside, not a place, and finds that track when it moves the other: a queue changed meanwhile — a removal on its way, another device's edit — still gets it right where a place counted on the list shown sent it one further (measured). In random order the handles are greyed: the order decides nothing. A tap on a track's row — anywhere but on its handle, its arrow and its padlock, with the mouse too — plays it, the tracks above it skipped; a track a box without Pro skips leads to the licence panel, as its padlock does. While the music goes out through HQPlayer or a network speaker, which do not play the box's own queue, the tap is refused and the message names where the music goes. A guest account can play a track, as it can skip to the next one; it cannot move one. A Roon queue, changed from the Roon app, has neither. New: POST /library/queue/{queue_id}/move with { "before": id } or { "after": id }, and POST /library/queue/{queue_id}/play.
  • [site] Christmas — a cartoon scene in the snow, every 24 and 25 December. By the visitor's own clock — or any day with ?christmas in the address — the landing dresses up: at the foot of the phone a decorated tree whose bulbs blink, two elves, a snowman, a penguin and a polar bear stand in a bank of snow; snow falls around the phone, baubles swing on red ribbons, holly sits in two corners, stars twinkle on the dark page, and Santa's sleigh crosses the band above the first line now and then. Down the page, snow falls in the margins beside the text, a bauble hangs from each section's top and a figure stands at its foot on a mound of snow — on a phone, a small bauble above each section's label; on the footer's edge, a last scene and a gold bauble. The page opens dark unless the visitor chose a theme, and a fir tree heads the tab's title. Nothing covers a line of text or a control (checked at fifteen widths, in both themes); for a visitor who asks for less motion the snow and the sleigh go and the figures stand still. A part of the page out of sight is not drawn at all: with a CPU slowed four times, the first screen kept the browser no busier than on an ordinary day, the rest of the page at most ten points busier (Chromium). Nothing is fetched on another day.
  • [site] Christmas — the strip plays the Nutcracker, and an excerpt to hear. Those two days the strip that shows a track playing shows the Dance of the Sugar Plum Fairy by the Berliner Philharmoniker and Sir Simon Rattle, and its round button, Christmas red, plays a 30-second excerpt, fetched at the first press only; while it plays, the strip shows what is heard — AAC, 44.1 kHz and the excerpt's own time.
  • [site] Christmas — a hidden gift to find on the 24th, five lifetime licences to win. On 24 December 2026, Paris time, the licence server wakes five lifetime licences one at a time, each at a secret moment in its fifth of the day. The first visitor to open the page after one wakes is shown a gold gift — unlike the decoration's red and green ones — in the band above one of the sections, and has twenty minutes to open it; the page then shows the licence key, keeps it on the device and says how to activate it. The announce bar says how many licences are left, then that all were found; the rules are on hidden-gift.html. Halloween's orange ghost and the gift are one game in the page (assets/days.js), each with its own prize, words and rules.
  • [site] The hidden gift, announced before the day. Until 23 December the announce bar gives its place every six seconds to the news of the 24th — a gift hides on this page, five lifetime licences to win — which leads to the rules; it takes the place of the orange ghost's, over since 31 October. A phone reads a shorter sentence, on one line from 320px, and the bar keeps its height with either.
  • [lic] The landing's game runs on 24 December 2026, the hidden gift (GHOST_DAY's default). The keys of 31 October nobody found play again, under the same [email protected], and a key found then — activated or not — never does. A deployed .env needs no change: it names no game day, and one that did would hold the server to that day.
  • [ui] Christmas — the installed app opens in the snow, every 24 and 25 December. By the device's own clock, the screen the installed app shows while it opens dresses up as the landing does: snow falls on its black ground, Santa's sleigh crosses the sky once, a fir tree whose bulbs blink stands beside the icon with two gifts set down at its foot, the two waves turn red then gold, and the line reads Your audio chain, merrily controlled. The sign-in page's screen too; nothing else in the app changes. The screen stays no longer than on another day: with a CPU slowed four times it began to fade at the same moment, the extra work spent while the app loads, every animation run by the compositor (Chromium). For a viewer who asks for less motion the snow and the tree stand still and the sleigh does not fly. The drawings are the landing's, fetched from the box on those two days only. The sleigh flies in the middle of the sky, clear of the tree, a phone held sideways included. The script that decides the days, halloween-boot.js, is now holiday-boot.js.
  • [core][ui][manual] A media server's album or playlist plays whole, or from one of its tracks. Browsing a media server such as MinimServer, an album only opened, and a track tapped inside it played alone: the album stopped after it. An album or a playlist of the server now carries Play and + on its row, which play it or add it to the queue whole — a tap on the row still opens it —; inside one, a track tapped plays the album from that track on, the tracks before it left out, as on HIGHRESAUDIO and Qobuz, and its + adds that track alone. A folder only opens, as does a server's photo album. The box reads the album page by page, where one read stopped at 200 tracks and would have queued a longer album cut short; past 5,000 tracks it refuses, saying how many the album holds, rather than queue part of it. On MPD, HQPlayer and a network speaker alike; measured on a private MPD with MinimServer: an album of 13 tracks queued in order, 11 from its third track. POST /library/queue takes a media server's album or playlist, and start_index / start_item_id for it.
  • [ui] A media server's long lists go on as you scroll. Every list of a media server stopped at its first 200 entries: MinimServer's 474 albums showed 200, ending on IRM, and nothing after it could be reached. The next 200 now come as the reader nears the end; past 1,000 entries they wait for a tap on Show more at the list's foot, since every row stays in the page — MinimServer's 6,472 tracks scrolled to their end took the page from 14 MB to 42 MB (Chromium). An answer for a list already left is dropped, where it could write itself over the one opened since. The library's album grid shares the mechanism, and asks for its next page by itself when the one just listed leaves the end in sight.
  • [ui][manual] In Chrome or Edge on a computer, signing in to Tidal no longer needs a paste. On the box's https address, coming back to the app after copying Tidal's return address is enough: the card reads it from the clipboard — once the browser is allowed to — and completes the sign-in. It reads only while it waits for Tidal, and takes nothing but Tidal's own return address (https://tidal.com/android/login/auth?code=…): a password copied to sign in, or another site's address, is left alone and never sent. A code the box has already answered about is not sent again on its own, even after a reload. In Safari — so on an iPhone or iPad —, in Firefox, or with the permission refused, the paste remains.
  • [ui][manual] Reopen Tidal page starts a Tidal sign-in afresh. While the card waits for Tidal, the button asks the box for a new sign-in and opens Tidal's page again: after a closed tab, a refused address or a box restarted meanwhile, only reloading the page used to help. It waits while an address is being sent, so as not to replace the sign-in that address completes. When the browser blocks Tidal's page, the card says so and offers it as a link, which a tap opens.

Changed

  • [ui] The manual's copy button keeps its 28px on a phone. The touch rule made it 44px there, and every code block grew to hold it: it took a fifth of a phone's code line, and a one-line block stood 56px tall instead of 45 (measured in Chromium at 390px). It now keeps its 28px, as on a computer — above the 24px WCAG 2.2 sets as a minimum target —, as the dialogs' buttons keep their 29px.

  • [manual] Installing the app on a phone, step by step and device by device. First run's steps 7 and 8 are numbered steps for the iPhone or iPad, Android, the Mac and Windows, and a line for Firefox, each certificate ending on a check; their tables are gone — on a phone, the answer sat off screen. The certificate is fetched on the box's name, http://<name>.local:8081/ca.crt (served whatever name the box is reached by, measured), with its address as the way out. The words are the devices' own, read in Apple's help for iOS 27 and iPadOS 27 and in Chrome's: on an iPhone, Share behind ⋯; on an iPad, Share → View More; on both, the Open as Web App switch, without which the icon only reopens Safari; on Android, the app's own Install bar, or ⋮ → Install and create shortcut, and the certificate found by searching Settings for CA certificate. The quick start no longer sends an iPhone to https://<ip-of-your-box> before the certificate is trusted, which Safari does not open; Installation says to press Enter at the installer's two questions on a new install — HTTPS, port 443; the FAQ and the landing's FAQ follow. Troubleshooting adds a profile installed but not trusted, an icon that opens in Safari, admin / admin123 refused after installing again — the accounts are kept —, and an interface installed again in the other mode or on another port.

  • [site] The landing's players, Stats and Tuning show the app as it is today. Six of the carousel's screens are captured anew, in the light and dark themes: the HIGHRESAUDIO, Qobuz and Tidal players, the radio — another station —, Stats and Tuning. The players now carry the Cover | VU meter switch under the cover, which moved everything below it 4.7 % down, and Stats and Tuning changed: their labels are measured again, on the new captures, whose light and dark themes line up to the pixel. The Output → DAC row now sits under the phone's bottom edge, so the Qobuz screen loses its Switch output mid-track label and Tidal's Your DAC names the chain line, which ends on the DAC. HQPlayer's screen keeps its capture: its player offers no meter, which follows the box's own MPD only. The dark captures are named -dark, where they were -black.

  • [site] The landing's footer and top bar lead where their words say. Releases in the footer opened the GitHub page of the packages, which holds the install commands and nothing else: it is now Release notes, and opens them. Community / Issues share a line. The top bar's button, which said Releases and opened that same page, now says GitHub and opens the site's repository, whose README presents Audiogravity and whose Discussions, Issues and release notes are the footer's links.

  • [site] The social preview and the README's image show the landing as it is today. Taken anew from the page by gen-og-image.js: no Release candidate ribbon, the GitHub button and the Blog link in the bar.

  • [site][manual] EDITIONS.md and the manual's editions table list what each edition holds today. Starter gains AirPlay and casting — what an iPhone, iPad or Mac sends to the box, or an app casts to it, which the core leaves open to Starter —, and in EDITIONS.md the one-click updates from the Admin page and the eight push notifications the box sends, where it named four. Pro's Level meter is the VU meters, as the app names them, now in the manual's table too; Performance adds its network tests (ping, iPerf3). EDITIONS.md also gives the platforms (x86-64 and Raspberry Pi, aarch64) and the licence reset on request when moving to another machine, as the landing does.

  • [site] The README says what 1.0 does, where to read about it and where to ask. Its features add the VU meters and their showcase, the player on a phone held sideways, HQPlayer as the output — on the box, or on another computer with the box as its NAA — with the streaming services going through it, Qobuz's purchases and your own playlists, Tidal's playlists, HIGHRESAUDIO's Vault, sections, playlists and search, and the radio's lossless filter. Documentation lists the guides, the blog and the press release; Support sends questions, ideas and setups to GitHub Discussions, bugs to the issues, and gives the contact address; Contributing sends feature ideas to Discussions → Ideas.

  • [ops][site] UPnP Bridge on an ARM box is upmpdcli 1.10.1, the author's latest. Its author publishes no ARM package, so Audiogravity builds one: it was 1.9.17, and 1.9.18, 1.10.0 and 1.10.1 have followed since — according to the author's notes, mostly for the media server (its Tidal and Subsonic plugins, icons for plugins, caching), with a crash fixed in OpenHome. Built from the author's signed sources, it is offered as an update in the Audio Software tab of an ARM box where UPnP Bridge is installed, at once with CHECK UPDATES or after the box's daily check; the two libraries it needs are unchanged. An x86 box gets it from the author's own repository, as before. An ARM box still on 0.9.37 or older can no longer install UPnP Bridge: it looks for the 1.9.17 package, no longer published, and updating Audiogravity first lifts that.

  • [ui][manual] Pasting Tidal's return address completes the sign-in: the Finish login button is gone. The card recognises the address by the core's own rule and completes on the paste; an address without a code — the sign-in page's own, say — is answered at once, and nothing is sent. Enter still sends whatever the field holds. A refusal by the box points to Reopen Tidal page; a box out of reach asks to retry, the sign-in being still pending there.

Fixed

  • [core][ops][manual] A service's output names its sound card: a card that wakes under another number no longer takes the music elsewhere, nor makes the DAC disappear. The kernel numbers the sound cards in the order their drivers wake, and that order moves: on the x86 test box the USB DAC was card 0, 1 or 2, and the two built-in cards swapped numbers. MPD, AirPlay and HQPlayer Embedded were given hw:<card>,<device>, and played on another card, or on none, once the numbers moved; and the number the guided setup reserved for the DAC (/etc/modprobe.d/ag-audio.conf) made the kernel refuse the DAC altogether when a built-in card took it first. Every output Audiogravity writes or compares now names its card — hw:CARD=Abacus,DEV=0 —, which ALSA finds whatever its number: measured with MPD in PCM, DoP and native DSD, HQPlayer Embedded in DSD128 and AirPlay, and MPD still playing on the DAC after a reboot moved it from card 0 to card 1. The guided setup no longer writes the reservation. At the update, the installer names the cards in the files the guided setup wrote and in HQPlayer Embedded's, and takes the reservation out once nothing that may be the DAC's is numbered with it. A number only says which card had it the day the file was written: an address is named only once its card is confirmed — by the kind of output its MPD block was written for, by the card's own name, by the card the guided setup chose — and nothing contradicts it; one that is not keeps its number, the installer says so, and the next update tries again. The step that decides whether MPD turns DSD into PCM before its volume no longer reads such a number either: the x86 box's files, written when the DAC was card 0, would have had the DAC's DSD turned into PCM, card 0 being the built-in card today. Measured on copies of the three test boxes' files: the x86 box's left numbered and untouched, the two ARM boxes' named. The manual no longer says the DAC's number is pinned: First run says outputs are known by their card's name, and Troubleshooting what a flagged line of the support report's Audio stack means now.
  • [core][ui] The output switch is offered for the services whose output can change, and only for them. The UPnP renderer — which plays through MPD — and Roon Bridge — which audio-config.json declares as roon — were said not to be steerable, though a switch of either worked: neither got its output pills on the phone. The NAA was said to be: its output is chosen in the HQPlayer that streams to it, and written in no file of the box. A switch of the NAA's output now says so — The NAA's output is chosen in the HQPlayer that streams to it, not on this box — instead of writing nothing; the computer's diagram lists under Steer → only the services whose output can change; and the UPnP renderer's pill lights MPD's output, paused too. GET /steering/status says which services can be steered, and each service of the pipeline's nodes carries steerable.
  • [core] The signal path no longer shows Spotify Connect. The model topology Audiogravity copied into a box until 2026-07-10 declared a Spotify Connect service, and the installer copies the model only to a box that has no topology: a box set up before still showed it, though Audiogravity has never played Spotify. The installer takes it out — only that entry, as Audiogravity wrote it, and only when nothing else in the file names it —, the previous file kept beside it. Comments that still spoke of it as a source are corrected.
  • [core] The signal path's Controllers view lights a phone app only while it plays through the box. It lit JPlay whenever MPD played, JPlay being declared as commanding the box's UPnP renderer, which plays through MPD: seen on the x86 test box with the renderer stopped and the music started from Audiogravity. The app is now lit only when the renderer carries the track — running, and MPD playing what it was handed, never one of Audiogravity's own streaming addresses nor a file of the library. For a media server's track or a radio, which Audiogravity and an app may both play, the renderer is asked: it hands back an app's own description of the track, and marks the one it makes for a track another player queued (measured with upmpdcli 1.9.17, through UPnP AV and OpenHome alike, an app that sends no description included). One question per play, shared with the cover-art lookup, the diagram never waiting for it; the lookup now asks the box's own renderer even when HQPlayer Embedded's answers too, and no longer takes a cover from an answer about another track. One case stays out of reach: a track Audiogravity plays again at the very address an app played it at, which the renderer still describes as the app's.
  • [core][manual] Choosing another USB drive as the music library in Guided takes effect when it is applied. Guided rewrote the library's mount for the new drive and left the previous one mounted in the folder MPD reads, until that drive was pulled out or the box restarted (measured on the x86 test box, with two disk images on a throwaway mount). The previous drive is now unmounted first — MPD stopped for that moment when it is what holds the drive, and started again once the mount names the new one —, and again once the new mount is in place, should something have read the folder in between and mounted the previous drive anew. A drive still in use refuses the change, saying what to do: close what reads it — a terminal open in its folder, a renderer playing from it — then apply again. The library's own drive still mounted from before it was pulled out while in use is unmounted the same way. A full guided setup refused so leaves the output as it was: the one chosen is recorded once the library is in place. The manual's First run says it.
  • [core][ops] With the library's USB drive unplugged, opening its folder gives up after 5 seconds, not a minute and a half. Whatever opened the music library's folder while its drive was pulled out — MPD playing a track or reading a cover, a screen of the app, a terminal — waited systemd's 90 seconds for the drive before No such device (measured on the x86 and ARM test boxes). The library's drive is now waited for 5 seconds. A drive is seen about 3 seconds after it is plugged in: plugged in and opened at once, it is still waited for; a hard disk slower to spin up when the box starts can miss that time, and the next opening of the folder waits for it again. Guided sets the wait when it chooses the drive and takes it off the drive the library leaves; a box whose library drive was chosen before gets it at the core's next start, and uninstalling removes it.
  • [ui][manual] A track that cannot be reached or read no longer reads as an unavailable output. When MPD could not open a track — an internet radio off the air, a box offline, Audiogravity restarting while a streaming track was due, a damaged file —, the notification and the fullscreen player said Output unavailable, though the output was fine. They now say Track could not be reached — told by the words of the library MPD fetches with, the same on the x86 and the ARM box — or Track could not be read; an output that fails keeps its message. The manual's troubleshooting explains both.
  • [core][ui][manual] A track a streaming service cannot provide no longer reads as a failed play, nor puts the core's key on screen. On a HIGHRESAUDIO playlist whose first track the service gave no address for — answering, for a subscribed account, with no address at all, every time —, the app said Playback failed while MPD went on to play the next track, and its message quoted the box's own stream address with the core's API key. The addresses MPD is given carry no key any more — those paths are public, the key bought nothing —, and a key in what MPD says is masked wherever it is shown: a play's error, the player's output, the support report. Measured on a private MPD 0.24.5: playing or paused, MPD refuses the first track and plays the next; stopped, it accepts and moves on; with nothing it can play, it stops. The play's verdict is now read from where MPD went, woken by MPD's own changes of track (idle player, through one watch the library rescan's wait now shares): the play succeeds and says HIGHRESAUDIO could not provide “Jolene”: the next track plays. — POST /library/queue answers refused, refused_titles and a new refused_by; a play of which nothing could be had answers 503 in words, HIGHRESAUDIO could not provide any of the 16 tracks.; in random order only the refused track is named. From a stopped MPD, such a play said nothing at all: the error MPD stopped on now stays on the local output for two minutes, and the app shows its sentence where it said Output unavailable. Tidal's 404 for a track it does not make available is read the same way. The journal no longer blames the subscription (no address given, subscription active). The manual's troubleshooting says what the message means, and to check the service in Library → Sources when every track fails.
  • [ops] The interface's installer gives the app's install steps as they are today. Its last screen told an iPhone owner Share > Add to Home Screen — Share now sits behind ⋯, and without Open as Web App the icon only reopens Safari —, an Android owner to accept an Install app prompt Chrome no longer calls so, and left Android's certificate to menus that vary by brand. It now gives the manual's steps, shortened: remove an older icon, fetch the certificate on the box's name with its address as the way out, install it, then turn on the line named exactly as the iPhone lists it — the authority's name is read from the certificate, which keeps the hostname of the box's first install —; Android's steps; the iPad's; and a link to the manual's step 7 for every other device. A test holds the words the installer and the manual share. In the next release of the interface.
  • [site] EDITIONS.md no longer promises what the Systemd tab cannot give. It said the tab's priorities eliminate scheduling jitter for bit-perfect, glitch-free playback: priorities keep the processor for the audio services when the box is busy, and bit-perfect — the data reaching the DAC unchanged — does not depend on them. The VU meters' showcase is no longer said to be for a tablet only: it opens on a phone too.
  • [ui] A track removed from the queue leaves it at once, and the list no longer flashes. Removing a track — swiped, by its arrow, or with Clear — read the whole queue again behind Loading…: the list blanked, every cover was drawn again, and a reader who had scrolled down was sent back to the top of it (measured in Chromium at 390 px). The track now leaves under the finger before the box answers, the rows around it stay where they are, and the list is read again behind them; a refused removal puts the track back, with the box's reason. A track whose removal is on its way stays out of every reading, however close two removals or a Clear come. Refresh greys while its reading is out, the list no longer flashing to say so. Loading… now stands in only while there is nothing to show — on the queue, and on Outputs and Sources, which blanked the same way after a switch of output or a connection, Sources rebuilding every card below over about a second; the output asked for says SWITCHING… until the list is back.
  • [ui] The button beside the track playing in the queue plays and pauses, and the queue keeps up with the music. It showed a pause sign whatever the player did, and opened the full-screen player. It now plays and pauses the player that holds the queue — the box's own, stopped or paused, even while a phone plays over AirPlay, or the Roon zone shown —, shows what that player does, changes at the tap, and stays greyed until the player's state is known; a play refused for want of Pro says why. The full-screen player now opens from the track's cover, on that same player. It takes the finger over 44 px, drawn at 36. While the queue is on screen it follows the player: it is read again a moment after the track changes, and when the app comes back on screen, where it went on showing a track long over until Refresh. A queue changed from another device still needs Refresh.
  • [core][ui][manual] The queue's Up next lists what comes next, not what has played. It listed every track but the one playing, the tracks before it in the queue included — played or skipped, and not to play again: the queue's first track headed the list while a track far below it played. It now lists the tracks after the one playing; in random order, every other track, since any can come next; with repeat on, the tracks after it, then those from the top of the queue up to it, the order they will play in — measured on a private MPD 0.24.5: repeating, it plays the first track after the last; not repeating, it stops. The played tracks stay in the queue, and Clear still removes them. Turning random order or repeat on or off in the player updates the list, as a change of track does. GET /library/queue gains shuffle and repeat.
  • [site] The landing and the README no longer say Audiogravity tunes interrupts. Their five-layer stack listed IRQs in the kernel layer, which nothing in the app does: the layer now reads RT scheduling · CPU pinning · ALSA. On a box tuned that way today, interrupts are confined by hand, on the kernel's command line; doing it from the app is in the backlog.
  • [ui][manual] The mini-player's progress bar no longer moves the track. It took the finger on a band 14 px high along the very bottom of a phone's screen — the 4 px bar and an invisible margin —, where the phone's own gestures start: a touch there, or a swipe of the mini-player that began there, moved the track without being meant to (measured in Chromium at 375 px; on a computer, the band also covered the top of the page's footer). The bar now shows the position only, and moving in a track is the full-screen player's. A swipe that starts on it — down to fold the mini-player, up to open the full-screen player, sideways to change source — is a swipe. The manual says where to move in a track.
  • [ui] The album-details bubble no longer flickers when the cover is tapped. The bubble was drawn inside the mini-player's cover, which shrinks and fades under the finger like every button on a touch screen: the bubble shrank and faded with it — see-through for a moment, its top jumping 23 px — as it opened or closed (measured in Chromium at 375 px). It now hangs beside the cover, which alone takes the press. The cover is a real button, which the keyboard reaches too.
  • [core] The album tracklist names each track once, disc by disc. The bubble's tracklist of a streaming album is read from the box's queue: an album queued three times listed each of its 19 tracks three times. An album kept in the library in two editions, a FLAC and a DSD, could list its tracks twice, and an album on two discs read 1-1, 2-1, 1-2, 2-2 (measured on the x86 test box's library). Each track now comes once — the same disc, number and title, case and accents aside; a track with no disc number is on the first —, in disc order, then track order. A track number that names none, a vinyl side's A1, no longer empties the list. GET /audio_pipeline/album-tracks answers so; the library reads a track's number the same way.
  • [ui] A missing cover shows a music note, everywhere. A track or an album without artwork showed a disc, close to the drawing of an output, and a playlist a list, where the player and the listening history show a music note; a media server's album without artwork showed a folder. The music note now stands for every missing cover — queue, search, albums, playlists, a media server's lists —; a folder keeps its folder. On a media server's lists, a track's Play and + touched and read as one button: they are now spaced as on the library's other rows.
  • [ui] The full-screen player's Queue button looks like the Queue tab. It drew the Browse tab's three lines: the two buttons that open the queue had two drawings, one of them Browse's.
  • [core][ui][manual] An update leaves each service as it was: running on the new version, stopped, or disabled. Each package's own scripts did something else to its service (read in them, and measured on the x86 test box): MPD's stopped MPD and started nothing — after an update from Audio Software, the box had no player until Restart required was pressed —; HQPlayer NAA's left the previous version running; Shairport Sync's turned the service back on at boot and started it, even one stopped or disabled, as upmpdcli's and HQPlayer Embedded's start one that was only stopped. The box now reads how the service stands before the operation and puts it back once it is over: one that ran runs again on what was installed — started, or restarted when it still ran the previous version —, and one stopped or disabled stays so. What is asked of it from the Services page or by a profile while the package downloads wins. An update that changed no version touches nothing; one that could not put the service back says so. The update's confirmation says This restarts … if it is running, and REPAIR — an install over a package still there — now says it too.
  • [core][ui][manual] Restart required restarts the right service, and shows only when the service could not be restarted. After Shairport Sync was updated, the badge answered Service 'airplay.service' is not managed: the card asked to restart the service's key in audio-config.json, airplay, which names no service — nor do the keys of HQPlayer NAA and Roon Bridge. It now restarts the service itself, shairport-sync.service. The badge was also lit after every install and update of a package with a service, though some packages restart their service as they update it: Shairport Sync's had, seconds before the badge asked again (measured on the x86 test box), and a press only cut the music. It now shows only while a service still runs the version from before because it could not be restarted, and the box asks systemd again each time it describes the package: the badge goes out once the service has restarted, whatever restarted it, and every device shows the same, where it lived in the browser that ran the update. Pressed on a second device, it restarts nothing the first already did. The API's packages gain service_unit and restart_needed, and lose restarts_after_install; the manual describes the badge.
  • [core] The services audio-config.json declares are read in one place, and each one's unit found one way. The Services page, the audio configuration and the profiles each read them and found a service's unit by a rule of their own: the audio configuration answered roon.service for Roon Bridge, whose unit is roonbridge.service. A rewrite of the file within the same few milliseconds was not read — its time alone could not tell it from the previous version —, and an entry naming its unit as something other than a name made every action on the services fail. The file is read once now, its versions told apart by more than their time, and an entry that names no unit is left out.
  • [ops][site] The ARM packages of UPnP Bridge come with their source code, and the third-party notices give their real licenses. Audiogravity builds upmpdcli and its two libraries for ARM and serves the packages, which makes it their distributor: upmpdcli and libupnpp are under the LGPL, version 2.1 or later, with parts under the GPL, and their source has to be offered with the packages — it was not. The archives they are built from, exactly as their author publishes them, and the script that builds them are now served beside the packages, in packages/upmpdcli-arm64/source/, and build-upmpdcli-arm64.sh --publish replaces them with each rebuild. THIRD_PARTY_NOTICES.md listed upmpdcli as GPL-2.0, said that every package came from its vendor, and named neither library; it now says which part is under which license, and reproduces the BSD notice of libnpupnp.
  • [site] The third-party notices name HQPlayer Embedded among the software the box can install. Their table listed six of the seven packages the Audio Software tab installs; HQPlayer Embedded, proprietary and fetched from Signalyst like the NAA, was missing.
  • [ui] Tidal's sign-in page could send the app's tab elsewhere. Opened without cutting its opener, the page kept a handle on the app's tab: a page from another origin, opened the same way, redirected it (measured in Chromium). The handle is now cut as the page opens — not with the noopener feature, under which a blocked page could no longer be told from an opened one.
  • [ui] Edit station opens its form in Favorites, and in sight. Every card of My Live Radio and Favorites offers Edit station, but the form was drawn by My Live Radio alone: in Favorites the tap showed nothing. And above a long list it opened out of sight — 1,348px above the screen for the 30th station (measured in Chromium) —, which read as nothing too. The form now opens above the list it was asked from, and the page goes back to its top to show it; it closes when the tab changes or its station leaves the list, and a save that ends after the reader moved on no longer closes the next form.

[1.0.0] - 2026-11-06

Added

  • [site] A blog post on the Rega Planar 3. How Rega's turntable of 1976 became the deck others at its price are measured against, and how it sounds by Stereophile's reviews of 1984, 1985, 1996, 2008 and 2017 and What Hi-Fi?'s of 2021 and 2024: its rhythm, its tone, its bass, what it gives up against far dearer decks, and the cartridge, on which Michael Fremer and John Atkinson disagree. Its five generations, from the Planar 3 to the P3, P3-24, RP3 and the Planar 3 of 2016, and the RS Edition of 2024; its rivals, the Linn LP12 first; a green P3 of about 2005 with an Elys 2, playing through a NAIT 2 into a pair of Rogers LS3/5A; setting it up from Rega's own guide. Rega's photograph heads it.
  • [site] Hi-Fi World's "NAITology" of October 2007, kept on the site as a source of the NAIT 2 post: the scan of its three pages, linked from the post. robots.txt asks search engines not to read it — a PDF cannot carry the noindex tag the archived web page does, and GitHub Pages sends no header —, and two tests keep every archived scan out of the crawl and every archived page readable, its noindex included.

Changed

  • [site] The blog's NAIT 2, LS3/5A and Quad 33/303 posts start from how they sound, and why they became references. Each opens on its standing, then gives most of its room to its sound as the specialist press heard it across the decades — where reviewers agree, where they disagree, and in what conditions they listened —, then its rivals and successors and what to pair it with; its history and figures come after, shorter. Read in full for them: Stereophile's reviews by J. Gordon Holt (1977), Robert Harley (1989), Sam Tellig (1990), John Atkinson (1989, 2007), Art Dudley (2016, 2019, 2020) and Herb Reichert (2021), The Absolute Sound's by Paul Seydor (2016), Robert Harley (2018) and Chris Kelly (2025), Hi-Fi World's of 2007 and Hi-Fi News' of 2025. Gone: the sections on streamers and on Audiogravity where the equipment has nothing to do with them, a comparison of a cartridge's output with a DAC's, and estimates the posts had made themselves, such as a listening level worked out from the BBC's figure. The titles follow: The Naim NAIT 2: the small amplifier that became a reference, The BBC LS3/5A: the little monitor that became a legend, The Quad 33 and 303: Peter Walker's transistor classic, and its return. The Quad post shows an original 33 and 303, Quad's photograph, and, at its head and on its card, a 33 of 2024.
  • [site][manual] The landing and the manual no longer call it a release candidate. The landing's corner ribbon and the « (Release candidate) » of its footer are gone, and so is the FAQ's « What does “Release candidate” mean? », in the manual and on the landing. The announce bar, the light / dark toggle and the GitHub button take back the room they kept for the ribbon; the announce bar still holds one line, cut at its own edge below 1100px. The release notes open on what Audiogravity 1.0 does, area by area, before what changed since 1.0.0-rc.4.

Fixed

  • [core][manual] Stats counts AirPlay and the radio in the right sound quality. Two causes. A station started straight after another track could count as Unknown for its whole listen: the box read it a split second before the player had decoded it, so without a format, and the format that came 0.5 to 2.3 s later (measured on a private MPD 0.24.5) went out only when something else changed or while a screen was open; France Inter, which names no song, kept none — both of its listens counted Unknown on the lab box had started straight after another track. A source's format now goes out when it arrives, its bitrate left aside: that moves at every read, and would send a state every second while a screen is open. A file's codec read from its header once it plays (an .m4a) now comes in a further state too, as the API said it would. And AirPlay counted as Hi-Res: the player showed ALAC written in and the 32 bits shairport-sync opens the sound card at, whatever it receives, where the boxes run classic AirPlay — DietPi's shairport-sync 5.6.2 without NQPTP, Debian's 4.3.7 built without AirPlay 2 —, which only sends ALAC at 16 bit / 44.1 kHz (during a session, shairport-sync received ALAC/44100/S16_LE/2 and opened the card at S32_LE). AirPlay is now read from what shairport-sync says it receives — its SourceFormat, else a classic service's only format — and never from the card. On the x86 box, the old core and the new side by side with no screen open: the same switch to France Inter counted Unknown against lossy (AAC, 220 kbps, 48 kHz), the same AirPlay track Unknown against CD quality. Listens already recorded keep their category. The manual says what Unknown means under Sound quality.
  • [core] A change seen as a screen opens now reaches Stats and the screens already open. The state a screen receives when it opens became the box's record of what had been announced, without being announced: a change it saw first — AirPlay stopped without a word, its service switched off during a session — reached neither the listening history nor the other screens until something else changed, and an AirPlay listen stayed open on the lab box. That state is now only read for the screen that asked; the box's watch, which a screen opening wakes, announces the change as it does any other.

[1.0.0-rc.4] - 2026-11-02

Added

  • [core][ui][site] A level meter: two needles that follow what the box's player plays, in step with the DAC. In the full-screen player, a Cover | VU meter switch below the cover puts two dials in its place, left above right — the choice kept on each device —, and a Showcase mode button opens them over the whole screen, side by side in landscape, with what plays below: the screen stays on and the portrait lock is lifted while it is open. On a phone held sideways the showcase fits the screen — it overflowed 2 to 49 px, measured in Chromium on phone sizes —: the dials take the height left and Exit moves to the end of the track's line. The dials are drawn in one ink, the theme's text colour, in light and in dark alike: the scale's red zone is a heavier band of it, its figures bold, and the peak lamp a ring that fills. The needles read the RMS level of each 40 ms of the sound, 0 VU at −10 dBFS, and swing as a VU meter's do, 99 % of the way in 300 ms; a lamp lights above +3 VU. The levels come from a copy of MPD's stream: Audiogravity gives MPD a fifo output, audiogravity-meter, off, after the sound cards — the guided setup writes it, and the installer adds it to an MPD configuration Audiogravity wrote whose socket is in /run/mpd, then restarts MPD; a configuration without Audiogravity's mark is left alone. The copy is on only while a screen shows the meter and MPD plays on the DAC: MPD reports a failed output only when all of them fail, so a copy left on would have kept it playing, silent and without an error, to a DAC that could not be opened — measured: with the copy on MPD said it played, with it off it paused on Failed to open ALSA device, which Audiogravity shows. A DAC another player holds — Roon, AirPlay, any program — does not count: the process holding the card, which the kernel names, must be MPD's. With the copy on for such a DAC, Play had MPD say it played, silent, where it says Device or resource busy — measured with the DAC held by another program, before and after. On, it cost MPD 0.4 % of its core at 44.1 kHz and 4.2 % at 384 kHz on the lab's x86 box, and the DAC's buffer stayed as full as without it. The copy runs ahead of the sound — 3.4 ms a minute at 44.1 kHz, 11 in DSD128, 42 at 384 kHz, measured on the lab's DAC —, so each window waits until the card has played it, read from the card's position in the kernel: the needles move with what is heard, through a seek, a pause or a new track. Only a stride of the samples is read: over a whole track, its levels were within 0.12 dB of every sample's on nine windows in ten, the median 0.00 dB. DSD is read from its density of ones. On a box where MPD turns DSD into PCM before its volume (audio_output_format "*:32:*"), the copy is read in the format MPD's outputs receive, which MPD's status does not name — it names the file's. The copy is left out of every list of MPD's outputs: the output switch, DoP and the DSD readings see the sound cards only. A copy left on — a core stopped while a screen showed the meter, MPD restarting when the last one closed — is turned off as soon as MPD answers. Paused, the meter does nothing until MPD plays again. With the copy on, MPD wrote Decoder is too slow; playing silence to avoid xrun at each start of a 384 kHz or DSD256 track (8 out of 8, none out of 2 without it), at the very start only; whether it can be heard was not checked. For what the box's own MPD plays — its music, the radio, the streaming services —: for AirPlay, Roon, HQPlayer or a network speaker the switch is not offered, an open showcase says why in the dials' place, and nothing is turned on. Pro: on a box without it, the switch shows a padlock that opens the licence panel. New: GET /player/levels (server-sent events, Pro), PlayerState.level_meter. The manual's Listening chapter describes it under The level meter, with the player and the showcase on a phone held sideways, and the landing lists it among the Pro features.
  • [site][manual] Search engines and link previews know what each guide, post and chapter is. A page pasted into a forum or a message shows a preview: a post's photograph, wide; the site's logo for a guide, a chapter or a list — not the landing's own preview, whose price and Install now would make a guide read as an advertisement. Each page describes itself to search engines in schema.org terms: an article — a post with the day it was published —, the latest day its Checked on line names, and its path from home, Audiogravity › Guides › …, which a result can show in place of the address; the lists of guides, posts and chapters describe themselves as collections. A chapter's snippet in search results takes the Contents' words for it — Listening — sources, transport, the Now Playing screen — where it gave its name alone. The first image of every page loads at once and the others as they near the screen: a post's photograph, the largest thing on screen when it opens and which search engines time, was fetched after the layout.
  • [site] Finding your way in the guides, the blog and the manual. The top bar links the guides and the blog to each other, the section being read underlined; the manual's names the manual alone, and a guide that has something to say about Audiogravity links to the chapter that says it, in its text. Below 360 px wide, the bar keeps the logo without the name. A page's parts are listed under its own entry in the sidebar, the part being read marked as the page scrolls and kept in view — a chapter runs to 33 parts, a post to 13 screens on a phone —; on a phone they are in an On this page menu under the title. Each page ends on the page before it and the page after it in its list — on the blog, the newer and the older post —, and on a link back to the list. The guides are filed under four themes, in their list and in the sidebar, each title on a line of its own with what it answers under it; the "Guide N" under their titles is gone. The blog's front page shows its posts as cards, each with its photograph — which grows a little under the pointer, and stays still for whoever asked for less motion —, its date written out and its reading time. Each guide's short answer sits in a box at the top, and the line saying what a guide or a post was checked or measured against, and when, moves under its title, after a check mark.
  • [site][manual] A community for owners, on GitHub Discussions. The landing's footer links it as Community, beside Issues; the roadmap's invitation to say which direction matters most now leads to the community's Ideas, instead of the issues; the manual's Getting help sends questions, ideas and setups there, and keeps bug reports for the issues. A pinned welcome says where each kind of post goes.
  • [site] A blog post comparing HIGHRESAUDIO, Qobuz and Tidal. Who runs each — a Berlin company led by its founder; Xandrie, a mostly French family-owned group, since 2015; TIDAL Music AS of Oslo, part of the American group Block since 2021 —, the size each claims for its catalogue, the countries it serves, what it streams and sells — FLAC up to 352.8 kHz in HIGHRESAUDIO's catalogue, read through its API, and DSD and DXD to buy; FLAC up to 24-bit / 192 kHz at Qobuz, which never streams DSD, an album bought in DSD streaming as PCM in the lab; the same on Tidal with Dolby Atmos, MQA dropped in July 2024, and music sold only by US artists who upload their own —, its prices on 2026-11-02, how you listen — Qobuz Connect, TIDAL Connect, one stream at a time per account —, what artists get — Qobuz alone publishes an average, $0.01873 per stream —, what each does about music made by AI, and which suits what. A short last section says what Audiogravity does with each: the quality it asks for, purchases, favourites, playlists. Everything is dated and sourced; Tidal's own pricing page could not be read, so its prices are its help centre's, in US dollars. Its photograph, by Vu Nguyen, is from Unsplash.

Changed

  • [ui] The full-screen player shows everything at once on a phone held sideways. It kept its upright layout there, one column to scroll through, 380 to 430 px of it out of sight on the phones measured. From 560 px wide it now takes the tablet's two columns: the cover — or the level meter — on the left, as large as the height leaves; the title, the format, the controls, the progress and the output on the right. The page dots move up under the header's title, the album joins the artist's line, the artist taking at most 60 % of it, and a long title keeps to two lines. Next shares a line with the output where the screen is wide enough; narrower — an iPhone SE or mini, a small Android phone — it moves under the cover, on one line. The page runs under the notch of a phone held sideways: the header and the columns keep clear of it. Measured in Chromium on phones on their side from 560 × 320 to 932 × 430 — the cover or the meter, with Pro or without, Next or not, short or very long titles —: nothing overflows. Under 560 px wide, a phone sharing its screen with another app, the single column stays, and scrolls.
  • [site] The blog post on the Quad 33 and 303 shows the inside of the new 33, and takes the new pair's size and power from Quad's manuals. Quad's picture of the 2024 33 with its cover off, its toroidal transformer in view, and links to Quad's user manuals for the 2024 33 and 303. By those manuals, the new 33 is within 2 mm of the original in every dimension and the new 303 is 1.7 cm taller — the post said, after Hi-Fi News, 15 mm deeper and 11 mm taller —, and the 303 gives 49.5 W into 8 ohms, not 50, and 170 W bridged into 4 ohms.
  • [site] The blog post on the Naim NAIT 2 shows the NAIT 50. What Hi-Fi?'s photograph of it at the High End show in Munich, in May 2023, at the head of its section, credited to What Hi-Fi? and linked to its news of the launch.

Fixed

  • [ui] The progress bar starts the next track at 0:00, and follows a track that goes back. After Next, the full-screen player's bar carried on from where the previous track had stopped and stayed ahead of the music until the player was closed and reopened — with the sources that say 0 as a track starts: Roon, AirPlay, a network speaker. Neither bar, the full player's nor the mini-player's, ever moved back on its own either: a track restarted, moved back from another device or app, or followed by another of the same name left both up to 2:36 ahead (measured on MPD, 2026-11-02), and a track played again with Previous resumed where it had been left. Both now follow the source back as soon as its position drops — within half a second on MPD; a title blank for a moment no longer counts as a new track, and on a source that gives no position (AirPlay through shairport-sync 4.3.7, Debian's) a new length under the same title does.
  • [core][manual] AirPlay no longer offers a seek it cannot make. The bar of an AirPlay track could be dragged, and the box refused every move: shairport-sync declares it cannot seek (CanSeek false, in 4.3.7 and 5.6.2); where it gives no position, the bar then fell back to 0:00. An MPRIS player's can_seek now follows what it declares — false for AirPlay, unchanged for a player that says nothing. The manual's AirPlay section says to move within a track from the iPhone, iPad or Mac sending it.
  • [core] Each AirPlay track shows its own cover. On a change of track, shairport-sync names the new album about 0.3 s before its picture; the previous picture, read in between, was kept as the new album's cover for a day and shown in place of the right one — 3 tracks out of 5 on 2026-11-02, and the same album again when replayed with Previous. A picture AirPlay gives is now read as it is, never stood in for by its album's; one already gone falls back on MusicBrainz and iTunes for its album, as before, without hiding the file when it comes back. A cover a browser already received wrong stays in its cache for up to a day.
  • [site] dev.sh serves a page asked for without its .html, as GitHub Pages does. The generated manual, guides and blog link to each other that way, and every such link answered 404 on the local server: the sidebar and the links between pages could not be tried before publishing. /page is served from page.html when only the latter exists, its query string kept.
  • [manual][site] The streaming services' section no longer says that Tidal sells nothing, nor that HIGHRESAUDIO streams up to 352.8 kHz. Tidal sells music since 2026, from US artists who upload their own; the manual now says only that Tidal, unlike Qobuz and HIGHRESAUDIO, has no Purchases shelf. HIGHRESAUDIO's catalogue lists FLAC up to 352.8 kHz, but a stream at that rate was never tried: the manual and the site's README now say what Audiogravity does — it plays each album's FLAC at its own resolution, as HIGHRESAUDIO sends it.

[1.0.0-rc.3] - 2026-10-31

Added

  • [site] A blog, beside the manual and the guides. Dated articles, the newest first, under a Blog tab after Guides in the landing's bar, on the manual's pages: each says when it was published and how long it takes to read, and the sitemap lists them. The first two are on Naim. The Naim NAIT 2: its inputs and power from Naim's manual and Stereophile's bench, what three reviewers heard, buying and servicing one, and every NAIT from 1983 to the NAIT 50 of 2023 on a timeline. How old is your Naim?: the year of each serial number from 1982 to 2022, from the table Naim's website gave until early 2024 — it now asks owners to write to its service department —, with its gaps and overlaps explained, and Naim's beginnings with its two founders. A NAIT 2 servicing guide whose site has closed, hi-fi AF's, is kept on the site as it was, credited to its author, with an address to ask for its removal; its 39 photos are shown in WebP at most 1640 px wide — 5.8 MB to read it all, where its original photos weighed 21.7 MB —, and the five it offered full size keep their original behind their link.
  • [site] Three more blog posts. The Heed Abacus: Heed's DAC of 2017, its inputs and formats from Heed's manual and from Linux, its streaming card and the Abacus S, and why its DSD stops when its USB volume drops below 0 dB — measured in the lab: DoP falls silent, native DSD turns to noise. The BBC LS3/5A: the BBC's monitor for outside-broadcast vans from its report of 1976, the 11-ohm revision of 1987 from Stereophile's measurements, and what a Rogers pair asks of a NAIT 2 and a streamer. The Quad 33 and 303: Quad's first transistor pre- and power amplifier, of 1967, from Quad's own manuals and circuit diagram — a DAC goes on the 33's Tape Replay input set to its highest level, not on Radio, which takes 2 V at most, a DAC's level at full scale —, Hi-Fi News' measurements of the originals, and what the versions of 2024 changed; the DIN guide points to it. Each photograph is credited to its source. The NAIT 2 post links to the LS3/5A one, and its tables now fit the column.
  • [site] A new guide: A streamer into an amplifier with DIN inputs — why a 5-pin DIN cable that plays on one amplifier may give one channel on another, and why a socket may carry power; how far a DAC's 2.1 V outruns a line input rated at 75 mV; where to set the volume — on the amplifier, on a PCM512x DAC's −6 dB analogue step, or in the player —; and how to check both channels with speaker-test. Like the others, it says what it was checked with and when, and lists its sources.
  • [ui][manual] The installed app's icon counts the notifications that arrived while it was out of sight. On an iPhone or iPad (iOS 16.4 or later), the icon of the app added to the Home Screen shows how many of the box's notifications arrived since the app was last opened on that device; opening the app or coming back to it clears it, the sign-in page included. A notification that arrives while the app is on screen is not counted; each device keeps its own count, in its service worker's IndexedDB; turning Badges off in the device's notification settings hides it. On an iPhone, a worker woken by a push sometimes failed its first opening of that store (UnknownError: Unable to open database file on disk), and three notifications arriving together showed 2: each change of the count is now tried twice, one change at a time — measured on an iPhone, three notifications at once on a worker asleep, the first opening failing, the second succeeding: 3. Chrome and Edge offer the same to an app installed on a computer (the API answers in Chromium 145); the number on a computer's icon was not tried. The manual says what the number means.

Changed

  • [manual] The certificate steps on an iPhone or iPad say where each switch is. The trust switch sits at the very bottom of About, iOS asks to confirm it with Continue, and a profile no longer offered at the top of Settings is found under General → VPN & Device Management — or downloaded again.
  • [core] Covers come from MPD in large pieces, at a fraction of the processor. MPD sends a picture in pieces of 8 KiB unless told otherwise, and opens the file again for every piece — for a cover inside an M4A, ffmpeg reading the file each time: on the lab's x86 box a cover of 844 KB took 104 requests and 510 ms, 476 of them MPD's own processor time. The core now asks MPD for pieces of 1 MiB (binarylimit): one request, 12 ms, 7 ms of processor; a cover of 44 KB went from 34 to 5.5 ms. Each piece is acknowledged at once: in pieces that large, a cover MPD sends in several pieces otherwise waited about 40 ms for Linux's delayed acknowledgement — a cover.jpg of 330 KB, sent in two pieces, took 43 ms, 1 ms with it (MPD 0.24.5). An MPD that refuses the command — older than 0.22.4, or with its max_output_buffer_size set under 1 MiB by hand, where its default is 8 MiB — keeps its 8 KiB pieces, and the picture comes whole as before. MPD 0.24.4 on ARM accepts it.
  • [ui] The Systemd editor's preset hint reads like the form's other hints, and a profile name takes the accent on hover. The hint under the presets is in italics, as the form's other help texts are. A profile name, which dimmed to grey on hover, takes the theme's accent — the colour its rule named from the start, under a name no theme declares.

Fixed

  • [site] The landing's bar keeps every label on one line, and its Releases button clear of the ribbon. Between 951 and 970 px wide and between 1161 and 1174 px, Signal path and Audio priority broke onto a second line; from 1161 to 1300 px — a 1280 px laptop among them — a corner of the Releases button sat under the Release candidate ribbon (measured in Chromium at every width from 320 to 1500 px). The button now goes below 1240 px, its link staying in the footer, and stands further in up to 1300 px; the links take a row of their own below 1030 px, where they did below 950. Measured again at every width: no label on two lines, the button 2.5 px clear of the ribbon at the worst, the theme button 4.3 px.
  • [core][ui][manual] The library's Refresh reads the music folder again, so music copied into it appears. Both Refresh controls of the local library — at the right end of the Library's bar and beside the Albums title — only rebuilt the list from what the player had already indexed: an album copied onto the NAS never showed, whatever was pressed. A Refresh now has MPD read the music folder first, then reloads the list: the icon turns and the title's link reads Refreshing… meanwhile, the albums already listed staying on screen — 9 to 26 s on the lab's 6,474 tracks on a NAS, music playing in DSD64 without a dropout. A Refresh always ends with a scan that began after it was pressed; presses that come together share one, and no more than one waits behind a scan already running. When the music folder reads as empty or does not answer while the library holds music — a NAS that is off, a share not mounted —, no scan is asked for: it would empty the library and give every album a new date in Recent (measured); the library is left as it is and a message says so. Moving to another source during a Refresh frees it at once. The API gains POST and GET /library/rescan; refresh=true on /library/albums keeps rebuilding the list from the player's database only. The manual says what to do when music is added, and what the message means.
  • [core][ui][manual][site] MPD's sound is no longer ranked below the rest of MPD. MPD makes its own sound threads real time, at priority 40; the Systemd tuning page's Audio Optimized preset put the whole service at SCHED_FIFO 80, so everything else MPD does — reading cover art, scanning the library — ran ahead of them. Measured on the lab's x86 box, MPD set above its sound by hand (RR 45) and its output buffer shortened by hand to 50 ms: while MPD read 40 album covers, five at a time as the core asks for them, during DSD64 playback, the sound threads went 47 ms without the processor and the DAC dropped out 16 times in 10.5 s; with them above the rest, none. The preset's FIFO 80 ranks them the same way, higher, and MPD's default buffer, 500 ms, which Audiogravity leaves as it is, would ride out such waits: neither was measured. On MPD the preset now leaves the policy and the priority at Default. A real-time priority of 40 or more on MPD is warned about by the core when it is saved, and by the editor as soon as it is typed, or on opening an MPD already set that way. The RT process monitor read MPD's main thread — NON-RT on an MPD as installed, whose sound is real time, which sent readers to the very setting that caused the dropouts —: it reads MPD's sound threads now, the weakest of them, and shows OUTRANKED when another part of MPD runs ahead of its sound. It also took a policy carrying reset on fork, as MPD sets it, for none. The manual, the real-time guide and the landing no longer give MPD priority 80. If Audio Optimized was applied to MPD: in Systemd tuning, open MPD, click Audio Optimized again, then Save Changes.
  • [ui] Systemd tuning's editor opens on the service's own settings. A value typed and then cancelled stayed in the form at the next opening — of the same service, or of another with the same setting — and Save Changes would then write it. Each opening now starts from the service's settings.
  • [ui] Four styles that named a token no theme declares take effect. The devices listed under Face ID / Touch ID in the configuration panel, with their remove button, the passkey list of a user's card — its key icon and No passkeys registered yet. — and the hint under the Systemd editor's presets took the colour of the text around them: they read in the themes' muted grey. The sound cards of the System page had square corners: they take the theme's small rounding, like the network panels beside them. The network test type chosen kept its label's colour: it is written in the accent.
  • [ui] The install banner no longer covers the mini-player. On a computer, the Add Audiogravity to your home screen banner sat 64 px from the bottom whatever the player bar's height, and hid the bar's title and controls (measured in Chromium at 1440 × 900); on a phone it bit 3 px into the bar. It read a height the player never published. It now sits above the bar — or its pull tab once folded — and the footer.

[1.0.0-rc.2] - 2026-10-27

Added

  • [site] dev.sh takes start, stop, restart and status. The landing's local server only ran in the foreground, whatever it was asked: ./dev.sh restart started a second server on the port the first one held, and stopped on Python's "Address already in use". start serves in the background, its output in dev.log; stop finds the server by its port, however it was started, and ends this directory's only; ./dev.sh alone serves in the foreground as before, and says so when a server already runs. Called by a relative path from another directory, it started and served nothing; it serves the site now.
  • [site] Halloween — ghosts, and the strip's Danse macabre to hear. On 31 October, ghosts in the page's ink rise through the margins beside each section's text, below its spider, and from behind the pumpkins on the footer's edge and at the foot of the phone — none on a phone, whose margins are too narrow, nor for a visitor whose system asks for less motion. A round button in the strip that shows a track playing plays a 30-second excerpt of the recording, fetched at the first press only; while it plays, the strip shows what is heard — AAC, 44.1 kHz and the excerpt's own time — and the box's again at the pause. A screen reader reaches and names the button, and putting it in moves nothing on the page.
  • [site] Halloween — an orange ghost to catch, ten lifetime licences to win. On 31 October 2026, Paris time, the licence server wakes ten lifetime licences one at a time, each at a secret moment in its tenth of the day. The first visitor to open the page after one wakes is shown an orange ghost in the band above one of the sections, and has twenty minutes to catch it; the page then shows the licence key, keeps it on the device and says how to activate it. The announce bar, in turn with the release's sentence, says how many licences are left, then that all were caught; the rules are on orange-ghost.html. With the licence server out of reach, the page stays as on any day; the page decides nothing, and can neither make a ghost appear nor know a key before it is given.
  • [site] The orange ghost, announced before the day. Until 30 October the announce bar at the top of the landing gives its place every six seconds to the news of the 31st — an orange ghost, ten lifetime licences to win — which leads to the rules; then the release's sentence comes back, its words untouched. A phone reads the news in a shorter sentence. The bar holds still while the pointer is over it or the keyboard is in it. On the 31st the game's own news takes that turn; from 1 November the release's sentence stays alone.
  • [site] Two new guides: “Undervoltage detected!” on a Raspberry Pi — what the kernel's message means, how to tell whether a Pi's supply has fallen short (vcgencmd get_throttled, the kernel's log), what it does — measured on a Pi 4 over six days: 68 drops a day, the processor at 600 MHz while Linux still showed 1.8 GHz, playback keeping up — and which supply and cable each model needs; Lossless internet radio — which stations stream FLAC and how to check what a station sends with ffprobe: of 333 streams read, 88 sent FLAC, 82 of them listed as Ogg by the Radio Browser directory, whose bitrates are what the stations announce; with a list of FLAC stations, how MPD plays HLS, and what its format readout says of a lossy stream. Like the others, each says what it was checked with and when, and lists its sources, with a short Audiogravity section at the end.
  • [lic] The orange ghost's two routes, POST /ls/portal/ghost-draw and POST /ls/portal/ghost-claim, answered to the landing's origin only. They hand out the pending lifetime keys the admin created under [email protected] and never create one; one chance per address per hour, one key per address, and a ticket that gave its key gives it again, so an answer lost on its way costs the winner nothing. Visitors' addresses are held in memory, for the day only. GHOST_DAY (2026-10-31) and GHOST_ORIGINS need no change to a deployed .env.
  • [ops] nginx — the orange ghost's routes are public, with their own request budget: 30 a minute per address and a burst of 10, turned away before they reach the licence server's single worker.
  • [ui] Halloween in the app, every 31 October. From midnight to midnight by the device's own clock — or any day with ?halloween in the address — the app dresses up: two spiders let themselves down on their threads in the screen's side gutters, under the bar and the tabs, and climb back; now and then a bat flits across the empty band of the title bar, right of the name (computer only); a ghost floats above the message of an empty page — no history, an empty queue, no result, no albums, an empty folder, no stations, nothing recorded in Stats — with Nothing here but ghosts under it, kept from screen readers; the fullscreen player's cover glows orange, like a lantern. The installed app's opening wears it too: a spider comes down above the icon as the wave, a pumpkin's orange that day, goes out, and climbs back as the screen fades; a bat crosses behind the name; the line reads Your audio chain, frightfully controlled. The sign-in page hangs the same spiders, behind its card, and a bat comes out from behind the name. Nothing covers a control or takes a touch, nothing makes a sound, and nothing moves on a device that asks for less motion, nor with the app's Animations setting off — the opening aside, as always; an app left open across midnight puts the dress on and takes it off by itself. Its cost is the browser's alone, the box only serving the files: in a browser without a graphics chip, a page left alone went from 11 % to 24 % of one processor core that day (Chromium, measured on 27 October 2026).
  • [ui] The radio's Halloween stations. On 31 October, a 🎃 Halloween button above the radio's search lists the stations the catalogue tags halloween — 13 on 27 October 2026, all MP3 or AAC — and clears the query, the country and Lossless only, any of which would leave few or none; the genre list offers Halloween that day, and keeps it while it is still chosen after midnight.

Changed

  • [site] The undervoltage guide names the supply of the Pi it was measured on: a third-party USB-C supply rated 5 V and 3 A, where Raspberry Pi asks for 5.1 V, with nothing plugged into its USB ports.
  • [ui][manual] In Minimal, the Stats view's sound quality and listening hours take the latency histogram's colours. The latency test's histogram draws its bands in grey, ink, amber and red; the Stats view drew every chart in ink and greys. In the Minimal theme, the sound-quality bar and its legend now show DSD in the histogram's grey, Hi-Res in ink, CD in amber and MP3/AAC in red, and the week's hours run from grey, the quietest, to red, the busiest — each drawn the histogram's way, the colour slightly see-through inside and full at the edge. Slate and Gravity are unchanged (compared pixel for pixel in Chromium, light and dark). The manual describes the hours by their legend, Less to More, where it said the darker the square, the more played, and its Stats figure is retaken.

Fixed

  • [site] The grid of the black panel beside Install in one gesture shows. Its lines were white at 4%: a lightness gap of 2.7 from their black ground, where the hero's grid shows 9.1 — next to nothing on screen. They now stand as far from their ground as the hero's do, in the light theme as in the dark one (9.3 and 5.9 against 9.1 and 5.7, measured in Chromium).
  • [ui][manual] The app's messages can be closed, and stay until they are read. A message — the toast that says how an action went — has an × now. An error or a warning stays until it is closed, where it left after four to eight seconds, whether or not its words had been read; a success or an information still leaves after a few seconds, but not while the pointer, a finger or the keyboard's focus rests on it. Three show at most, the one shown or said again longest ago giving way, and a message already on screen is not shown a second time. An action that takes a while — starting playback, saving the pipeline's topology, applying a service's systemd settings, checking or updating the audio software — no longer stacks its result under its "in progress" message: that message turns into the result, and two plays started together each keep their own. A screen reader is told each message once, with its type, two that come together included. The messages are a card on the page's ground, their type in a coloured left edge and icon; in Minimal the information is in black, where it was blue. The manual describes them.
  • [ui] The messages no longer cover the app's controls. In the app installed on an iPhone, a message lay over the top bar, and its menu, Library and settings buttons did not answer while it showed (measured in Chromium with an emulated notch); on a computer three messages hid the Library and Manual tabs, every tab in a narrow window, and on a phone the offline banner's RETRY. They now start below the top bar, the tab bar and the offline banner. On a phone they are centred, where they sat 6px from the left edge and 18px from the right, and their × answers a finger over 44px, still drawn at 28.
  • [ui] A favourite that cannot be saved says so, and why. The message was a blue box headed "ERROR", with nothing else: its call passed the words where the message's type goes. It reads Favorite not added — or not removed — with the reason. Every call that shows a message is now checked for that order.
  • [ui] The messages' words say what happened. About seventy are reworded: a title that gives the outcome — Topology saved, Settings not applied, Passkey not added — and a sentence that says what follows from it, where many repeated their title ("Configuration Saved — Configuration updated successfully") or showed an error as the core wrote it. A cancelled passkey and an empty log to copy are information, no longer warnings. The Update All confirmation says 1 package where it said 1 packages.
  • [core][ui][manual][site] The radio search keeps the stations that really stream lossless, and its box says so: Lossless only. Its Hi-Res only kept a station listed as FLAC or WAV, or announced at 256 kbps or more: an MP3 at 320 kbps passed, while the FLAC stations the Radio Browser catalogue lists as OGG — the wrapper they are sent in, which also carries Vorbis and Opus — passed only on the bitrate they announce. Ticked, Lossless only keeps the stations listed as FLAC or WAV, and reads the first 4 KB of each station listed as OGG, UNKNOWN or with no codec, to keep those that send FLAC; on the lab's box a search for "flac" took under ten seconds the first time, and was answered at once the next. A station read that way shows FLAC and its resolution — 24bit / 48kHz — in the search, My Live Radio and Favorites, and while it plays the player and the signal path show FLAC and its bit depth where they showed OGG, as lossy. The box reads four streams at a time at most, follows a station's redirections checking each address, never reads an address on the local network for a station taken from the catalogue — a station added by hand may be on it —, and tries a station that failed again a day later only. The API's lossless_only replaces hi_res_only, kept as a deprecated alias; a station gains verified_codec, verified_bits and verified_rate_hz. The manual, the editions page and the landing's radio card say so.
  • [ui][core][manual] A Raspberry Pi short of power shows its processor Slowed down, not at a speed it is not running at. While a Pi's supply falls short, its processor runs slower than the speed Linux gives — 600 MHz on a Pi 4 that Linux still showed at 1.8 GHz, measured on the lab's Pi —, and the Performance tab showed Linux's figure. While the THROTTLED badge shows on a Pi, the core's frequency and Current Freq above the cards now read Slowed down, and the support report says slowed down (undervoltage) in place of a figure — its audio_live gains cpu_slowed. An Intel processor's figure is the speed it measured, and stays, even while it is slowed for heat.
  • [ui][manual] The network test's verdict reads in every theme. EXCELLENT, GOOD, FAIR or CRITICAL was a solid block of colour with the page's ground as its words: under the 4.5:1 floor in 11 of the 24 combinations of theme, mode and grade, down to 1.94:1, and in Minimal GOOD was a black block that looked like a button. It is drawn as the app's other state badges are — a tint, and the word in its colour —, at 4.5:1 or more everywhere. The manual's figures of the network and latency tests are retaken.
  • [ui][manual] The performance tests' charts follow the theme. A UDP network test's jitter was a smoothed line in a blue no theme sets — it read a colour no theme declares —, which drew values between the seconds it never measured: it is one bar per second now, in the theme's network colour, as on the System tab. In Minimal, the latency histogram draws its bars under 50 µs in grey and up to 200 µs in black, and keeps colour for the delays from 200 µs. Both charts are drawn again when the theme or the light/dark mode changes: they kept the colours of the mode they were drawn in. The manual gives the histogram's bands and says the jitter is drawn second by second.
  • [core][ui][manual] The iperf3 network test takes a server's port. The field asks for an address and its port, and iperf3 read the whole as a host name — "unable to connect to server … Name or service not known", measured with iperf3 3.18 —: with a port, the test stopped at once. The port is passed to iperf3 on its own, 5201 when none is given; an IPv6 address takes brackets, [fe80::1]:5201; and an entry that names no server — empty, or with a port that is not a number or outside 1–65535 — is refused before the test starts (422), and no longer kept among the field's suggestions. The UPnP media server search, which split an address the same way, writes an IPv6 host in brackets in the addresses it probes.
  • [ui] The Pipeline writes a link's latency and buffer fill in colours that read. On a link, in its bubble and in the legend, the latency took the latency histogram's bar colours: 2.02 to 3.76:1 on its pill in every light theme. It takes its band's text colour now — in Minimal grey under 50 µs and ink up to 200 µs, as the latency test's histogram, then orange and red —, and the buffer fill beside it, ink at rest in Minimal, likewise: 5.07:1 or more in every theme and mode. The legend gives the bands in their colours instead of naming them — "Green <50µs · Blue …" was false in Minimal —, and the thresholds — the latency's, written in five places, legends included, and the buffer's, in three — are kept in one.
  • [site] The social preview and the README's image show the landing as it is today. The card networks and messengers show for audiogravity.app, and the image heading the site's README, still showed the public beta: its ribbon, an older capture of the app, the previous carousel controls — the README's even the opening line the landing dropped. Both are taken anew from the page by gen-og-image.js: the release candidate's ribbon, the HQPlayer screen the carousel opens on, with its annotations.
  • [site] Halloween — the spider beside the phone hangs from its thread in Safari. Safari drew the spider and not its thread, a pseudo-element sized with container units; each thread is now drawn in its spider's own box, its length in pixels.
  • [lic] A licence key no box holds is no longer written in full in the server's log. Created by the admin or by a PayPal payment, deactivated or restored, such a key is a bearer secret — the first box to send it keeps it; those lines now write only its first group (AG-XXXX…), and so does uvicorn's access log, which wrote it in the address of the admin's routes on an order.
  • [ops] nginx — the licence server's access log redacts licence keys, in the path as in the query (AG-XXXX-[key_redacted]), as it already dropped credentials.
  • [site] The README says what the app does today. Its opening line is the landing's; the editions table puts the box's own music, the radio and the player in Starter, where it put them in Pro; Tidal reads Hi-Res up to 24-bit / 192 kHz, no longer a "HiFi" plan Tidal has dropped; the interface's installer flags and --version are listed, and an omitted --vapid-email gives Audiogravity's own push address, not a placeholder with a warning; installing the app starts with trusting the box's certificate, and no longer says Chrome on iOS cannot; contributions follow the interface's own policy — no pull requests yet, issues on this site —; and the link to the core's repository, private, which a visitor could not open, is gone.
  • [ui] The README describes the repository as it is: the test count no longer reads 188 — the badge gives it —, npm test runs the Storybook tests in a browser as well, the signal-path diagram is the one component in a shadow root, and ESLint, npm run lint and ./dev.sh start --https are listed.
  • [ops] releases repository — the README's download check works. Its command fetched audiogravity-ui-*.tar.gz, a name curl does not expand, and got a 404; it now reads the file name from SHA256SUMS, which covers the three tarballs — not every file, as it said. It also gives the system required and the address the installer prints.
  • [manual] The login figure no longer shows the BETA badge. The app writes its stage in its corner — BETA up to 0.9.80, 1.0 RC now, nothing once 1.0 is out —, and the figure of the login screen still showed BETA. The badge is erased from it, the rest of the figure left as it was.
  • [ui][manual] Days and months are written in English whatever the phone's language, and dates read the same everywhere. On a phone set to French the Stats view wrote its week as lun. to dim., its bars as 28 sept. and a square as lun. 10:00, beside its English words; a configuration card's date and the performance tests' histories did the same, and the service and profile cards and the history panel wrote American English (Oct 27, 05:45 PM). One function writes them now, in British English — Mon 12 Oct, 27 Oct, 17:45, September as Sept —; a time keeps the phone's clock (17:45, 05:45 PM) and a date in digits its conventions. A moment from another year carries its year (27 Oct 2025, 17:45), a missing one reads -- where it read Invalid Date, and a configuration file saved at 23:50 reads yesterday at 00:10, where it read today. The manual's Stats and configuration figures are retaken.
  • [ui] The fullscreen player stays in place when a station's inset cover is tapped. Tapping the small cover in the corner of a station's artwork — the station's logo or the track's cover, to swap the two — threw the whole player off the screen and slid it back up, on an iPhone as in Chromium on a touch screen; a new track's cover did the same after a touch. Each new background tint rewrote the player's whole style and dropped what the touch had set, so its opening slide-in played again. The tint is now set on its own, and the slide-in is taken off once it has played: with a mouse, Close now slides the player down, where it stayed for a third of a second, then vanished. A drag the system interrupts — a call, Control Center — now puts the player back in place; it stayed where the finger had left it.

[1.0.0-rc.1] - 2026-10-25

Added

  • [site] Halloween on the landing, every 31 October. From midnight to midnight by the visitor's own clock, the page is dressed from top to bottom, never over a line of text. Bats fly across the hero and across the band above each section's label; on a screen wide enough to leave margins beside the text, more bats swoop through them, and in them spiders come down from each section's top on their threads, then climb back — on a phone, a spider hangs in each band instead. Webs hang in two corners of the hero. Jack-o'-lanterns stand at the foot of the phone, with a mist in the dark theme, and on the footer's top edge, under which a last spider hangs; their carved faces are lit from inside by a candle that flickers, each pumpkin at its own pace. On a screen too short to show the foot of the phone, its pumpkins rest on the bottom of the screen until it comes into view. That day the strip that shows a track playing plays Saint-Saëns' Danse macabre, and the tab's title starts with a pumpkin. The page opens dark unless the visitor chose a theme with the button, and nothing is stored. On any other day nothing of it is fetched; ?halloween in the address shows it on any day, to check it beforehand. Nothing moves in a part of the page scrolled out of sight, and for a visitor whose system asks for less motion only the pumpkins, their candles still, and the webs remain.
  • [site] Each guide says how long it takes to read, and the guides are numbered. Under a guide's title, a line gives its number and its reading time — Guide 10 · 5 min read —, and the list of guides ends each entry with its time: the words a reader sees, at 238 words a minute, the mean silent reading speed for non-fiction in English (Brysbaert, 2019), rounded up — 3 to 8 minutes today. The sidebar, the phone's menu and the list number the guides by their place in the list; a guide's address carries no number, so a guide added in the middle renumbers the list without breaking a link.

Changed

  • [site] The hero's carousel shows the app as it is today, in the page's theme. Seventeen screens captured anew, each in the app's light and dark themes; the page shows the capture of the theme it is in — the system's, or the one chosen with the button — and switches with it. Four screens are new: listening statistics, the menu, a service's network activity, and the audio software with its updates. The screen of Qobuz's new releases with a station playing is gone, for want of a new capture. Every annotation beside the phone is measured again on the new captures, and the captures weigh 1.9 MB where the originals weighed 32.9. The passkey login's capture, further down the page, is new too and follows the theme the same way.
  • [site] The hero carousel's controls are a pill of dots, and the current one fills over its screen's time. A dot per screen in a pill, beside a round play button: the dot of the screen shown is drawn long and fills from left to right over the 3.5 seconds before the next screen, a pause holds it where it is, and pressing a dot shows its screen. The arrows are gone. The dots are buttons now, reached with the keyboard and named for a screen reader, and the play button says what a press will do. The panel's grid, painted over the dots and the buttons until now, lies under them, and so does the phone's shadow, which greyed the controls in the light theme.
  • [site] The guides' sidebar gives their titles room. At 240px a guide's title — a whole question — broke onto up to four lines, at the browser's tightest line spacing, and the list read as one block. The column now takes what its titles need, up to 320px, and the lines of a title are spaced a little, less than two titles: three lines at most, the guide's text keeping its width (Chromium, 1440 × 900). The manual's chapters keep their 240px column; their numbers now sit on their title's line, where they sat a little high.
  • [site] The guides no longer carry the trademark notice: it speaks for the app, which the manual documents and a guide does not. The manual's chapters keep it.
  • [ui] The corner badge names the release stage: "1.0 RC". The purple badge in the bottom right corner of the app read BETA; it now reads the stage from the version itself — "1.0 RC" for this release candidate, BETA for a 0.x release, none for a final release or a development build —, where a build setting kept by hand on one machine decided it.
  • [site] The landing says "Release candidate". The corner ribbon, the announce bar, the footer and the README name the stage; on a phone the ribbon reads "1.0 RC" and keeps the old ribbon's place, the long words covering the end of the links and leaving the theme toggle no room beside the logo at 320px. On a wider screen the announce bar's sentence, the theme toggle and the GitHub button keep their distance from the longer ribbon; the GitHub button leaves the bar from 1160px instead of 1130px, where it found no room beside it.
  • [manual] The FAQ answers what "Release candidate" means, in place of "Public beta": the version proposed as 1.0, complete, taking only corrections until then. The landing's FAQ, generated from it, follows.

Fixed

  • [lic] A box on a release candidate is offered its final release. The update offer and the automatic publication compared the numbers alone, so 1.0.0-rc.1 and 1.0.0 were one version: the final release would never have been offered to a box on its candidate, nor published by itself. They now order versions as SemVer does — a candidate before its release, rc.2 before rc.10 —, and a development build keeps its offers as before.
  • [ops] self-update — the health check tells a release candidate from its final release. After an update it checks that the core answers with the version asked for; on the numbers alone, a release candidate still answering after an update to its final release passed for a success, and the update was not rolled back. The check that judges an update is the one the installed version shipped, so it leaves with the release candidate.
  • [ops] release — verify promotes a final release over its release candidate. It refused to move the latest release backwards with sort -V, which ranks 1.0.0-rc.1 after 1.0.0; it now orders versions as the licence server does, and stops when it cannot compare them.
  • [ui] Update banner — a failed attempt at a release candidate is no longer shown against the offer of its final release.
  • [lic] The admin's device table sorts versions in the order updates are offered: a release candidate before its release, rc.2 before rc.10. Reading every digit of the version put a release candidate after its own release.
  • [ops] release — the README version badges follow a release candidate. sync-version.mjs skipped every version with a dash, which left the ui, lic and core badges on the release before; update-test-report.sh wrote the landing's with a single dash, which a shields.io address reads as a field separator. Both write a pre-release with the dash doubled (1.0.0--rc.1), and "beta" after a 0.x release only.
  • [site] The pipeline drawing's moving dots no longer keep the whole page busy. They kept the page redrawing sixty times a second wherever the reader was, the drawing on screen or not — on a modest phone, more than half of what the page asked of it; they now move only while the drawing is on screen, and stay still for a visitor whose system asks for less motion, which they used to ignore.

[0.9.80] - 2026-10-24

Added

  • [core] Network renderer — the player shows the track's format. FORMAT, SAMPLE and BITRATE stayed empty for whatever a network renderer played, and CODEC for anything but a DSD file: a renderer reports none of them. They are now filled as Audiogravity knows them — the header of the bytes the box relays to the renderer (Qobuz, HIGHRESAUDIO), Tidal's stream description, the file's own header for a library or media-server file (FLAC, WAV, AIFF, DSF, MP4), a station's catalogue bitrate —, read once per track in the background, at no request of its own for a streaming service. A value nothing tells shows a dash.

Changed

  • [manual] Listening → casting to a network speaker: the hi-fi readout shows what Audiogravity knows of the track sent, where it said the readout was absent; a cell nothing fills shows a dash.
  • [ui] API.md documents format.codec: none when nothing tells it, read a moment after an .m4a/.mp4 starts (a further state event), and filled for a network renderer, with its format, sample_rate and bitrate.

Fixed

  • [core] HIGHRESAUDIO tracks show their codec. The CODEC cell stayed empty: MPD plays the box's relay, whose address names no format. The relay now records the codec of the file the service serves.
  • [core] Qobuz shows the codec it serves, read in the address Qobuz resolves for the track: the relay's address named none, and every track read FLAC — an MP3 included, on a box set to MP3.
  • [core] An .m4a or .mp4 file shows ALAC or AAC, as its header says. Both were shown as ALAC: an AAC — an iTunes Store purchase — passed for lossless. Read once per track, from the box's disk or the media server.
  • [core] Roon — the CODEC cell reads "—" when Roon names no codec, no longer "Roon".
  • [core] A file whose codec is not known yet is no longer shown as "PCM" — an MP4 the moment its header is read, or an extension Audiogravity does not know.
  • [core] DSD played on a network renderer no longer brings the box's DAC to full volume. The DSD protection took the one card playing on the box — AirPlay's, say — and set it to 100 % while the DSD went out to the speaker.
  • [core] A library file cast to a network renderer is recognised by its extension, which a signature follows in its address: a DSD file there was not seen as DSD.

[0.9.79] - 2026-10-23

Added

  • [core+ui] Push notifications — three new alerts. A new release of Audiogravity is announced once per version, from the daily check of the licence server, and its notification opens the Admin tab, where the update banner installs it; only an answer of that server withdraws it — a server that does not answer, at a restart before the network is up, withdraws nothing. The system disk almost full — the Disk Usage tile's figure, read every ten minutes while a device is subscribed — is announced at 90 % and again at 95 %, and once more only after it went back under 85 %. A Raspberry Pi whose power supply lets the voltage drop below the safe level is announced at most once a day: the core listens to the Pi's own under-voltage flag, through the lookup the Performance tab's THROTTLED badge already uses, rather than reading it on a timer. Measured on the lab's Pi 4, whose supply drops 65 times a day: every drop and every return seen within the second of the kernel's message, one notification for all of them, 0.003 s of CPU over 100 s of waiting. A box no device subscribed from runs none of the three checks, and a machine without the flag never has the third.
  • [core+ui] Push notifications — turning Notifications on sends a test notification, and says why when they cannot be turned on. The box sends that device, and that device only, a notification: the proof that the whole way works. When it does not go — notifications blocked for the site, an app not opened over HTTPS or, on an iPhone or iPad, not installed on the Home Screen, a box without its notification keys, a subscription the box refuses, a push service that refuses it or cannot be reached —, a message names the reason, the subscription is undone on both sides and the switch goes back off. The switch used to stay on after a failure, with nothing said. An app newer than its core, which has no test yet, turns notifications on as before and says no test could be sent. Measured in Chromium against the dev core: subscription, test accepted by Google's push service, Notifications on; without keys, The box cannot send notifications and the switch back off. POST /push/test sends it, with the API key.
  • [ops] Passkeys work on the box's name, set up by the installer. Passkeys (Face ID / Touch ID) were left off on a box installed without --public-url, and every page said they needed a real domain. They need a name, never an address — measured on an iPhone: on https://musics.local, with the box's own certificate trusted and no domain bought, a passkey registered and signed in, twice. The interface installer now gives the core that origin — the name it prints and puts in its QR code, with its port — when it runs on the same box in HTTPS, the certificate carries the name and the core has none; a box whose name another device already holds is left alone, its name can change. The origin it set is marked as its own and follows the box's port or name at the next install; one given with --public-url, or by hand, is never touched.
  • [manual] Administration → Push notifications lists the seven alerts, says Profile Activated goes to the other devices, and explains the test notification and the messages shown when notifications cannot be turned on.

Changed

  • [manual] A real domain is optional: notifications and passkeys work on the box's name. Installation, First run, Administration, Troubleshooting, the FAQ, the glossary and the quick start said push notifications and passkeys needed a real HTTPS domain and could not work over a bare IP. On an iPhone or iPad they work on the box's name, its own certificate trusted — measured for both, the notifications arriving with the app closed. Getting HTTPS — for passkeys and push becomes A real domain — optional: what a domain adds is the app, passkeys included, from another network than the box's, and no certificate to trust on each device. Also put right: add the app from the box's name only from the box's network — from another Wi-Fi or VLAN the name does not open, so add it from the address and reserve that address in the router; iOS 16.4 or later for notifications; turn them on from the installed app; the FAQ no longer says only Safari installs the app; --vapid-email and --public-url say what they do. A new Troubleshooting entry, Notifications never reach an iPhone, iPad or Mac, says to update the box.
  • [site] Landing — the install section no longer says passkeys and push need a public HTTPS domain: it says to trust the box's certificate once on each phone, offers plain HTTP for a box behind a reverse proxy rather than as a way round the certificate — plain HTTP gives no app, no notifications and no passkeys —, and presents a domain of one's own as optional. The FAQ is synced from the manual again, the Starter Edition answer included.
  • [site] Landing — the push notification card and the app section list every alert: disk almost full, a Raspberry Pi's power supply too weak and a new version of Audiogravity join service down, critical temperature and audio software updates; the app section names Profile Activated too, which goes to the other devices.
  • [ops] sync-landing-faq.py writes the landing's links in --warn-text, the colour the landing's rework gave them for contrast: run, it put back --warn in every FAQ link.
  • [core] POST /push/subscribe refuses an endpoint over 2048 characters (422), the bound the test notification and a profile activation share; push services hand out a few hundred.

Fixed

  • [core] Push notifications — deactivating a profile no longer sends Profile Activated. A deactivation runs through the activation with a profile that starts nothing, which reads as active once done: every subscribed device was told the profile had just been activated — reproduced with the real profiles code over stand-in services, one activation and one deactivation giving two notifications.
  • [core+ui] Push notifications — Profile Activated is no longer sent to the device that made the change, only to the other subscribed devices: the app names its own push endpoint with the activation (push_endpoint of POST /profiles/{profile_id}/activate).
  • [core] Push notifications — an update announced is no longer announced again after the box restarts. What had gone out was kept in memory, and the core restarts at every update of Audiogravity and every reboot: an audio software update left uninstalled was announced anew each time. What reached at least one device is now kept on the box, beside its subscriptions, and is forgotten only once the update is installed or withdrawn — never on a pass that could not read the installed versions. A notification no device received is not recorded, and is tried again.
  • [core+ops] Push notifications reach iPhones, iPads and Macs. Every box the installer set up signed its notifications with mailto:[email protected], a contact Apple's push service refuses: measured against Apple with an invented device, a .local or localhost contact gets 403 BadJwtToken where a real address is accepted — so an Apple device never received a single notification. --vapid-email could not help: it wrote into .env, while the core signs with the copy in vapid_keys.json, which the installer always filled with the placeholder. The installer now writes the contact given, else [email protected], where the core reads it; an update replaces the old placeholder — keys kept, so subscribed devices stay subscribed — and keeps a contact set by hand; the core itself replaces a contact that cannot work, and says so in its journal. Measured on a copy of an ARM box's own keys: refused before, accepted after. A key file without a contact no longer leaves push off.
  • [core] Push notifications reach Android and Apple devices of the same box. The library that sends them writes the push service's address into the signature it is handed, and the core handed it the same one every time: after a notification to Google, every one to Apple carried Google's address, and Apple refused it (403, measured) — a box with both kinds of device reached only one. Each push now gets a signature of its own.
  • [ops] Installer — --public-url is applied on an update too. It was read on a first install only: a box given a domain later kept its first passkey origin.
  • [core] Manual install guide (docs/INSTALL.md) writes the push contact into vapid_keys.json, and says passkeys need a name, not a real domain.
  • [core] Push notifications — a box without notification keys says so. GET /push/vapid-public-key answered 500 there instead of its 503 — the dev core of the lab's x86 box —, and the app could only report that the box did not answer.

[0.9.78] - 2026-10-22

Added

  • [site] Four new guides: What your USB DAC accepts on Linux, and how to read it — where Linux lists the formats, rates and DSD a USB DAC takes, why the list can stop at 96 kHz, and which settings copied from forums the kernel now ignores; Software or hardware volume on a Linux streamer — what a player's, a DAC's and an amplifier's volume each do to the music, MPD's settings, its slider in decibels (50 % is −18.5 dB), and DSD; AirPlay to a Raspberry Pi: what quality actually arrives — CD quality in classic AirPlay, lossy AAC in AirPlay 2 from an iPhone, iPad or Mac, and a Debian 13 receiver that plays classic AirPlay only; Your streamer on the network — the bandwidth music takes, Wi-Fi or cable for a streamer and an NAA, and how AirPlay, UPnP and NAA devices find each other across VLANs. Like the others, each gives the steps on Raspberry Pi OS and Debian and the messages as they show, says what it was checked with and when, and lists its sources, with a short Audiogravity section at the end.

Changed

  • [site] Four guides rewritten for every reader, Audiogravity user or not: Your NAS music on a Raspberry Pi streamer, DSD on a Linux streamer, Qobuz, Tidal and HIGHRESAUDIO on a Raspberry Pi and HQPlayer NAA on a Raspberry Pi. Each now gives the steps on Raspberry Pi OS and Debian — the commands, how to check, the messages as they show —, says what it was checked with and when, and lists its sources; Audiogravity keeps a short section at the end, a tenth of the text or so where it was up to half. A guide that explains how to do something no longer walks through the other products; the streaming guide, whose subject is the choice, compares the ways to play the three services on a Pi — upmpdcli, Volumio, moOde, Lyrion, HQPlayer, Roon — with what each costs and allows. The NAS guide's commands and ten error messages were measured against a Samba and an NFS server, and so was a trap it now steers around: an MPD update run while the share is not mounted empties its library, without a message, where a share mounted when its folder is first opened is refused and the library kept. The NAA guide gives Signalyst's rule — an NAA 6 needs HQPlayer 5.16 or later — and how DietPi installs the newest NAA whatever the HQPlayer. Statements found wrong on the way are put right: Linux recognises native DSD from a list of DACs and of makers, not by a generic detection; HIGHRESAUDIO names Volumio's own players, not Volumio.
  • [site] Five more guides rewritten the same way: How to check bit-perfect playback, HQPlayer Embedded on a Raspberry Pi 5, “Device or resource busy”, Roon on a Raspberry Pi and Real-time priority and reserved CPU cores. The HQPlayer Embedded guide installs Signalyst's package step by step and says what to set by hand after it — the web page's password, the output, the log, the half-hour trial; the Roon guide installs Roon Bridge with Roon's own script, or on DietPi; the real-time guide shows how to read a player's priority, pin it to a core and reserve that core on a Pi; the busy guide gives MPD's message as measured. Audiogravity's sections come down to about a tenth of each text, where they took from a quarter to almost two fifths. Put right on the way: a 2023 remark by HQPlayer's author on the Pi 5, made the day the board was announced, was read as a measured limit — his 2025 findings on its USB output replace it; of the ten arm64 versions Signalyst lists for Debian 13, only three can be downloaded, and version 5 cannot install there, with apt's message given.
  • [manual] Troubleshooting → Manual NAS mount mounts the share when its folder is first opened, with nofail, the account written in a file that never passes through the shell's history, and a check that says why a mount failed. The share was mounted at start-up: with the NAS off then, its folder stayed empty, and a library scan could empty MPD's index. It no longer says that kernels disable SMB1 by default: Debian 13's still accepts vers=1.0.

Fixed

  • [core+ui] Audio Software — a vendor's site that does not answer no longer turns software into Not Supported. On 21 October the site of upmpdcli's author answered HTTP 503 to every client during an emergency maintenance at its host, for over a day. The boxes updated meanwhile recorded UPnP Bridge as published for no machine at all, and its card read Not Supported while it was installed and playing; on ARM, AG also stopped offering its own build of it, which that site does not serve. A check that gets no answer now keeps the vendor's last answer for the box — same address, same Debian release, same machine — for up to five days, and asks again every day; on ARM, AG's own build of UPnP Bridge is used whenever its author offers none for ARM, whether his site says so or says nothing, and his site is still asked every day, for the day it does. The card no longer calls a check that could not be made Not Supported, and a list of machines left empty reads Unknown. Measured on an x86 and an ARM box of the lab, the author's site still down: the x86 one keeps UPnP Bridge available for amd64 and i386 from an answer of 6 October, and the ARM one goes back to AG's own build from the configuration it has installed. A box with no answer to keep — the lab's x86 one, updated during the outage — reads as unchecked until the site is back, no longer as unsupported. In GET /packages/, unknown now means the source could not be reached and gave no answer for this box in the last five days.
  • [core+ui] Support report — a package that cannot be installed is given its real reason. It read not supported on this architecture whatever stood in the way: the lab's x86 box had UPnP Bridge (its author's site down) and Roon Server (Roon Bridge installed) in that case, both supported there. It now gives the reason the software card shows — cannot be checked: repository answered HTTP 503, unavailable: Roon Bridge is installed — the two cannot run together, not available here: … —, and keeps not supported on this architecture for the one case it describes. The report's packages entries give availability and availability_reason, as GET /packages/ does.
  • [site] On a phone, the manual's and the guides' list of pages is a menu. Below 860px it was a strip scrolled sideways: fine for the manual's one-word chapters, unusable for the guides, whose titles are whole questions — on a 375px phone not one title showed whole, and the guide being read sat far off to the right. An All guides or All chapters line now opens the list, every title on its own line and the page being read marked; closed, the page starts on the first screen. On a wider screen the column stays as it was. The Guides and User manual labels in the top bar now lead to their contents, which no guide page linked back to.
  • [site] On a phone, Pricing and Install in the landing's bar bring their section to the top of the screen at the first touch. The page stopped short, the section showing near the bottom of the screen, until a second touch: the app's sign-in screenshot further up took its room only once it loaded, during the scroll. It now has its size from the start; nothing else on the page moves.

[0.9.77] - 2026-10-21

Added

  • [core+ui] Library → Stats shows covers. Most played and Recently played show each row's cover — an album's, a track's, a station's logo — where they showed a music note, which stays when none is known. An artist shows the cover of the album of theirs listened to most over the whole history, the same whatever the period. A cover is kept as its address, never as an image, and read when Stats shows it: of the addresses an album's or a track's listens gave, the surest — the picture where it is kept, before a search by name, before a file AirPlay deletes once the track is over —, the latest among them; a track with none of its own takes its album's. What was recorded before keeps the note until the same track or album is heard again. Measured on ten years of synthetic listens: the history file grows from 33.0 MB to 39.7, and 12 months takes 104 ms to read instead of 101 on x86, 256 instead of 248 on ARM. GET /listening/summary gives top_artists[], top_albums[], top_tracks[] and recent[] a cover_token, for GET /audio_pipeline/cover.
  • [manual] Library & streaming → Stats says each row shows its cover, an artist the cover of the album of theirs played most, a station its logo.

Fixed

  • [core] Library → Stats no longer counts a radio station under a word that names no genre. A station's tags in the radio directory name its codec, its country or its owner as readily as its style: France Inter's are aac, general, public radio, radio france, and its listens were counted under a genre called Aac. A station — and anything played live — now counts only under a genre of AG's common list, from its tags as from what its stream announces, and under Unknown otherwise; a file still keeps its own word when it names a genre no service uses. The listens already recorded are put right at the box's next start, and so is what an older version records should the box go back to it: on the lab's box, Aac (1 h 41) and Eclectic Programming (29 min) moved to Unknown, the time listened unchanged.
  • [manual] Library & streaming → Stats says a radio station counts as Unknown when its description names no genre of the list.

[0.9.76] - 2026-10-20

Fixed

  • [core+ui] Library → Stats counts what is playing. A listen was recorded once it ended, and Stats counted only what was recorded: a station left on all evening showed nowhere until it stopped, and Stats read Nothing recorded yet while it played. What plays now counts as it stands when Stats is read — once it has played 30 seconds — in every figure and in the count of listens, and comes first in Recently played, marked IN PROGRESS where the others show when they began, with its time so far; a track on pause keeps its time, without the mark. Nothing more is written for it, and it costs nothing while the music plays. While Stats is on screen it is read again when what plays changes, once a track just started counts, and when the app comes back to the foreground — from the player stream the mini player already holds open. The history's reads and writes now run one at a time, in the order asked, so a track ending as Stats opens is counted once, never left out or counted twice. A history file the core can no longer write (a card gone read-only) shows what it holds instead of an error. GET /listening/summary gives recent[] its in_progress; listens of GET /listening/settings counts the listens in progress.
  • [core] A device that no longer takes notifications is forgotten at its first refusal. One that had unsubscribed — the app removed from a home screen, its notifications turned off — was sent every notification again: the core never read the status of the push service's refusal. It is now forgotten as soon as the service answers that the subscription is gone — Apple's 410 Unregistered, or an expired subscription's 404 —, and kept on any other refusal. On the lab's box, 7 of its 10 subscriptions had been refused 133 times over 19 notifications since 10 October, none forgotten; they go at the next notification.
  • [manual] Library & streaming says what is playing counts in Stats, marked IN PROGRESS, and Troubleshooting (A listen is missing from Stats) no longer says a listen shows only once it ends.
  • [core] HQPlayer — every internet radio station sent to HQPlayer goes through the box's own relay, on a port of its own (8010). Fetched by HQPlayer itself, Radio France's MP3 stations stopped after 30 to 40 seconds; served by the core's web server, the same stations played one evening and failed five times out of five the next morning (HQPlayer Desktop 5.28.1). The relay answers in HTTP/1.0 — the end of the body is the end of the connection —, the shape a separate test program held in. A station in MP3 or FLAC is passed on as it is sent; the others — AAC, or a format the catalogue does not name — are converted to FLAC as they arrive, as AAC stations already were. A station whose server answers the old SHOUTcast way (ICY 200 OK, which the core's HTTP client refuses) goes through ffmpeg, its frames copied, unconverted. A connection HQPlayer has stopped reading for a minute is closed, and what it held with it. Measured: France Inter and FIP, in MP3 and converted from AAC, held all their plays and resumes after a pause — 13 and 13 on HQPlayer Desktop, 6 and 6 on HQPlayer Embedded 6.2.6. A station already in HQPlayer's list when the box is updated is to be sent again.
  • [core] HQPlayer — Play on a paused station picks up the live broadcast. A bare Play resumed it where it was paused, then stopped once HQPlayer's buffer ran dry; AG now selects the station's entry again, then plays — the sound back a few seconds after the press —, and reports a refused selection instead of resuming the past.
  • [core] HQPlayer — Previous and Next are greyed where they lead nowhere: Previous on the first entry, Next on the last, both on a station played alone. HQPlayer answered either with an error and stopped the music. can_next / can_prev of the player state follow the entry HQPlayer plays.
  • [core] HQPlayer — a station or an album started after HQPlayer sat paused on a station for a while now plays. The Play that followed the new list was ignored after a pause of two minutes; the first entry is now selected before it.
  • [core] HQPlayer — what AG sends to HQPlayer says whether it is a live broadcast (live of the player state, true for a station), as for a network renderer; what HQPlayer plays from its own remote stays unknown (null).
  • [ops] installer — the core's settings carry HQPLAYER_RELAY_PORT=8010, added to an existing box's at its update, never over a port changed by hand; another program on that port makes the relay take a free one and say which in the journal. The interface's installer refuses that port when the core runs on the same box.
  • [manual] Installation → the interface installer's --port refuses 8010 too, where the core serves HQPlayer its radio stations.
  • [core] HQPlayer — the player's signal path names the HQPlayer that plays, and reaches the DAC. With the chain description a new box arrives with, an HQPlayer on another computer read hqplayer — the internal name — with no DAC, and the box's own HQPlayer Embedded had no DAC either; with a description of one's own, the step took whatever name it gave another machine's HQPlayer (on the lab's box, HQPlayer Embedded while the Mac's HQPlayer Desktop played). The step is now named as HQPlayer calls itself — HQPlayer Desktop, HQPlayer Embedded, or HQPlayer until it has said — asked once when it answers a status read, never while the music plays, nor again of one that went silent. Another computer's HQPlayer goes through the box's NAA to the DAC (HQPlayer Desktop → HQPlayer NAA → DAC), the box's own straight to it; one whose sound the box does not play is shown alone, with no DAC. A service of the box that the description leaves out — HQPlayer Embedded or the NAA, on the one delivered — is placed on the box, when the description has a single streamer, and reaches the DAC as MPD does.
  • [manual] Outputs & engines → Use as output names the HQPlayer in the signal path: Library → HQPlayer Desktop → HQPlayer NAA → your DAC.

[0.9.75] - 2026-10-19

Added

  • [ops] installer — the interface's installer takes --http, --https and --port. They choose how the interface is served and win over an existing install, so re-running the installer with one of them is how a box changes mode or port; --http or --https alone means that mode's own port (8080, 443), and nothing is asked. Without them a re-run, or a self-update, keeps the mode and port in place. The all-in-one installer passes them on, and both hosted installers check the port before downloading anything. An installer from before them refuses them instead of ignoring them.
  • [ops] installer — the interface's installer ends on the box's name, and its QR code carries it. http(s)://<name>.local comes first, the address below it should the name not open: without a fixed address, the name is the only way back to the box, and an app added to a home screen keeps for ever the address it was added from. The name is the one avahi announces, once settled — a box renamed because another device holds its name (audiogravity-2.local) shows the one it answers to, with a warning in yellow to give the box a hostname of its own and run the installer again, since which device keeps the plain name can change at a restart; in HTTPS it is offered only when the certificate carries it. The address is repeated under the QR code. The addresses to open stand out in green, and what blocks when skipped — trusting the certificate, its step 3, the limits of plain HTTP — in yellow.

Fixed

  • [site][manual] The install instructions open the streamer by its name, and offer plain HTTP. The landing's install block and Installation → all-in-one sent the reader to https://<ip-of-your-streamer>; they now point at the address the installer prints, https://<name>.local, and the landing gives the --http install for those with no certificate to set up on their phone.
  • [ops] installer — the core installer announces admin / admin123 only when it has just created that account. Run over an existing install it keeps the accounts, yet it ended on "Default login: admin / admin123", so an owner who had changed the password was refused without a word. It now says the accounts were kept, and gives the two commands that bring admin / admin123 back; the password itself cannot be shown, the box keeps only its hash.
  • [ops] installer — the interface's certificate names the .local name the box answers to. It was built from the hostname, so a box renamed by avahi because another device holds its name got a certificate for that other device's name, and https://<name>-2.local was refused. It now carries both, compared exactly; the interface's start never waits for avahi, only the installer does, a few seconds, and each question to avahi is bounded to one second.
  • [ops] installer — in plain HTTP, the summary no longer says the home-screen icon receives notifications. A browser grants notifications and the offline start to HTTPS only; the summary now says to open the interface through the reverse proxy's https:// address for those.
  • [manual] Installation, First run and Troubleshooting say which .local name to use: the one the installer prints first. A box that takes -2 because another device already answers to its name is to be given a hostname of its own — Troubleshooting says how.
  • [ops] installer — the interface's installer asks again when an answer is not one it can use. "2" typed for the mode and landing on the port question gave an interface on port 2: a port is now 80, 443, or 1024 to 65535 — not the core's port on the same box, nor in HTTPS the one that hands out the certificate (8081) — and the mode 1 or 2.
  • [ops] installer — the core installer shows the API key once. A new install printed it in two boxes in a row.
  • [manual] Installation → Getting HTTPS switches the interface to plain HTTP with --http. It said to re-run the ui installer and pick plain HTTP, but an installed interface was never asked again. The flags have their own table.
  • [ops] uninstaller — after an uninstall without --purge, it names what it kept and the command that removes it. It ended on "Run with --purge to remove them", but it lives in the directory it has just removed: there was nothing left to run.
  • [manual] Installation → Uninstalling runs --purge in place of the plain uninstall, not after it. It gives the command that removes what a plain uninstall kept — now naming the processor setting of the Performance page among them — and says that the interface keeps the box's certificate authority, and why.

[0.9.74] - 2026-10-19

Added

  • [core+ui] Library → Stats: what the box played. A sixth tab in Library shows the time listened over the last 7 or 30 days, 12 months or the whole history — one bar per day or month, tap one to read it —, where it came from and went to, its sound quality (DSD, Hi-Res, CD quality, lossy), its genres, the most played artists, albums and tracks, and the latest listens. A listen counts from 30 seconds actually played, pauses left out; a station is one listen for as long as you stay on it, its genre taken from its tags. The box records on every edition, from what its player already announces — the player reads nothing more for it, with or without a screen open (measured: 8 reads in two minutes of playback, none following a screen); the figures come with Pro. Roon and HQPlayer are not counted. A listen whose source vanishes from the player for a moment, or that plays across a restart of the core, stays one listen. The history is one file on the box, readable by the core's account only, never sent anywhere; ten years of listens open in 75 ms, 12 months in 280 ms at most, on the ARM box (200,000 synthetic listens, measured on a 1.8 GHz board, the week's hours included). New routes: GET /listening/summary (Pro), GET/PUT /listening/settings and DELETE /listening/history (every edition, refused to a guest account).

  • [core+ui] Library → Stats shows when the box plays. When you listen maps the week: a row per day, Monday to Sunday, a square per hour, the darker the more the box played at that hour, on average per week over the period chosen; tap a square to read its time. A listen counts in the hours it actually played — one started at 20:50 for half an hour fills 20:00 and 21:00, a pause fills none — on the box's clock, including the nights the clocks change and in the countries half an hour off UTC. A track replayed after a pause is dated by its own start. It costs nothing while the music plays; opening the view takes 14 ms more on the x86 box and 35 ms more on the ARM box, at most, with ten years of listens. GET /listening/summary gains week_hours.

  • [core] Streaming tracks have a genre, from one list shared by every source. Qobuz, Tidal and HIGHRESAUDIO each file music under genres of their own — the same style as Soul/Funk/R&B, R&B/Soul or R&B — and their tracks reached Library → Stats as Unknown. The box now counts and shows them under one list of 23 genres, at the level of the genre, never of the sub-genre: Pop, Rock, Pop/Rock, Alternative, Metal, Electronic, Dance, Hip-Hop, Soul & R&B, Jazz, Blues, Country, Folk, Classical, Soundtrack, World, Latin, Reggae, French, Easy Listening, Children, Spoken Word, Religious. Qobuz tracks and HIGHRESAUDIO album tracks carry their album's genre from the moment they are queued; a Tidal track, or a track of an HIGHRESAUDIO playlist, has it asked of its service once it starts playing — one request per track, in the background, answered in 0.17 s for Tidal and 0.30 s for HIGHRESAUDIO (medians over 40 tracks and 15 albums). The track playing shows its genre in its details, HIGHRESAUDIO's album genre rather than the one its file names. A file's or a station's genre tag is read against the same list, each genre it lists in order: Pop & Rock and Pop/Rock are one genre, and a name no source uses is kept as written. Listens recorded before keep their genre. LibraryTrack gains genre; sources[].genre of the player state carries the common genre for a streaming track.

  • [ui] Recording can be switched off and the history cleared on every edition. At the bottom of Library → Stats, Keep a history of what plays and Clear history… (asked first) work on a Starter box too, where the figures stay behind the padlock.

  • [ops] installer — the listening history is enabled on an existing box at its next update.

  • [manual] Library & streaming describes Stats: the periods, the charts, what counts as a listen, switching the recording off and clearing it. The editions table, Troubleshooting (A listen is missing from Stats) and the backup list (listening.db) follow.

  • [manual] Library & streaming says the genres in Stats follow one list whatever the service, with an example, and that files and radio stations are read against it.

  • [manual] A figure of Stats, and the tab menu with its Stats tab. Library & streaming shows the Stats view over 30 days — the time listened, a bar per day, the week's hours — with an example history, set in the capturing browser only. First run's two figures of the tab menu, Pro and Starter, and its list of Library's views now include Stats.

  • [site] The landing has a Pro card for the listening history, beside the self-update card, now half width.

  • [site] Search engines are told where the site is and what it holds. The site had no sitemap, the landing named no address of its own while www.audiogravity.app served the same page, its search snippet ran past what a result shows, and Bing returned nothing for the domain (measured 2026-10-17). The landing now names https://audiogravity.app/ as its address, and www.audiogravity.app redirects there; its snippet says what Audiogravity is in 145 characters; structured data gives the site's name and what the product is, with no price. A sitemap.xml lists the landing and every page of the manual, and robots.txt points to it. The sitemap is written with the manual's pages, from the same Contents list, so a new chapter reaches it in the same run, at the address its page gives.

  • [ops] release — the manual's regeneration commits the sitemap with it, so the landing is not left modified for release.sh ui to refuse.

  • [site] Guides, with their own tab in the landing's bar. Practical answers to questions asked on audiophile forums, each checked against the app and dated, the other products' facts sourced: HQPlayer Embedded on a Raspberry Pi 5, Qobuz, Tidal and HIGHRESAUDIO on a Raspberry Pi, HQPlayer NAA on a Raspberry Pi, How to check bit-perfect playback on a Linux streamer and Audiogravity, Volumio or moOde: which to choose?. They are written in Markdown, listed in docs/guides/README.md, and rendered by the manual's generator on the manual's page shell, each with its own address for search engines and its entry in the sitemap. With one more link, two labels of the bar broke onto a second line from 1101 to 1118 px and from 851 to 934 px wide: the bar's two thresholds move to 1130 and 950 px, and no label breaks from 320 to 1400 px (Chromium).

  • [ops] release — the manual's regeneration commits the guides with it.

  • [site] Six more guides: A DAC HAT on a Raspberry Pi: why it is not detected, and the line that fixes it, Your NAS music on a Raspberry Pi streamer: mount the share, or use a UPnP server, Roon on a Raspberry Pi: Roon Bridge, and Roon through HQPlayer, DSD on a Linux streamer: native, DoP, or converted to PCM, Real-time priority and reserved CPU cores for audio on Linux and “Device or resource busy”: two players, one DAC. What they say of Audiogravity is checked in the app and measured on the lab's boxes where it could be — the DAC held and released, DSD as the DAC receives it, wake-up delays on an isolated core and on a shared one —, and what they say of other products cites its sources, dated. The guides' list now follows the order of building a streamer. How to check bit-perfect playback on a Linux streamer no longer says that the guided setup turns DoP off: until this version, it did not.

  • [core+ui] A DoP switch in MPD's guided setup. Config → MPD → Guided has a DSD field: Send DSD as DoP (DSD over PCM), off by default, for a DAC that decodes DoP but that Linux cannot send DSD to as it is. It is set on the output MPD plays on — in that output's block of MPD's file alone, its old dsd_usb alias taken out with it — and MPD takes it without a restart, from the next time it opens the DAC — restarted instead when the switch moves MPD's DSD conversion before its volume (below). The switch shows what MPD applies, or what its file says while a Roon or HQPlayer profile has stopped it; it waits while another output is chosen and not yet applied. Reset to default keeps it on the same output, and leaves it off for another. An output with no DoP setting (not a sound card) is refused before anything is written.

  • [core+ui] The player says how a DSD track reaches the DAC. The format shown is the track's: a DSD128 file read DSD128 even when the DAC received PCM. A couple of seconds into a DSD track played by the box's MPD, the mini player and the full-screen format strip add a DoP tag, or Converted to PCM, when it does not reach the DAC as DSD — read once per track from the sound card and MPD's DoP setting, again when the track starts over.

  • [manual] Outputs & engines explains the three ways DSD reaches the DAC — as DSD, as DoP, converted to PCM — and the DoP switch: when to turn it on, when to leave it off, a HAT, DSD256 too fast for DoP. Listening, First run, Administration and the glossary (DoP) follow.

  • [core+ui] Full DAC volume during DSD, a switch per DAC. Config → MPD → Guided has a second switch in its DSD field, under DoP's: Full DAC volume during DSD, on by default. On, the DAC MPD plays on has its own volume brought to full while it receives DSD — as DSD or as DoP, from any player of the box, HQPlayer included — and given back when the DSD ends. Off, for a DAC that lowers DSD properly, its volume is left as set. The setting belongs to the DAC, not to MPD: it is kept in audio-topology.json, found again by the DAC's USB id, else by its name, and it applies at once, a DSD track playing included. Like DoP's, the switch waits while another output is chosen and not yet applied; Reset to default leaves it as it is. New route: POST /audio-stack/dsd-full-volume; GET /audio-stack/status gives the mpd entry its dsd_full_volume.

  • [manual] Outputs & engines says what happens to your DAC's volume during DSD, and how to turn it off, with a figure of MPD's DSD field. First run's figure of the Guided view is retaken with that field; First run, Administration, Introduction, Listening and the glossary (DSD) follow.

  • [core+ui] HQPlayer's volume on its card. In Library → Sources, the HQPlayer card shows HQPlayer's volume in dB, with a slider that sets it — for the box's own HQPlayer as for one on the network, whether it plays through the box or not; it sat in the DSP menu, closed by default. The slider covers the range set in HQPlayer, from its Volume min to its Volume max, and the card follows a change made in HQPlayer or by Roon every 5 seconds while it is on screen, asking nothing while it is out of sight. A volume fixed in HQPlayer's settings shows its level and no slider; a volume that cannot be read shows a dash, where it read 0.0 dB. GET /hqplayer/status gives volume_min_db, volume_max_db and volume_fixed.

  • [manual] Outputs & engines describes HQPlayer's volume on its card, and its two figures of the card are retaken.

  • [core+ui] The volume works on DSD on a box whose DAC does not take DSD. A HAT on a Raspberry Pi, a computer's built-in outputs or a USB DAC Linux sends no DSD to receive DSD converted to PCM, and MPD converted it in the last filter of its output, after its volume: the volume did nothing to DSD, and the player showed a padlock. Where none of MPD's outputs takes DSD as it is and DoP is off, the guided setup now writes audio_output_format "*:32:*" in MPD's file: DSD becomes PCM at the decoder, before the volume, and the player keeps its slider. Measured on a HiFiBerry DAC+ Pro on a Raspberry Pi 4 (MPD 0.24.4): DSD64, DSD128 and DSD256 reach the card as 32-bit PCM at 352.8 kHz, MPD taking 3.9, 37.4 and 68.7 % of one core, against 3.0, 36.6 and 65.1 % with the conversion in its output; PCM goes out in 32-bit words, its values unchanged at 100 %. Whether a card takes DSD is read without opening it — a USB DAC's formats as Linux lists them; a computer's built-in sound, and a Raspberry Pi's HAT on its I2S, jack and HDMI, take none —, and a card that is not known, or not there, changes nothing. The line follows the outputs: written or taken out when the guided setup writes MPD's file, when the DoP switch moves — MPD then restarts, since it reads that line only when it starts — and when a box with one output is given another card. Config → MPD → Guided says so under DSD. GET /audio-stack/status gives the mpd entry its volume_on_dsd; the player state and its sources[] entries carry volume_on_dsd for a DSD track.

  • [ops] installer — boxes already set up get it at their update, and a box given a DAC that takes DSD since loses it. Only in a configuration the guided setup wrote, once the DoP alias is taken out; the previous file is kept among the configuration editor's saved copies, and MPD is restarted. A USB DAC switched off during the update leaves the file as it is.

  • [manual] Listening, Outputs & engines and the glossary (DSD) say the volume works on DSD on a box whose DAC does not take DSD, and that turning on DoP there restarts MPD.

Changed

  • [ui] Stats says when recording is off. The switch sat at the bottom of the view, under the rankings: with it off, Stats simply stopped growing and nothing said why. A line at the top now says Recording is off: nothing new is recorded, with a button that takes you to the switch. Its words are in the text colour on the warning tint — the warning's own text colour read 4.35:1 there in Slate light; 12:1 to 19:1 now, measured in the six palettes.
  • [ui] In the tab menu, a Library tab sent to a new line starts it at the left. With six tabs, the sixth took the menu's whole width alone under the five others. The page's own Library bar is unchanged: it scrolls.
  • [ui] A switch names what it turns on to screen readers, and a filter bar says which option is chosen. The switch's checkbox had no name of its own; the filter buttons said their choice by colour only (aria-pressed).
  • [core] During DSD, the player's volume is left alone, and the DAC's own volume is what goes to full. Until now a DSD track set MPD's volume to 100 — and that of every other source of the box — then the DAC's, and put them back when it ended. Measured with a DSD128 file on MPD 0.24.5: MPD's volume changes nothing to DSD, sent as DSD, as DoP or converted to PCM, so it now stays where it was; a DAC's own volume does — a USB DAC lowered by 10 dB played DoP as silence and native DSD as noise. The DAC goes to full only while it receives DSD: when the DSD reaches it converted to PCM, its volume is given back as soon as the player shows Converted to PCM, and left alone the next time the same player sends it DSD at that rate — or from the first second, when MPD's configuration converts DSD before its volume. Lowered while DSD plays, paused included, it goes back to full at once: while it holds the DAC, the box listens to the sound card's events rather than reading its volume every second — which it still does when it cannot listen — and it stops listening when the DSD ends.
  • [core] The DAC set to full is the one the DSD plays on. It was the first sound card playing, whatever held it, and the box's first card when none was playing yet. With several DACs playing, it is now the one held by the player that sends the DSD; none is touched when that cannot be told.
  • [ui] The DSD padlock tells screen readers that the volume has no effect on DSD, and to set the level on the amplifier — it said the volume was locked at 100 %.
  • [site] The editions page says DSD protection, and the landing's comparison DSD volume protection — they read DSD bit-perfect and Per-source bit-perfect lock (DSD volume protection); the other products' cells are unchanged.
  • [site] The DSD and bit-perfect guides say what the volume does to DSD, with the measurements above, and how to turn off the DAC's full volume.
  • [site] The DSD and bit-perfect guides say when MPD converts DSD before its volume, with the HiFiBerry's measurements, and that MPD then works in 32-bit words: S32_LE on a card that takes them, the values unchanged at 100 %.
  • [core] The player's volume covers the range set in HQPlayer. While HQPlayer plays through the box, the player's slider ran from −60 to 0 dB whatever HQPlayer's settings: on an HQPlayer whose Volume max is −3 dB, the top of the slider did nothing and −6 dB read 90. It now runs over HQPlayer's own range, read with its status on the same connection, and over −60 to 0 dB for an HQPlayer that does not say one; it is not shown while HQPlayer's volume is fixed.
  • [core] PUT /hqplayer/volume keeps the volume within HQPlayer's range, and answers the volume it sent. It refused anything outside −60 to 0 dB, which a range set in HQPlayer may pass; it now brings a value past either end back to it, and to −60..0 dB while HQPlayer has not said its range.
  • [ui] The HQPlayer card shows the output mode and rate beside the name as soon as it opens. They appeared once the DSP menu had been opened.
  • [core] shairport-sync's settings are read with libconf instead of pylibconfig2. The values read are the same — compared on the guided setup's file, on two boxes' files and on shairport-sync's own sample with every setting switched on — and they now keep the file's order. A 64-bit value written with its L, such as airplay_device_id, no longer makes the file unreadable. A setting named twice is still refused, as shairport-sync refuses it, and so is a line that includes another file, which is never opened. The third-party notices list libconf.

Fixed

  • [core] Saving AirPlay's settings in Structured mode writes a file shairport-sync can read. A value holding a quote, a backslash or a line break, a true/false setting and a list were written in forms that libconfig — the library shairport-sync reads its file with — refuses; they are now written in its syntax.
  • [ui] Config → Guided preselects the right output of a card that has several. The jack and HDMI of one card share its name: the output chosen for a service was matched without its device number, so the other one could be shown as chosen. The first-time setup matched the same way.
  • [manual] Listening no longer says the format readout shows exactly what reaches the DAC: it shows the track's format, and for DSD how it reaches the DAC. The glossary no longer says the DSD lock shows for native DSD only, nor does the padlock's description for screen readers.
  • [core] DSD goes out as DSD when the DAC accepts it, and as PCM otherwise — no longer as DoP. The guided setup wrote dsd_usb "yes" beside dop "no", meant as native DSD; MPD still reads dsd_usb as the old name of dop, so DSD left wrapped in PCM (DoP). Measured with a DSD128 file on MPD 0.24.5: the DAC received PCM at 352.8 kHz with that line, and native DSD at 176.4 kHz without it. A DAC that cannot decode DoP received PCM full of markers, with the volume held at 100. The setup no longer writes the line.
  • [ops] installer — boxes already set up lose that line at their update. Only in a configuration the guided setup wrote, and only where the line sits beside dop "no": an output that asks for DoP itself, and a file written by hand, are left as they are. The previous file is kept among the configuration editor's saved copies, where it can be restored, and MPD is restarted so that it reads the corrected one.
  • [core] Closing a subscription the event bus had dropped as too slow no longer raises.
  • [site] The landing says what Starter and Pro each bring. The configuration card counted Roon among what Starter edits — Roon's settings come with Pro, as HQPlayer's do. The library card said the music on the box was only browsable on Starter: Starter browses, searches and queues it, and casts it to a renderer. The FAQ now says that listening to a streaming service, and connecting one, come with Pro, and the introduction to the features names what Pro adds: the streaming services, the media servers, Roon and HQPlayer.
  • [manual] The manual says what comes with Pro wherever it sends you there. Roon's settings come with Pro, alongside HQPlayer Embedded's and the NAA's (Introduction, First run, Administration, and the editions page). Troubleshooting marks the tools it suggests from the Systemd and Performance tabs as Pro, and Administration says the 30-day trial is the time to tune the box — the Systemd overrides, the disk and network counting among them, stay in force once it ends. The FAQ says connecting a streaming service comes with Pro, Listening describes the padlock a streaming track wears in the queue on Starter, and Outputs & engines marks the Pipeline tab as Pro where it first describes it.
  • [site] The landing's comparison table matches each vendor's own pages, checked on 17 October 2026. The streaming row no longer says "official APIs": Audiogravity is a partner of none of the three services. Roon offers Tidal and Qobuz, not HIGHRESAUDIO; HQPlayer streams Qobuz and HIGHRESAUDIO, and JPLAY all three; moOde offers Qobuz only, through upmpdcli, and has not browsed UPnP media servers since 2022; Volumio's Premium needs an account; HQPlayer runs on Windows, macOS and Linux and costs €319 for version 6, not €265; JPLAY runs on Apple-silicon Macs too; Volumio's monthly price is added.
  • [site] The guides describe the box as Audiogravity installs it. Settings made by hand on the test box — a shorter audio buffer, processor cores reserved at boot, its own music share — no longer read as if they were the reader's: the buffer is the one the configuration Audiogravity writes leaves to MPD, and a measurement says where it was taken and what it does not represent. The HAT guide now says to edit the two audio lines Raspberry Pi OS already has, rather than add two new ones beside them.
  • [site] A guide's title carrying the brand reads in one piece in the sidebar. Audiogravity, Volumio or moOde: which to choose? was split into three pieces set side by side like columns; every title of the manual's and the guides' sidebars is now one block.
  • [site] The landing no longer promises CPU cores reserved for the player. Audiogravity pins each audio service to the cores you choose for it; it does not keep the other programs off them. The card and its Dedicated cores line now say so: "on the cores you choose for it".
  • [manual] The manual no longer says the guided setup's configuration does no volume math. First run said the configurations send audio to the DAC "untouched — no resampling, no volume math in the digital domain", and the glossary that bit-perfect means "no software volume": the volume Audiogravity sets is MPD's software volume, which recalculates every sample below 100. Both now say MPD's configuration is bit-perfect with the volume at 100 — a 16-bit file padded to 24 or 32 bits, its values unchanged — except on a DAC that accepts 16 bits only, and that AirPlay is not bit-perfect: it adds or drops a sample now and then to stay in time with the sender. Quick start, Introduction and Outputs & engines say bit-perfect with the volume at 100.
  • [site] The landing's configuration card and the README say the same: bit-perfect for MPD at full volume. The README's Bit-perfect lock with DSD volume protection now reads DSD protection — your DAC's own volume at full while it receives DSD, from any player.
  • [site] The bit-perfect guide says what MPD's volume does towards a DAC that accepts 16 bits only. MPD widens 16-bit audio to 24 bits for its volume, then brings it back to 16 bits with dither for such a DAC: measured on MPD 0.24.5 (2026-10-18), half the samples of a 3-second test came out changed, by up to 3 steps, at 100 % too.
  • [ui] The HQPlayer card no longer keeps what it read from the HQPlayer it showed before. When the box's own HQPlayer started or stopped, the card showed the other instance with the previous one's filter, shaper and mode lists, and a filter is sent as a position in the list: one picked there could reach the new HQPlayer as another filter (the Mac's Desktop 5 and the x86 box's Embedded 6.1 name different filters at 54 of the 70 positions they share).
  • [core] A filter, shaper, mode or volume changed on the HQPlayer card shows at once: the status read back right after could be the one kept from before the change, for up to 2 seconds.
  • [site] The third-party notices list what the core's binary carries besides its Python packages. The Python interpreter, compiled with the core, whose licence asks for its text to go with it; OpenSSL, libffi, Expat, Zstandard, libuuid, bzip2 and XZ's liblzma, copied from the machine the binary is built on; the runtime of Nuitka, the compiler, under AGPL-3.0 with the exception that lets a program compiled with it be distributed under its author's terms; and SQLite, which the binary carries from this version for the listening history. Each comes with its copyright and licence, read from the official sources of the versions the binary carries.

[0.9.73] - 2026-10-16

Added

  • [core+ui] On a Starter box, the queue shows which tracks will not play. The streaming services' tracks left in the queue — Qobuz, Tidal, HIGHRESAUDIO, queued during the trial — wear the Pro tint and a padlock in place of their length, in Library → Queue (the track playing included) and on the full-screen player's Next, a network speaker's too; a tap on the padlock opens the licence panel, and the track can still be removed. The core marks each queued track from its address — a station saved with a streaming relay's address is marked as well: pro on /library/queue, on the player state's queue_next and on a renderer's status (queue_next_pro).
  • [core+ui] Play says why when only streaming tracks are left in the queue. On a Starter box, Play on such a queue did nothing and said nothing: the box's player skipped every track and stopped, a network speaker replayed a refused one. It is now refused with Only streaming tracks are left in the queue…, shown by the full-screen player; the box's queue is judged whole, a speaker's from the track it is on. A queue holding anything else plays as before, the streaming tracks skipped. Measured on MPD 0.24.5, and on a UPnP renderer (upmpdcli) driven by the box.

Changed

  • [core+ui] The free Starter edition plays the music on the box and the internet radio. Starter opened the box's administration only, and playing anything needed Pro. It now has Library's local library — browse, search, queue — and Radio, the choice of output — the box's DAC or a network speaker —, the mini-player's buttons and the full-screen player with its sleep timer. Pro adds the streaming services (Qobuz, Tidal, HIGHRESAUDIO), the UPnP media servers, Roon and HQPlayer — as the output, its settings, HQPlayer Embedded's and the NAA's output and file — and keeps the Pipeline, systemd tuning and Performance. On a Starter box what Pro adds stays in view behind a padlock — a tab, a Library source and its chip in Search, a Config tile, the outputs of a Roon or HQPlayer source — and a tap takes you to the licence panel, with the activation window open.
  • [core] The box refuses what Pro opens to a box without it, whoever asks. The padlocks were in the interface alone: a request that skipped it got through, and Library, the players and HQPlayer refused nothing. The core now answers 403 to the Pro routes and to a Pro source, whether the request names it with or without its src_ prefix; to a player command that reaches Roon or HQPlayer, judged on the device it reaches; to the address of a streaming service's track handed over as a file, a radio station or a speaker's track; to HQPlayer's and Roon's output, and to HQPlayer Embedded's and the NAA's file. The relays of the streaming services refuse too: a track queued while the box had Pro no longer plays once it has not — on the box's player, on a network speaker, or sent by another app — and the box's player skips it.
  • [core] A Starter box never sends anything to HQPlayer. HQPlayer chosen as the output stays chosen for when the box has Pro again; meanwhile the music plays on the box's own player. While the box's own HQPlayer Embedded runs and holds the DAC, a play from Library, from Radio or from the player's Play button is refused with what to do: stop HQPlayer Embedded from Services, or activate a license.
  • [ui] What Pro locks wears one tint, from the theme's brand colour. The locked tabs wore a rosy tint of the error colour, all but invisible on a near-black bar (1.07:1 in Minimal and Gravity dark) — and red marks an error. The locked tabs, the Pro sources' cards and chips, the outputs of a Pro source and Config's HQPlayer tiles now take the theme's brand colour at 8 % in light, 16 % in dark: 1.10 to 1.47:1 against their ground, measured in Chromium in the three themes. In Search, a chip's padlock no longer touches its name.
  • [core+ui] A play refused for want of Pro is said where it was asked. The full-screen player's Play button kept such a refusal silent, as it does a device that declines, and the Radio screen showed one neutral sentence for every failure. Both now show the box's own words — HQPlayer Embedded holding the DAC, a station that plays a streaming service's relay, a queue left with streaming tracks only — which the core answers as every refusal for want of Pro, with 403 and a mention of the licence, so the padlocks follow. A device that declines stays as it was: the button simply does not move. Only the Play button shows them: a volume drag sends one command per movement.
  • [manual] Three figures show what a Starter box sees. The tab menu with its badge and the padlocks of Systemd, Performance and Pipeline (First run); the Library's sources, the Pro ones behind their padlocks (Introduction, under the editions); the licence panel once the trial has ended (Administration, between the trial and an activated licence). Taken on 2026-10-15 from the live interface, at the same iPhone 13 mini size as the others.

Fixed

  • [core+ui] A box reporting tampered no longer shows the Pro tabs open. The interface kept its own list of the statuses that open Pro, and it disagreed with the core's: a tampered box showed every Pro tab open while the core refused them all. The core now publishes the verdict — pro on /license/status, on each source, on the player state and on each Config service — and the interface reads it.
  • [core] A licence with no end date is no longer checked against the clock. The check meant for a trial and a licence with an end date ran for a perpetual one too, and its verdict is kept an hour: a box starting far off the right time — a Raspberry Pi has no battery-backed clock — could lock its paying owner out of Pro.
  • [core] The Pro gate no longer reads the licence file at each request. Every Pro request — each page of a streaming service, every HQPlayer call — re-read the file and re-checked its signature. The verdict is now kept 30 seconds and forgotten at once when a licence is activated, imported or removed; a refusal still reads the status itself, so a box that has just become Pro is never refused.
  • [ui] The interface reads the licence again when the box comes back, and when the box refuses something for want of Pro. It read it once, when it opened: a reading that failed then — the box restarting — left a box with Pro without its Pro sources for the whole session, and a trial that ended while the app stayed open showed no padlock. It now reads it again whenever the connection to the box comes back — the app back from the background, the box restarted — and at each refusal for want of Pro.
  • [ui] A Starter box is no longer told to buy a licence "to continue". Its licence panel said Purchase a lifetime license to continue, as if the box stopped with the trial — it carries on in the Starter edition, for free. It now says A lifetime license brings the Pro features back.
  • [manual] The manual says what comes with Pro wherever it describes it. Library & streaming never mentioned Pro: the media servers, Roon, the streaming services and the albums Search finds in them now say they come with Pro, as do HQPlayer Embedded, the quick start and the introduction's overview. Troubleshooting's streaming entry opens with the end of the trial and the message the player shows, and the FAQ says what happens to the queue — and that what comes with Pro is using Roon and HQPlayer from the app, not installing or starting them.
  • [manual] Troubleshooting names the message a Starter box shows when HQPlayer Embedded holds the DAC. The list of messages under No sound / wrong output did not have it: the box is on Starter, HQPlayer comes with Pro, and the entry says to stop HQPlayer Embedded from Services or to activate a licence.
  • [manual] The box receives AirPlay; it does not send to it. The introduction, the quick start and Listening listed AirPlay among the places music can be sent to: an iPhone, an iPad or a Mac streams to the box.
  • [site] The landing promises Starter only what Starter does. The unified transport, internet radio and UPnP control point cards are Starter, with Roon and HQPlayer marked Pro in the cards and in the FAQ. The Auto-discovery card named AirPlay receivers, MPD instances and Roon Bridges, none of which the box looks for; it now names what the box finds — the UPnP media servers, HQPlayer, the Roon Core and its zones.
  • [core] The box reads MPD's long answers to their end. It asked MPD to close the connection along with each question, and MPD hung up before it had sent a long answer: some 16 KiB came back. Library → Queue showed 282 tracks of a queue of 1,000; on a library of 3,000 albums the album list showed 245 of them, an artist sent to a network speaker 38 of its 96 tracks, and a search lost its last results. The box now reads each answer up to MPD's final OK, then closes the connection itself. Measured on MPD 0.24.5.
  • [core] Building the album list no longer holds MPD. The box sent all of the album list's questions at once, in lists of 50, and MPD answers a list to its end before anyone else: on a library of 3,000 albums, the box's other questions to MPD went unanswered for 5 seconds, and with 6,000 albums the album list failed. The lists now go one after the other — MPD answers in between (181 ms at worst with 3,000 albums, 23 ms with 474), the list takes the same time, and 6,000 albums are listed in 9.8 seconds. Measured on MPD 0.24.5, x86.

[0.9.72] - 2026-10-15

Added

  • [site][manual] Every page of the online manual has the light / dark toggle. The manual's pages on audiogravity.app followed the theme chosen on the landing, else the system's, and offered no way to change it: a reader who arrived on a chapter from a search was left with their system's. Each page's top bar now ends with the landing's toggle — the moon for dark, the sun back to light — and one choice holds across the site, switched on the landing or in the manual. Without a choice, a page still follows the system. The landing and the manual share one small script for it; in a private window it still switches, for as long as the page is open. On a 320 px phone the bar's gaps narrow so that User manual stays on one line. Tried in Chromium from 320 to 1440 px; not in Safari.

Changed

  • [ui] Every badge you can press has the small button size, wherever it sits. CONFIG (Pipeline), NEW USER (Admin), an Audio Software tile's password and Restart required pills, and a Services tile's ENABLED / DISABLED were 18 to 20 px high in 10 px type, beside 24 px buttons in 11 px — a list of places held the others to the size, and missed them. One rule now sizes every pressable badge, a button or a clickable tag. An Audio Software tile's footer wraps on a phone instead of running past the card, which it already did by 36 px at 320 px. The password pill reads Set pwd; a screen reader hears "Set pwd: the web interface password", the words shown first, so a voice command naming them finds it.
  • [ui] CRITICAL sits at the right edge of a Services tile, under the state it qualifies (UP, IDLE…), apart from the buttons. On a narrow phone, or a computer showing three columns of tiles, it takes its own line, at the right.
  • [manual] The HQPlayer Embedded section names the card's Set pwd.

Fixed

  • [ui] One device can no longer freeze the interface for all the others, and a refused certificate shows in the journal. On a box installed in HTTPS — the installer's default — the interface's server negotiated each connection's encryption in the one loop that accepts them all: a device that opened a connection and stayed silent kept every other device waiting for as long as it stayed open, and with no limit if it dropped off the network without closing (measured: still frozen after 60 s). Each connection now negotiates on its own, within 30 seconds, then has 30 more to send its request; past that it is let go, and the journal says so. A device that refuses the box's certificate, or an address typed with http:// instead of https://, now leaves a line with its address and the reason, where it left none. Nothing changes once a request is in: live updates, the terminal and large files are not cut. A box installed in plain HTTP also lets go of a connection that sends nothing for 30 seconds. Measured on x86 and ARM (Python 3.13) with a silent and a vanished client; the refusals read from Python, curl and Chromium — not tried in Safari.
  • [ui] A screen reader hears the main bar as a navigation, not as tabs. The bar declared itself a list of tabs and held more than tabs — the wordmark, the user, Manual and, in the phone's column, the library's Queue, Sources and Radio — so a screen reader counted those among the app's tabs. It is now a navigation named Main navigation: its entries are buttons, the page shown is marked as the current one, and the library's views under Library form a group of their own, Library views. With a keyboard, Tab now reaches every entry and Enter opens it, as on any button; the arrow keys no longer move from page to page. Nothing changes to a finger or a mouse. Checked in Chromium's accessibility tree; not tried with a screen reader.
  • [site][manual] A page no longer meets a stylesheet or a script from before an update. The host lets a browser keep the site's stylesheets and scripts four hours, and its pages not at all: after an update, a visitor of the past four hours got the new page with the old files — the manual's new light / dark button sat against the bar's label in Safari until a forced reload. Every page now asks for each of its stylesheets and scripts with a stamp of the file's content (style.css?v=881223e621): a changed file has a new address, which no cache holds, and Cloudflare fetches it from the source; one unchanged stays cached. The manual's generator stamps its pages, and its workflow now also runs when a stylesheet or a script changes; the landing and the 404 page are stamped by scripts/asset_versions.py, and their tests refuse a stamp out of date. Fonts and images are left as they are.
  • [ui] After signing in again, the app opens the page and tab you were on. A session that ended — expired after its 12 hours, or ended by the box — sent you to sign in, and back to the first tab: the sign-in page read the destination where nothing wrote it, and an expired session erased it before it was read. You now come back to the page and its tab, query included — only ever to a page of the box itself, read by the browser's own address parser; a sign-out forgets it, and a sign-in page that finds you signed in already in another tab takes it too.
  • [ui] An address no longer opens a tab your role cannot see. /#admin opened the Admin tab for a guest — hidden, its calls refused — and coming back after signing in could take an admin turned guest there. Such an address now opens Profiles. The guard meant to move a former admin off the Admin tab called a function that did not exist, and threw; it works.
  • [ui] The audio pipeline diagram's buttons react to a finger. On a tablet or a phone held sideways, zoom, FIT, RESET, LEGEND, MINIMAP and NETWORK showed nothing under the finger, unlike every other button: the diagram draws in a part of the page the stylesheets do not reach. They now shrink and fade while pressed, from the same file as every other button.
  • [ui] A paused station's Live marker is greyed like the buttons beside it. In the full-screen player, a paused station's Live took the secondary text grey, a shade darker than the greyed Repeat, Previous, Next and Shuffle just above it. It now takes theirs — the text colour at 30 % opacity — in every theme, light and dark.
  • [ui] Chart bars stand out from their ground in Slate and Gravity, the faded ones too. On System, Performance and an expanded Services chart, every bar but the latest is faded so the current one stands out. In the light Slate and Gravity themes the faded bars read 1.5 to 2.1:1 on their ground, and the amber (temperature) and green (network) ones stayed under 3:1 even at full strength — the floor for a chart's marks (WCAG 1.4.11). The light palettes take a darker shade of the same colours, and past bars fade to 80 % in light and 70 % in dark instead of 55 %: measured in Chromium, 3.5:1 at the least in light, 3.4:1 in dark. Minimal, already above, is unchanged. The latest bar stands out less; the figure above the chart gives the current value.
  • [ui] The full-screen player's sleep timer keeps its clock while it counts. Once set, the timer replaced its clock with the time left, and a bare 14:52 beside the queue button read as the time of day. The clock now stays, with the time left beside it.
  • [core+ui] Library › Sources lists AirPlay and the UPnP bridge only when they are installed, and shows one switched off as Off. Both inputs were listed on every box, Idle whether their software was missing, stopped or waiting for a stream. An input whose software is not installed is now left out; one installed but stopped reads Off, with an empty ring in place of the dot and its icon and name faded as an unavailable tile is; one running, or starting, still reads Idle. The screen follows the box live, without being reopened: measured on the x86 box, a stop or a start shows within 2 s, an install or a removal from Audio Software in the same second. Installed means systemd has the unit's file — or its template's, for an instance unit — the test the profiles now share.
  • [core+ui] A webradio started from another app is shown as live. Played on the box's MPD from a phone's MPD client or a UPnP app, at an address the radio module does not know, it had Previous and Next greyed but no Live marker: the full-screen player drew a progress bar stuck at 0:00, and Repeat and Shuffle stayed on though they act on nothing there. The interface decided by itself, from the source and the length; it now reads the box's answer — the rule Play already followed to reopen a paused station: a station the radio module knows, or a stream that names itself and has no length. Measured on MPD 0.24.5, an unknown webradio names itself about 0.6 s after it starts and a file never does, so a track whose length is not known yet never reads as live. A station AG cast to a network renderer is live whatever length the renderer reports, and the playback bar no longer draws a progress bar for it. For HQPlayer, AirPlay, Roon and a stream another app cast to a renderer, the box says it cannot tell and the interface keeps its own reading.

[0.9.71] - 2026-10-13

Changed

  • [ui] In the Minimal theme, what you type into or press has an outline you can see. Fields, lists, framed buttons, selectors, filters and switches were drawn by the cards' hairline, #D0D0D0 — 1.45:1 on a panel, where accessibility guidance (WCAG 1.4.11) asks 3:1 for what tells a control apart. They now take a darker grey of their own, at least 3:1 on every ground of the theme: #858585 in light, #707070 in dark. Cards keep their hairline, the sign-in page is unchanged, and so are the Slate and Gravity themes.
  • [ui] In the Minimal theme, colour marks a state, never an action or a figure at rest. Some twenty green or orange marks showed at once on Services alone, where red, orange and green mean something only while they stay rare. START and ACTIVATE, a service's figure at rest, the running share of a health bar, the stripe and CRITICAL tag of a service the chain needs, the stripe of an installed package and the actions written as badges (REFRESH, CHECK UPDATES, NEW USER, CONFIG) now draw in the theme's text colours; UP, ENABLED, failures, warnings and what deletes keep theirs. The Slate and Gravity themes are unchanged.
  • [manual] Twenty-seven figures show the Minimal theme's outlines and colours. Retaken on 2026-10-13 from the live interface, at the same iPhone 13 mini size and with the same music as before: fields and buttons outlined in Minimal's darker grey, and actions and tags drawn in black (START, CONFIG, TEST, REFRESH, CHECK UPDATES) or grey (CRITICAL). They cover Audio Software, Services, Config, the latency and network tests, the guided editor, the network share form, the audio stack's set-up, Settings, a user card, the system actions, the CPU cards, Qobuz browsing, the radio, HQPlayer Embedded's card, install window and output, the five licence figures, the full-screen player, the busy output, a cast to a network renderer, the mixed queue and the signal chain. Two sentences no longer give a colour Minimal does not show: a profile's health bar is no longer described as green while running, nor a critical profile's tile as edged in orange. The Listening chapter says which transport buttons are greyed, and the Administration chapter that a service's charts unfold in bars, two for network and disk.
  • [ui] A service's expanded charts are drawn in bars, in the theme's chart colours. As the System tiles are: one bar per measurement, the newest one full, the highest measurement written above. Writes were drawn in red and egress in orange, as if they were errors; each chart now takes a colour of the theme's chart palette. The figure above each chart is in the text colour — in the chart's colour it read 2.0 to 3.8:1 in the light themes — and shares one line with its label.

Fixed

  • [core+ui] The players grey the buttons that would do nothing, and no longer remove any. On a station played alone, the full-screen player offered Repeat, Previous, Next and Shuffle as if they worked: Next stopped playback, Previous only reconnected the stream, and Repeat or Shuffle set a mode that did nothing there and stayed on for the next queue; the playback bar offered Next too. The box now works out from its player's queue what Previous, Next and seeking would do — whether another entry follows or comes before in play order, repeat and random included, and whether the entry has a length (measured on MPD 0.24.5) — and both players grey what would go nowhere: Previous and Next on a station alone, Next on the last track of a queue, both while the player is stopped, which then refuses them. On a station alone, Repeat and Shuffle are greyed too, unless one is on, so it can still be turned off; with tracks queued behind a station they stay lit, since they act on those. Paused, a change of what can be skipped reaches the screen within seconds instead of waiting for something else to move. No button is removed any more: the playback bar dropped Previous and Next when a source could not use them, and Play slid into their place — on the last track of an album, under the finger that had just pressed Next; for Roon or AirPlay they are greyed as well. On a phone the playback bar still leaves Previous to the full-screen player. With HQPlayer as the output, Previous and Next stay lit.
  • [ui] Systemd's I/O Acc and IP Acc tags are grey while the accounting is off. They were red, the colour of a failure, for an option turned off; they are now grey and still when off, green and beating when on, in every theme.
  • [ui] UPDATE ALL is orange. The button Audio Software shows while updates wait was red, the colour of a failure; it now takes the orange of what waits on you.
  • [ui] Library › Sources' Scan renderers and Scan network are as wide as their label. Both stretched across their card — 327 px on a phone, 860 px on a computer — and a long bar at the compact buttons' 24 px read as lower than the DSP and Disconnect buttons beside it, though it measured the same height and type. They now keep the width of their label, aligned left like the page's other buttons; the renderers a scan finds, listed in the same column, keep the full width.

[0.9.70] - 2026-10-11

Changed

  • [ui] On a touch screen, a button fades under the finger. With hover kept for a mouse, the only sign of a press was a 4% shrink — under 2 px on a 40 px button, hardly seen on an iPhone. While the finger is on it, a button now also fades to 60%, and comes back as the finger lifts; a dimmed button fades in proportion rather than brightening. A busy button does not react. A card that holds buttons of its own — a radio's star — stays still while one of them is pressed, and the Now Playing cover stays still while its album-details bubble is touched. A radio being swiped away stays opaque, so the red Remove behind it no longer shows through. On a computer nothing changes. The audio pipeline diagram's buttons, shown on a tablet or a phone held sideways, still show nothing under the finger.
  • [ui] Five windows' buttons are written as the confirmation's. The User, Systemd override, JSON configuration, operation log and Support report windows kept the older grey buttons — Cancel filled grey, Save outlined green — once their size was aligned. Shown on mock-ups and chosen: Cancel, Close and the tools (Copy, Download, Upload) are outlined, the window's main action is filled — Save, and the support report's Copy, which is made to be pasted into a message — and Cancel Operation is orange, the colour of what interrupts or removes. A disabled or unavailable framed button no longer reacts under the mouse: a disabled Save still flipped to its outline, because a second copy of its rule, loaded later, won over the first; a disabled orange button and, with this change, a disabled Cancel turned black. The JSON window's Upload and Edit, unavailable while its editor loads, read as such again. A disabled button in these windows shows at 30 % opacity, as every framed button does, where the grey ones showed at 50 %.
  • [ui] Every field you type into has one size and one font. Measured in Chromium, the fields came in a dozen looks — 24 to 42 px tall, text at 11 or 12 px, the network share's in a monospace face: the library search, the HIGHRESAUDIO search filters, the radio search and new-station form, the UPnP server address, the Tidal code, the HIGHRESAUDIO sign-in, the HQPlayer address, the network share and library path, a passkey's name, the network test. Shown on mock-ups and chosen: every one is now 29 px tall, its text at 12 px in Inter — the size of its label — with the same padding, border and corners, on a computer and under a finger alike. Lists keep their own look, save the three of the HIGHRESAUDIO search filters, which share the fields' row; the library search keeps its magnifier; a compact button beside a field keeps its 24 px. The web password an install shows in clear stays in a monospace face, to tell an l from a 1. The active field is marked by its border turning to the text colour, as the dialogs' fields already were: the fields that showed a 3 px focus ring no longer do. The sign-in page does not change.
  • [manual] Eleven figures show this release's fields and titles. Retaken on 2026-10-11 from the live interface, at the same iPhone 13 mini size: the guided editor, the audio stack's set-up, the network share form and the network test, with their fields in the shared look; Audio Software without its DOWNLOAD button; the space under the titles of the latency test, of the licence panel — on trial and activated — and of HQPlayer Embedded's configuration; the HQPlayer Embedded install window offering 6.1.1-5; and the sign-in page of 0.9.69, which still showed 0.9.54. The licence panel's Device ID and Order ID stay blurred.

Removed

  • [core+ui] Audio Software no longer offers to download the package catalogue. The DOWNLOAD button in the tab's header saved packages-config.json, the list of audio software the box works out by itself — what is published, from where, for which machine. Nothing could be done with the file: the box rebuilds it on its own (REFRESH), and nothing takes it back in. The button is gone, and with it the core's GET /packages/config/view, which served it to no one else. The support report still lists every package with its state, its version and whether it runs on this machine.

Fixed

  • [core+ui] Play starts the queue again once it has played through. When the box's player had stopped with its queue — the last track over, or stopped by another app — Play did nothing: the box sent it a pause/resume switch, which that player ignores when stopped, and once the box had restarted it answered that no source was active. Play now starts the current track, or the top of the queue; on an empty queue it changes nothing and says so. Measured on the box's MPD 0.24.5. While nothing plays, Play reaches the selected network player, else the box's player when it was the last one used — not when Roon, AirPlay or another source was, nor while HQPlayer is the output.
  • [core+ui] Play and Pause do what the screen shows. A tap sent a switch, which the box turned into a pause or a play from the player's state when it arrived: a state that had moved in between — the queue running out a second earlier, another source taking over — turned a Pause into a Play. A tap now sends Play or Pause, as shown, to every player — the box's, AirPlay, Roon, HQPlayer, a network player; repeated, either changes nothing. The sleep timer sends a Pause the same way: firing as the queue ended, it could have started the whole queue again. A network player between two tracks is paused too, where it was left playing, and HQPlayer is asked its state at that moment rather than from a reading up to 2 s old; out of reach, it says so.
  • [ui] A screen opened while nothing plays shows the player's state at once. The full-screen player opened from the Library waited for the next change to show anything, so its Play went out with no player to reach.
  • [ui] The player's state no longer arrives twice after the box was out of reach. A screen opened while the app was reconnecting opened a second connection, never closed, and every state then came in twice until the page was reloaded.
  • [core] Play on a paused radio station picks up the live broadcast. After a long pause, the station came back at the moment it was paused, then stopped a few seconds later: the box kept the station's connection open without reading it, the station dropped it, and only the seconds already received were left to play. Measured after a 30-minute pause: the box found the connection gone 20 s after Play, and the sound stopped at 60 s; a network renderer did the same after a 2-minute pause. Play now reconnects to the station — after a short pause too, so what was said during the pause is not kept. A paused track, from the library, a media server or a streaming service, still resumes where it was. This covers the box's own output and network renderers; with HQPlayer as the output, a paused station still resumes where it was paused.
  • [ui] On a touch screen, a button no longer stays lit after a tap. A phone or a tablet has no pointer to hover with, so the browser gave the hover look to the last thing tapped and kept it there until something else was tapped: a tile, a filter, an album, a dialog's button stayed in its hover colour after it was pressed, as if still held or chosen. Every hover effect now waits for a mouse or a trackpad; one that shared its rule with another state — the chosen tab, a pressed button — keeps that state on a touch screen. Nothing that hover revealed is lost: the cover's actions in the Library already show by themselves on a touch screen. Measured in Chromium on the stylesheets before and after: on a computer, every style is the same; on an emulated phone, 111 hover styles are gone and nothing else. The offline page's Retry button follows.
  • [ui] Text in the accent colour reads clearly in every theme. Links — the manual's and the update banner's too —, the manual's current chapter, a favourite's star, the source badges of Now Playing and of the mobile pipeline, the high-resolution value of the format strip, the pipeline diagram's volume and source format were written in the accent colour, under the 4.5:1 contrast text needs: 3.6 to 4.47:1 in Slate and Gravity light, 3.5:1 on grey panels in Slate dark, 3.9:1 for the badges on the accent's own tint. Text in the accent now takes a shade of its own — a deeper indigo in the light themes (#4F46E5 for #6366F1), a lighter one in Slate dark (#A5B4FC for #818CF8); Minimal and Gravity dark, already clear, keep theirs. The lowest is now 5.1:1. Fills, outlines and underlines keep the accent as it was; an icon drawn in its label's colour takes the label's shade.
  • [ui] On a phone, Audio Software's title is the size of every other page's. Held upright — a screen 480 px wide or less — it dropped to 14 px, against 18 px for the others, from a rule of its own that put it on a line above its buttons. It keeps its own line, at 18 px. A card title, on any page or panel, can no longer be given a size of its own.
  • [ui] Every card title stands 24 px above what follows it. The space under a title ran from 12 to 30 px depending on the page, measured on every tab: Services, Audio Software, Profiles, Performance and the logs sat at 30 px, Systemd, System and the licence at 12, Config at 28, and the HISTORY and EVENTS panels at 36. It is now 24 px under every page title, panel header and Performance test, on a phone and on a computer; on a phone, the Pipeline's title, alone in its own card, stands 24 px above the chain too.
  • [ui] The User and Systemd override windows draw their fields. Their 19 fields had no style at all and showed the browser's own field: 20 px tall, no padding, an inset border, and a width of the browser's choosing — 192 px for a name in a 336 px column, 37 to 87 px for a number. They now take the shared field, across the full width of the window.
  • [ui] A multi-line field no longer makes an iPhone zoom in. Safari zooms the page into a field declared under 16 px and does not zoom back out. The app kept single-line fields at 16 px under a finger, but a multi-line one lost to any rule giving it a size: the config editor's multi-line settings were declared at 12 px (measured in Chromium with a touch pointer). They now get the 16 px every field does, drawn at the size of their label. On a computer nothing changes.

Security

  • [ui] A confirmation shows the names in it as text. The confirmation dialog read a text message as markup, so a name inserted into one — a passkey's, a network share's label and the box's message about it, a service, a backup file, the CPU governor, the version an update offers — was interpreted, not shown: a tag in it formatted the question, or could carry a script. Every one of these names is typed by the box's owner or comes from the box itself, so none could be used against it. The dialog now shows every text message as text, whatever it holds; the few questions that need bold or a list — removing a passkey or a user, a refused systemd setting, Update All — are built so that the names and versions in them are text too. A text message that holds a tag is reported in the browser console.

[0.9.69] - 2026-10-09

Changed

  • [ui][manual] On a computer, the tabs stay a bar across the top. The Switch button at the end of the bar turned them into a sidebar and back, and the browser kept the choice; it is gone, and a sidebar chosen before turns back into the bar. A phone, and any screen whose short side is 768 px or less — a small tablet, a 1366 × 768 laptop — keeps the sidebar, as before. The layout now follows the screen both ways: a window carried from a small screen to a large one gets the bar back, where it used to keep the sidebar until Switch was pressed; the sidebar comes back as it was left. The manual says which screen gets which.
  • [ui][manual] Manual stands at the right-hand end of the tab bar. On a computer it followed the last tab; it now stands apart, at the far end, as it stands at the foot of the sidebar. When the tabs fill a narrow window it simply follows them.
  • [ui] The name at the head of the sidebar is a little larger. Audiogravity, beside the licence badge and the connection dot, goes from 12 to 14 px — the largest step that keeps the three on one line with the longest licence label, Version expired; at 20 px that label touched the name and pushed the dot out of the sidebar.
  • [ui] Where the tabs are a sidebar, the connection status shows once. Beside the menu button, a dot and Connected repeated the dot the sidebar carries next to the logo — on a phone held sideways, a 1366 × 768 laptop. They now show only when the box stops answering: Connecting... comes back there, the one place left to say it while the sidebar is closed. On a computer's tab bar, which has no other, nothing changes.
  • [ui][manual] Removing a station, a UPnP server or a renderer asks first. A left-swipe past the threshold removed it at once, so a swipe carried a little too far while scrolling was enough to lose one; Remove from My Live Radio, under a station's ⋯, did the same. Both now open a confirmation that names the item and the list it leaves — Remove FIP from Favorites? — with Remove in orange and Cancel selected; cancelled, the row stays where it was. This covers the swipe in My Live Radio and Favorites, on Sources' UPnP servers and on Audio Output's renderers; the queue's swipe still removes a track at once. When a removal would erase a station added by hand — no other list keeps it, and its address goes with it — the question says so, and in that one case the star asks too; anywhere else it still removes in one touch. The manual's Library and Outputs chapters say so.
  • [ui] Held sideways, the sign-in card is two columns. On a phone held sideways — a screen 500 px high or less — the brand, the version, the core and the licence stand on the left and the form on the right, the footer under the brand; the light/dark switch moves to the empty top-left corner. The automatic passkey panel does the same: who and where and the core on the left, the key on the right. Measured in Chromium at 956 × 440, 812 × 375 and 667 × 375, the card fits without scrolling, and it keeps clear of the notch, the Dynamic Island and the home indicator. Upright and on a computer nothing moves.
  • [ui] The passkey key is filled, as the Login button. Black with a white key in a light theme, the reverse in a dark one; it was outlined, and filled only under the pointer. Under a mouse it opens up as Login does; on a touch screen it stays filled.
  • [ui] The app no longer makes an empty frame at every launch. The API reference's window, in the page from the start, held a frame even closed and with nothing to show, so a blank document was made each time the app opened — a probe took it for the app loading twice after a sign-in. The frame now comes with the reference and leaves once the window has faded out; the reference loads afresh at each opening, as it already did.

Fixed

  • [ui] The sign-in page scrolls when its card is taller than the screen. Held sideways, or upright on a small phone with the passkey button, the card was centred on a page that could not scroll: its top and its bottom were out of reach. It now starts at the top and the rest scrolls into view; a card that fits does not move.
  • [ui] On the sign-in page, the licence stands as far under the core status as the lines above. It sat 4 px under CORE · CONNECTED, read as part of that line; it is now 8 px under, in the header and in the passkey panel.
  • [ui] After signing in, the raw tab bar no longer shows before the app. The tabs written in the page appeared on a blank screen while the app loaded — horizontal even on a phone — and could be taken for another screen; they stay hidden until the app draws them.

[0.9.68] - 2026-10-08

Added

  • [manual] The manual says how to activate a licence key, and how to check one. The Licence section of the administration chapter walks through the License activation window's three steps — key, host, activate — and every refusal it can show, at each step; then Verify existing license and what each status means: ACTIVE is a key already activated on this box. Three figures show the second step, a successful activation and a checked key, taken with an example key and a made-up Device ID.
  • [manual] The introduction lists what each edition opens. Sixteen rows, Starter against Pro, under the names of the app's own tabs. They come from what the app locks: the Library, Pipeline, Systemd and Performance tabs, the playback controls and the full-screen player. The short table above them no longer gives a partial list of its own.

Changed

  • [ui] The small buttons share one size: 24 px, 11 px type. The tiles' buttons — Start, Stop, Restart, Install, Edit Config… on Profiles, Services, Audio Software, Systemd and Config — stood 18 px tall at 10 px; the filters 20 px; Audio Software's Refresh, Download and Check updates 20 px; and the compact buttons of the licence, Admin, System and Performance panels 24 px, but with twice the side padding and, in one of their two families, a lighter weight and square corners. They now share one height, padding, type, weight, letter spacing and corners, on a phone and a computer; each keeps its colours. In a Services tile, the Disabled badge beside Stop and Restart takes their height, and so does the network test's service state beside its Start button. The selectors follow, from heights of 18 to 28 px: Library's Browse pills, Search sources and Radio views (with the custom-station form's Cancel and Save), the log's level buttons and the Pipeline diagram's panel, whose buttons also leave Arial for the app's typeface. The small actions follow, from heights of 18 to 26 px: Radio's + Add custom station, the events' Running / Stopped, the Clear of the history panels, the log's refresh and Live, and an account's Enabled / Disabled. The queue's Refresh and Clear, words of 12 px in the line of the track count, are now buttons framed like the source filters below them; a station's Edit station and Remove from My Live Radio, 44 px tall on a phone, take the small size too.
  • [manual] Ten figures show the new button sizes. Retaken on 2026-10-08 from the live interface, at the same iPhone 13 mini size: Audio Software, Services and Config with their tiles' buttons and chosen filter; the Library's Radio, Browse and mixed queue; the HQPlayer Embedded install window and both HQPlayer cards; and the update banner, which still showed 0.9.55.
  • [ui] The Pipeline diagram's panel rules its sections in every theme. The lines between its groups were white at 10 %, which a light theme did not show; they take the theme's border colour.
  • [ui] The chosen filter is filled black, as in the Library. The filters of Profiles, Services and Audio Software, and the queue's source filters, showed the chosen one in a light grey; it is now filled with the text colour, like the chosen category, source or view of the Library's Browse, Search and Radio.
  • [ui] A dialog's buttons share one size, and a confirmation that deletes says so in orange. They stood 18 px tall on a computer in the confirmations and 29 px elsewhere, and 44 px on a phone; they are now 29 px everywhere, in 12 px type of weight 600, with the small buttons' corners. In a confirmation, Cancel is outlined and OK filled. OK turns orange, the colour of every button that deletes, on the thirteen confirmations that delete or remove something or throw away unsaved changes (Clear History, Remove Passkey, Delete License, Delete User, a network share's removal, Systemd's Remove Override and Restore Backup…), and so does the uninstall dialog's Uninstall, like the tile's Uninstall that opens it.
  • [ui] The installed app opens on its icon, its name and a wave. On a phone it opened on a drawn key beating under its name, at least 2.5 s at every launch, after iOS had shown the name alone in an older typeface. It now opens on its icon, Audiogravity and Your audio chain, fully controlled, coming in one after the other, then a wave going out of the icon for as long as the app loads — at least once, 2.1 s from the first motion. On a black ground whatever the theme: the plain ground iOS now shows before it, so one leads into the other without a jump; in a light theme the app comes out of black in the fade. The app's own Animations setting leaves it animated; with the device's Reduce motion on, nothing moves and the icon's light pulses slowly instead.
  • [ui] The opening screen shows once per opening of the app. It played again after a sign-out, before the sign-in form, and after a sign-in, before the app; a launch with no session cut it short, then played it again on the sign-in page. It now shows at launch only — on the sign-in page when there is no session, the same screen as the app's — and not after a sign-out, a sign-in or a reload. In a browser tab it no longer flashes up before the page.
  • [ui][manual] The licence panel's EDITIONS & LICENSE button is gone. It opened a copy of the editions and of the licence agreement kept by hand in the app. The editions are in the manual's introduction, and the licence agreement is linked from the sign-in page and from the footer. LICENSE KEY moves up beside the panel's title, and the manual's two figures of the panel are retaken without the button.
  • [manual] A locked Pro tab is said to open the activation window. On Starter, tapping one takes you to the licence panel with the License activation window open over it; the manual said it opened the panel.

Fixed

  • [ui] A confirmation opens with its default button selected. It was meant to select OK, but looked for a button by a name none of them wore, so Enter did nothing. It now selects OK, Cancel when OK deletes, or the password field when there is one.
  • [site][manual] Every edition is said to receive updates. EDITIONS.md gave the trial and Starter no updates, and the FAQ, the manual's and the landing's, named Pro alone; the landing's Starter card now says it too. The trial, Starter and Pro all receive the updates of their major version, as the licence server offers them; Pro holders keep a preferential price for the next major version.
  • [ui] Leaving the app no longer stops on an error. On a reload or a sign-out, the app closed its live stream with a call the stream does not have, and stopped there. It now closes it as it does when sent to the background, and the stream comes back if the page stays.
  • [ui] The activation window shows its last step done. Once the licence was active, 03 · ACTIVATE stayed black under a green Activation successful: a step is done once a later one is shown, and none comes after the third. All three now turn green.
  • [ui] The activated licence's Device ID stays inside its frame. Sixty-four characters with nowhere to break, it ran past the certificate's edge on a phone; it now wraps.
  • [ui] In the top bar, each name sits on its figure's line, beside its own figure. Wherever the names read inline, from 430 px wide, a name sat 2 px above its figure and 10 px from it — the uptime's 4 px — and on a phone or a tablet a figure stood nearly as close to the next name as to its own. Each name now shares its figure's line, 8 px from it, and two pairs stand 16 px apart. A figure turning orange or red still moves none of its neighbours. Under 430 px, where each name sits over its figure, nothing changes.

[0.9.67] - 2026-10-06

Changed

  • [ui][manual] In the tab menu, a tap on Library unfolds its views. It opened the page, with the library's views always spread under it. On a phone, and wherever the tabs are a column, Library now carries a chevron: a tap unfolds Browse, Search, Queue, Sources and Radio without leaving the page shown nor closing the menu, a second tap folds them, and the menu always opens with them folded. Coming from another tab, the view you were on opens as you left it, an album or an artist still open. On a computer's horizontal bar, a swipe or an arrow key, Library still opens the page. The manual's first-run chapter shows the menu, Library unfolded.
  • [manual] Sixteen figures show the interface of 0.9.66. Retaken on 2026-10-06 from the live interface, at the same iPhone 13 mini size: the queue's mini-player without Previous, its title and artist on a line each; CRITICAL, CONFIGURED, ACTIVE, READY, MOUNTED and the licence's tags as tints, no longer framed like buttons; DELETE LICENSE in orange; the buttons of the Audio Software and Pipeline headers in their own colours; check boxes and radio buttons in the theme's. The playback figures are taken with their original tracks.

Fixed

  • [core][manual] A station that names no song is shown under its name. France Inter sends no song title: the full-screen player read Nothing playing while it played, the Pipeline tab's card Stream active and the queue franceinter-hifi.aac. The player, the mini-player, the Pipeline card and the queue now show the station's name — the radio list's, else the one the stream announces, for a station another app started — as they already did through HQPlayer or a network speaker. A title made of dashes alone counts as none.
  • [manual] The Library tab's views and the way to the outputs read as the app shows them. The views are listed in the order of their bar, Radio among them; the list of outputs opens from the Outputs button of Sources, or from Switch under the full-screen player — there was no Library → Outputs view. On a computer, Audio events sits under the Pipeline tab's list, beside it only in a narrower window. The glossary gains Lossy, and says the HI·RES badge appears from 88.2 kHz.
  • [ui] The manual's capture tool frames a radio's Pipeline card without cutting it. The figure of the stream-origin badge stopped 22 px under the title, through the format row a radio's card now carries there: it stops before that row, keeps the card's text whole, and takes its boxes from one card.

[0.9.66] - 2026-10-06

Added

  • [ui][manual] The app can follow the device's light or dark mode. Settings → Appearance offers Automatic, Light and Dark in place of the light/dark switch. Automatic, the default, follows the device — light by day, dark at night on a device that switches by itself — and so does the sign-in screen, from its first paint. A device set to light or dark with the old switch keeps that choice.
  • [ui][manual] A live radio shows LIVE instead of a progress bar. A live stream has no end and no position: the bar gives way to a LIVE marker under the transport, which dims while the stream is paused.
  • [ui][manual] On a computer, the Pipeline page lists the streams beside the diagram. What plays, through what, to which output — the list a phone shows — sits in the right-hand column, with Audio events under it; on a narrower window, the list and the events sit side by side under the diagram.
  • [ui][manual] The tab menu opens the library's views directly. Under Library, the ☰ menu carries the library's own tabs — Browse, Search, Queue, Sources and Radio, three for the radio, which has no albums to browse — in a tray set in from the edge, smaller than the entry they belong to. A tap opens that view, from any tab, and closes the menu; the view shown is marked by an orange tint and bar, and told to a screen reader. Only where the tabs are a column: the horizontal bar of a computer keeps the library's tabs under the Library tab.

Changed

  • [ui][manual] On a phone, the mini-player gives the title its room. Title and artist shared one line, in the width four buttons left them — 115 px on an iPhone mini — and had to scroll to be read. Previous leaves the bar (the full player keeps it; volume stays), and title and artist take a line each: 159 px of text.
  • [ui][manual] A radio card's actions open from ⋯. In My Live Radio every card carried a green tick that meant Remove from My Live Radio: touching it removed the station. A ⋯ button now opens Edit station and Remove from My Live Radio.
  • [ui] The settings panel and the tab menu are opaque. At 40 % with a blur, the page showed through them: coloured smudges and stray words between the switches and under the tab names.
  • [ui] On a phone, every page's title sits at the same distance from the tab menu. Six tabs left 2 px between the menu and the title, three left 6 px; all now leave 6 px.
  • [ui] Text faded at rest reads again. The cards of stopped services, the dates of configuration files and the PERSIST label were faded down to 2.8–3.5:1, and the number of backups on a Config card and the card of a package that is not installed — one INSTALL away — under the floor too; they now keep the contrast of the text around them.
  • [ui] Check boxes and radio buttons take the theme's colours. They took the browser's blue, and in the dark palette an unchecked box was a white square: they now take the text's colour, drawn in the light or dark palette in force.
  • [ui] System figures are written as they are. Values were set in capitals — X86_64, and MHZ, which is no unit; they now read x86_64 and MHz, the unit smaller on the number's line.
  • [ui][manual] Plainer words in Admin. The live feed's tile is Live updates, no longer SSE Stream, and the cockpit at the bottom of the tab is the UI performance cockpit.
  • [ui] On a computer, the library's lists are centred. The queue, sources, outputs and radio lists stretched across the whole screen; they now keep to 900 px, centred, under a tab bar that stays in place from one view to the next.
  • [ui] The full-screen player uses a landscape screen's width. From 900 px wide in landscape, the cover sits on the left and the track, its controls and what comes next on the right, instead of one column under the cover.
  • [site][ui] The editions speak of a simulation, not a dry-run. The Starter edition's software line, on the landing and in the app.
  • [ui][manual] The top bar names its figures. CPU, Temp and RAM were hidden under 1,025 px, which left three numbers no one could name. They now read inline — CPU: 12% — wherever there is room, and over their figure on a phone under 430 px. A figure takes a colour only past a threshold, orange then red, where all three were always green. A phone keeps those three and leaves the uptime to a computer's bar; on a small tablet, the word Connected gives way to the figures, as on a phone.
  • [ui][manual] Audio Software simulates, and offers UPDATE only when it does something. The DRY-RUN switch is now SIMULATE. REFRESH and DOWNLOAD carry their word next to their icon. UPDATE shows when another version is offered — or for Roon, whose installer always fetches the current build — where it showed on every installed package and answered Already Up-to-Date.
  • [core] The audio pipeline goes out when it changes. It went out every 30 seconds to every screen, identical but for its time. A screen that connects gets the last one at once — to it alone, without the box computing it again — and the pipeline's time is now written in universal time.
  • [ui] The phone's pipeline list stops asking for its outputs every 15 seconds. It reads them again when an output changes.
  • [ui] In landscape, the full-screen player's column is centred on the cover whatever it shows, and a phone held sideways has less to scroll.
  • [ui] CRITICAL reads as a tag, not as a button. On the Config, Services and Systemd cards, the badge that marks a service the audio chain depends on was outlined like the buttons beside it — next to EDIT CONFIG it read as a second one. It is now a tint without a frame, like STOPPED at the top of the same card, and reads CRITICAL on all three.
  • [ui] Every button that deletes is orange. Deleting the licence, removing a station from My Live Radio, confirming a playlist's deletion and removing a passkey were red, beside the orange of deleting a user, uninstalling a package or removing an override. All are orange now, their label too: on a light page the orange border alone read at 2.1:1. The controls that take a colour only under the pointer — a passkey's cross in the settings, a network share's, Clear in a test's history — turn orange as well, and a disabled button no longer fills on hover.
  • [ui] A frame means a badge can be pressed. Every badge that is not a button is a tint without a frame, wherever it sits: FAILED and NOT INSTALLED on the Services cards, UNAVAILABLE and N failed on the profiles, Not configured, Not Supported and Test Package on Audio Software, which sat among buttons outlined like them; and with them the real-time monitor's SCHED_RR and STOPPED, the network test's service state beside START, the counts and results of a validation, ACTIVE and READY on the library's output cards, the performance monitor's and the test history's tags, THROTTLED, the mini-player's source and output, the transport of the Pipeline list. A badge that is pressed keeps its frame, and Restart required, the one button drawn without one, has it back. A tag with no colour of its own takes a translucent grey, which also shows on the grey cards of the guided configuration.
  • [ui] A page title's badges look like what they are. In the titles of System, Pipeline, Admin and Audio Software, every badge was painted the same blue: LIVE, which is not pressed, looked like NEW USER or REFRESH beside it, and white on that blue read 3.5:1 in Slate light and 2.5:1 in Gravity light. LIVE is a tint now, and the buttons are outlined in their own colour — orange for CONFIG, NEW USER and REFRESH, grey for DOWNLOAD, green for CHECK UPDATES, red for UPDATE ALL.
  • [ui] A pulse no longer fades a badge's words. LIVE, I/O Acc, IP Acc, Update available, Restart required and THROTTLED pulsed by turning half transparent, words included: 2.1 to 4.1:1 at the bottom of the beat. The tint of a tag, or the frame of a button, breathes now, and the words stay; a filled badge does not pulse.
  • [ui] The tiles of software that is not installed fade only what is inert. On Services, Profiles and Systemd the whole tile was faded, and with it the NOT INSTALLED or UNAVAILABLE badge that says why; the name, the state and that badge now read, and the buttons keep the look of a disabled button.
  • [ui] The log's levels read in every theme. A level sat on its own tint over its row's, and the level and the time beside it fell under 4.5:1 in four theme modes. The levels are words in their colour, an error or a warning row is marked by a stripe on its left, and DEBUG, at 1.5 to 3.6:1, takes the secondary grey.
  • [ui] Red text reads on its tint in the light Minimal and Slate themes. On its tint, the red of FAILED read at 4.3:1 in Minimal and 4.1:1 in Slate, under the 4.5:1 floor; the red of those two light themes is a shade darker and reads at 4.6:1, and Slate's light information blue, under the floor on its tint too, follows.
  • [ui] The Pipeline tab opens on the pipeline the page already has. The stream hands every screen the current pipeline as it connects; the tab, loaded at its first opening, missed it and asked the box to compute it again. The page now keeps it from the start.
  • [ui] The library's Library tab is now Sources. It opens the list of sources to play from — local, streaming, Roon, media servers — and read twice in the tab menu, under the Library entry it belongs to. The manual already called it Library → Sources.
  • [ui][manual] Library is the last tab, after Admin. In the menu and in a computer's tab bar alike; for an account that does not see Admin, nothing moves.
  • [ui] Manual sits at the foot of the tab menu, apart from the tabs — just above Switch on a computer whose tabs are a column.

Fixed

  • [ui] The System tab shows the memory in use. It read two fields the core does not send, and showed 0.0 GB / 0.0 GB on every box; it now shows the memory in use and the total, such as 2.8 GB / 3.7 GB.
  • [core][ui] A simulation no longer passes for the real thing. It answered Successfully installed …, its log ended the same way, and a simulation that failed left its card in error. It now answers the steps to install … went through — or a step … failed — nothing was changed, its log ends on that sentence and its window on Simulated: nothing was changed, the card keeps the state it had, the history reads Simulate install …, and a simulated uninstall that would delete the settings says so.
  • [ui] Restart required appears only after an install or an update that went through, on the device that ran it. A simulation, an update the box refused and an operation run from another device lit it too, and a press restarted the service — cutting the music — for nothing.
  • [core] A failure repaired outside Audiogravity stays repaired. After a simulation or a refused operation, a card whose package had been repaired from a terminal went back to ERROR until the core restarted.
  • [ui] Audio events records what the players did. When the diagram's first reading of the pipeline failed, the speakers, the amplifier, the turntable and the phone all "started playback". A line is now written when a player starts, stops or moves to another track — a pause and a resume included.
  • [ui] The Pipeline page no longer goes back in time. A reading that came back late, or the pipeline kept for offline use, could put an older state back on screen.
  • [core][ui] On a box with MPD on several outputs, the output pills and the Config tab name the output it plays on. They named the first output of its configuration while the diagram showed the one playing. Asked of MPD, the answer waits two seconds at most: an MPD that stopped answering held the diagram for good.
  • [ui] The update window's title said Updateing.
  • [ui] An arrow key pressed on Manual or Switch in the tab bar no longer opens the next tab. Only a key pressed on a tab moves to another.
  • [core] An MPD that stops answering no longer holds the screens for good. An MPD that accepts a connection and never answers — a suspended one does exactly that — held the Pipeline page, the choice of an output, the library's indexing indicator, the tracks of the album playing and a cover read from its files. The Pipeline page now waits a few seconds at most, the others ten — a cover, asked twice, twenty.
  • [ui] Started without a network, the Pipeline tab shows the last pipeline kept. It said Pipeline Inactive — No audio sources detected.
  • [core][manual] A radio station shows the format it broadcasts in, not PCM. The box read the format from the stream's address: when the address does not name it — …/stream/733765 — the player showed PCM, read as lossless, for a 128 kbps MP3, and the Pipeline tab's list showed no format at all. Both now show the codec the radio catalogue gives the station, the one the radio list already showed. A web stream whose format no one can tell is no longer called PCM; a CD still is. The manual's live hi-fi readout names the codec.
  • [core] A station played from a search keeps its name, logo and format while it plays. The box remembers the last thousand stations seen and let go of the oldest — the one playing among them, after enough searches.
  • [core] Changing a station's address drops the old stream's format. The radio list and the player kept the codec and bitrate of the previous address.
  • [core][ui][manual] A compressed stream shows no bit depth. MP3, AAC, Ogg and Opus store none: the 24bit shown for an MP3 radio was the resolution MPD decodes at — 16bit for France Inter's AAC. The player's FORMAT now reads Lossy there and the Pipeline tab writes the bitrate in its place, MP3 · 128kbps · 44.1kHz; a lossless file keeps its bit depth. A lossy stream never earns the Hi·Res badge, whatever its rate.
  • [core] Tidal's compressed tiers show AAC, not FLAC. Every Tidal address read FLAC; the box now believes the codec Tidal names for the stream.
  • [core] A station the catalogue lists with a video codec — AAC,H.264 — shows its audio codec, and is known as compressed.
  • [core] The radio search's Hi-Res only takes an Ogg station on its bitrate, like an MP3 or an AAC one. Every Ogg station passed as lossless; in the catalogue they broadcast Vorbis or Opus.
  • [ui] The Pipeline tab's list shows the format the box computes, the one the player shows: a CD reads PCM | 16bit | 44.1kHz there, where nothing showed.
  • [core] Simulating an install introduces itself as Audiogravity. Its check that each download answers went out as a bare Python client, on a connection of its own; it now goes out as the download itself does.

[0.9.65] - 2026-10-04

Added

  • [site][manual] A button takes the reader back to the top. A round button in the bottom-right corner of the landing and of every page of the manual on audiogravity.app, shown once the first screen is behind, with a ring round it that fills with the reading. It scrolls up without writing to the address or the history, so Back still leaves the page; it jumps instead of scrolling for whoever asks for less motion, and moves the keyboard focus to the bar's first link. The ring is drawn by a small script on every browser: drawn by the browser itself, it stayed empty in Safari. The landing's logo already led to the top, but nothing said so, and on a phone the top of the screen is far from the thumb.
  • [ui] The Manual window has the same button back to the top. Its chapters scroll inside the window, out of the website's button's reach. A round button in the chapter's bottom-right corner, shown once a full window of it is behind, with a ring round it that fills with the reading: it brings the chapter back to its top, jumping instead of scrolling for whoever asks for less motion, and moves the keyboard focus to the chapter, so the arrow keys keep scrolling it. It sits inside the chapter: a wheel or a finger that starts on it still scrolls the chapter, and at the end it rests under the last line instead of over it — chapters end 70 px lower to make room. On an iPhone, the end of a chapter now stops above the home indicator (not tried on an iPhone).

Changed

  • [ui] The interface no longer loads anything from the Internet. Chart.js, CodeMirror and xterm.js came from jsDelivr, a public CDN: 14 requests at every cold start, whatever the screen, and the device had to reach the Internet, not only the box. The CodeMirror stylesheets held back the first paint, and the module that starts the app waited for Chart.js: with jsDelivr answering 3 s late, the app ran after 6.2 s instead of 0.4; with jsDelivr not answering at all, after 20 s (measured in Chromium on an installed 0.9.64). And the eleven CodeMirror scripts, run in every page without an integrity check, could read the session. The three libraries now ship with the interface and come from the box — 0.2 to 0.5 s to start, whatever the Internet does (measured) — and each one is loaded by the screen that needs it, when it opens: the Performance tab's charts, the configuration editors, the terminal. The service worker keeps them with the rest of the version. The pages' security policy now lets scripts and styles come from the box alone. The terminal moves from xterm.js 4.19, deprecated under that name, to 6.0.
  • [ui] The configuration editors say when they are not ready. CodeMirror now arrives when an editor opens. Until it is there, the Expert editor says Loading the editor…, its text cannot be edited, and Save answers Editor not ready instead of saving; the topology window keeps Edit and Upload unavailable, and they no longer light up under the pointer. If CodeMirror cannot be loaded, both say so, and send to a reload of the page: the browser keeps a failed load for the life of the page (measured in Chromium), so opening the editor again would not load it. The terminal says the same when xterm.js cannot be loaded, instead of offering RETRY.
  • [ui] The tab strips no longer glide for whoever asks for less motion. When the tab changes without the strip being touched — a swipe across the page, a link — the main tabs on a phone and the library's tabs scroll themselves to show it, and the arrows of the library's filters scroll their row. They glided unless the app's Animations switch was off; they now also jump into place when the device is set to reduce motion (Reduce Motion, on an iPhone or a Mac), as the app's other animations already did. A scroll the page asks to be smooth ignores that setting (measured in Chromium).
  • [site] The landing's Now playing strip shows the player screen. Twelve dancing bars suggested a visualiser the app does not have, two chains told one path two different ways, the chain went from Roon straight to the HQPlayer NAA — leaving out HQPlayer, which makes the DSD256 — and began at iOS, which the sound never leaves. The strip now shows what the app's player screen shows, in the page's style: the album cover, the signal path (Qobuz → HQPlayer → HQPlayer NAA → USB → Heed Abacus), the four format cells (DSD256 · 11.3MHz · 22.6Mbps · DSD), the progress with the time left and the output, and a link to the signal-path diagram below it. The track is Come Together, from Abbey Road, 4:19 on Qobuz.
  • [site] Every text on the landing meets the contrast floor, in both themes. 29 texts read under 4.5:1 in the light theme and 26 in the dark one (measured in Chromium, element opacity counted): green and orange as text on white (2.3 and 2.2:1), the section labels and the palest grey (3.2 to 3.5:1), the footer, the code comments, the ghosted numbers of the five-layer stack (1.7:1), the bar's date. None does now: green and orange take a darker shade as text in the light theme, the palest grey goes one step up, the always-dark blocks keep their own greys. The signal-path diagram's lines, arrows, frames and moving dots, and the carousel's dots, are above the 3:1 a graphic needs.
  • [site] The landing's layout holds together. The backgrounds alternate down the page again: the figures under End the jitter now share its band. The feature cards are ordered by height, which cut their empty space from 1,378 to 820 px at 1440 px. The board-and-software list reads at 75 characters a line, where it ran to 165; the installation list's values start on one line. The signal-path diagram fills its frame, its smallest labels grow from 9 to 11.3 px at 1440 px, it lists HQPlayer Embedded among what runs on the box, and it names the same DAC as the strip, the Heed Abacus. The brand's raised ty no longer pushes its line down, and stays lowercase in capitalised labels, as in the logo.
  • [site] The landing's first screen on a phone shows the product. The first capture of the app came 1,070 px down. The bar now holds one line, the links row fades at its edge to say it scrolls, the install button comes under the title, the capture right after it — its top half on the first screen — then a shorter hook that says nothing the long one does not. The green Public beta over the button is gone; the ribbon stays, and no longer covers the bar's text or the theme toggle under 1100 px.

Fixed

  • [core][ops][ui][manual] An update that does not go through says why, under the update banner. Whatever stopped it, the banner said Update failed — previous version restored, and the reason the updater recorded — reinstall failed for any failure of the installer — reached no screen. It now gives the installer's own words — a system turned down (This system (Debian GNU/Linux 12 (bookworm)) has glibc 2.36; this core needs 2.38 or later…), a download that failed, a damaged package — or says that after the install the box was not running the new version, or that the updater could not be started. The reason stays under the banner and is read back when the Admin page is opened again or from another device, for admins only: they alone can act on it. The manual says the box tells why.
  • [ops] An installer that never arrives no longer restarts the core. When the installer cannot be fetched — no network, or an answer that is not a script, a captive portal's — nothing on the box has changed yet: the update now stops there and says so, without stopping and restarting the core.
  • [ops] The update log keeps about the last twenty updates. Each update added 10–15 KB to it and nothing ever shortened it — 749 KB on a box updated 52 times. It now keeps its last 256 KB, cut before each update, in a way a full disk cannot empty.
  • [ui] The update banner no longer acts once it is off the page. An answer that arrived after the banner was removed still set its state and the Admin tab's badge.
  • [ops] Updating the interface no longer leaves the old versions' files behind. The installer copied each release over the last one and removed nothing, and a release names its files after their content: every update left about 17 files (1.2 MB) for good (measured on the 0.9.62 → 0.9.63 → 0.9.64 packages), and a validation box held 944 files, 94 MB, for a version that ships 25. The installer now keeps two versions — the new one, and the one it replaces, which a page left open during the update still loads from — and removes the rest: the files it recorded for older versions, whatever else lies in the folders the interface ships, and the source maps older versions left. What the box writes for itself — the interface's settings file, its certificates — is never touched. The first update made with this installer records what it installs, and removes the older files at the next one; the source maps go at once. Nothing is removed when what is in place is not what the installer recorded — a copy made by hand, an older installer — nor when the new package cannot be read whole, and two installs at once wait for each other.
  • [ui] A page left open through two updates no longer opens empty screens. The box keeps the files of two versions: a page older than that asked for files that were gone, and the screen it opened stayed empty — without a service worker, nothing reloaded it. When a screen's code is missing, the page now asks the box which version it serves and, if it is a newer one, reloads onto it, once a minute at most. When the box serves the same version — a network blip, a server restarting — nothing is reloaded, and a message says that part of the screen could not be loaded, and to reload the page.
  • [ui] Reconnecting the terminal no longer stacks a second one. RECONNECT and RETRY opened a new terminal under the one still on screen, whose size kept being adjusted for nobody.
  • [ui] Edit, pressed right after the topology window opens again, keeps the editor editable. Pressed within its first moments, it was set back to read-only when the file was shown.
  • [ui] A new version of the interface reaches an open app in seconds, no longer five minutes later. After an update of the box, the app said Updating… A new version is being applied and nothing followed: the service worker relayed the app's two live streams — the player's and the dashboard's — and the version in place, kept busy by them for as long as the app stayed open, could not hand over. Chromium forced it after five minutes, and the page reloaded then, with no word (measured in Chromium 145 without a debugger attached: one changes these timings). The streams now go straight to the box, and the page reloads onto the new version within 2 to 10 s, 12 s when it has to reload itself (below) — 19 attempts out of 19, the app open during the update, opened right after it, or in two tabs. A new version already waiting when the app opens — installed while no page listened — is told too: it waited until every window of the app was closed, a reload included. Told, a new version still stayed waiting in a third of the attempts with the app opened right after an update, with nothing left in its way; a reload let it in every time, so the page now reloads itself 10 s after Updating… when the new version has not come, once in five minutes at most, and tells a version still waiting again at its five-minute check. A player stream that loses the box still spaces its attempts out — 3, 7, 15, then 31 s, as before. Safari not tried. The update to this version still meets, on each device, the service worker of the version before, which relays the streams: it applies when the app is reloaded or closed, or after those five minutes.
  • [site][ui] The third-party notices list what the interface ships. Chart.js and CodeMirror were listed as loaded from a CDN, xterm.js was missing, and Cytoscape was still there.
  • [site] The signal-path diagram's dots no longer wait in its corner. For up to 1.8 s after the page loaded, three of the moving dots sat at the drawing's top-left corner before setting off; they now start already under way.
  • [manual] The manual's links read in the light theme. On audiogravity.app they were the green that reads 2.3:1 on white; they now take the darker green (5.0:1), and the active chapter's mark the green kept for lines (3.3:1).

Removed

  • [ui] Cytoscape and cytoscape-dagre are gone. Nothing in the interface used them any more — the audio chain is drawn in SVG — and they brought lodash, dagre and graphlib into the build.
  • [core][ops] SELF_UPDATE_BOOTSTRAP_URL is gone. The updater fetches the installer from the address it carries, which its root launcher lets no caller change: the key steered nothing, and edited, it would have lied. Updated boxes lose it — the core refuses a key it does not know — and the support report shows the address the installed updater uses.

[0.9.64] - 2026-10-02

Added

  • [ui][manual] A Settings switch hides the top bar's figures, on this device. Top Bar Metrics, under Animations and on by default: switched off, the top bar no longer shows uptime, CPU, temperature and memory, and no longer handles them — they still reach the device, on the stream every live update travels on, and the System tab keeps showing them. Switched back on, it shows at once the last reading of the live stream, or dashes if there is none yet, and asks the box nothing: the box's own answer is partial, and its CPU figure covers the few milliseconds since the stream's last one. Each device keeps its own choice. The manual lists it in the Settings panel, whose figure is retaken.

Changed

  • [site] The landing promises what the app does. Profiles are no longer sold as atomic — a service that fails does not undo the switch — but by their live health bar. The service tiles' sparklines are live, not 60 s, and the throughput reference the tile quoted is dropped: it went with the INFO windows in 0.9.63, and its figures were eight times too high. The configuration editor offers a diff and an automatic backup restored in one click, not live validation. The passkeys card no longer mentions a session history the app does not keep. The login screenshot is retaken.

  • [ui] What acts as a button is a button. The Enabled/Disabled badge of an account, the four actions of the Audio Software title (refresh, download, CHECK UPDATES, UPDATE ALL), the metrics icon of the Services title, the API icon of the Settings panel, the footer's icon, the player's source dots, in the bar and full screen, and the pull tab were spans and divs made to act as buttons, with code of ours for the keyboard. They are real buttons now, and look exactly as before — measured in Chromium, on a phone and a desktop, light and dark: same boxes, same pixels. The browser itself gives them Tab, Enter and Space, and announces them. A test fails on an element announced as a button that is not one, bar the two below.

  • [ui] The CPU cards of the Performance tab draw their load in bars, as the System tiles do. Sixty samples, the time they cover written beside Load, a gap where the stream paused — and one 0–100 % scale for every core: each card used to scale its curve to its own values, so a core idling between 1 and 3 % drew as tall as one at 80 %. CPU n and its socket and core now share one line, and the chart takes the line this frees: nearly twice as tall, on a card of the same height (measured in Chromium at every width the grid gives a card, with and without the THROTTLED badge, which cuts the socket and core short on a narrow card rather than push them onto a second line). A measured zero keeps a 1 px bar, here and on the System tiles, so a core idling at 0 % no longer looks like a moment nothing was measured.

  • [ui] The volume slider sends its value four times a second at most. A range slider fires on every value it crosses, and each one became a request: 32 for a third of the course dragged in a second, 101 for the whole course in two (measured in Chromium) — on HQPlayer, two connections to the player each. It now sends the first value at once, then the latest every quarter of a second, and the value it stops on as soon as the gesture ends: 8, 3 and 13 requests for the same drags, the last one always where the slider stopped. The step buttons and the keyboard still send each press at once, and nothing is sent once a gesture has ended, so a late value cannot reach a source switched to in the meantime. Not measured on an iPhone.

  • [core] HQPlayer's volume is left to HQPlayer. Audiogravity kept it with the DSP settings and sent it back each time the HQPlayer card reconnected: the level last changed through that Audiogravity — two instances watching the same HQPlayer kept −10.2 and −5.4 dB — so a volume lowered in HQPlayer itself was set back at once. It is no longer kept nor sent back, and HQPlayer's settings file is no longer rewritten at every step of the slider. A snapshot of HQPlayer's settings that matches none of its lists is no longer taken again every time the card is shown.

  • [site] The landing's Settings view no longer offers the configuration export. Its caption still read Export the whole configuration to a file, and a note pointed at the export, which left Settings in 0.9.62. The caption now says what the panel holds; the capture itself, from 0.9.45, is still to be retaken. The push notifications card says an audio service where it said a critical service — in the app, critical is a kind of profile, and the alert covers every audio service Audiogravity follows. The HQPlayer card says that each HQPlayer keeps its own filter, shaper and mode.

  • [manual] The manual says what DSD does to the volume, and what Audiogravity writes in the topology. Listening: while a DSD track plays, the volume is set to 100 and the volume control gives way to a padlock, then the volume goes back to where it was — with the app closed too, so the listening level belongs on the amplifier; through HQPlayer or a network speaker, the volume stays adjustable. Outputs & engines: each HQPlayer keeps its own filter, noise shaper and mode. Administration: a value of the wrong kind blocks the topology's save; you can always roll back promised what no screen offers, so the advice is now to Download the file before a big change; and the topology file is no longer said never to be rewritten — the output chosen for each service is written there.

  • [manual] The Performance chapter shows what it describes. It had no figure, and never said what a CPU card holds: it now shows two cards — frequency, temperature and its colours, load in bars, governor — a latency test after a run, and the network test. The Terminal says that closing it, or the end of the session, stops what runs in it, a command started with nohup included.

  • [core] Reading the temperatures no longer holds the core up. It happened in the core's own loop: on the x86 test box, whose Apple chip exposes 60 sensors, each reading stopped the core for up to 56 ms — no request answered, no live update sent — every twenty seconds or so while the app is open. It now happens aside: 6 ms at worst, 1 ms in 99 cases out of 100 (measured on that box). Nothing changes on a Raspberry Pi, whose single sensor took about 1 ms. Callers arriving together share one reading.

  • [core] Playing no longer keeps systemd loading services. While something plays and a screen follows it, the audio chain is rebuilt every second, and each rebuild asked the box's systemd about every service the topology names — a share on the NAS, an app on the phone, Roon Server on another machine — and read the details of every known audio service, running or not. Asked about a unit it does not hold in memory, systemd loads it, then drops it again: 35 loads a second on the x86 test box while it played with two screens open, 5.9 % of a core in systemd and 2.2 % in the system bus (measured). Only the box's own services are asked now, all in one call that makes systemd load none of them, and a service's process is looked up only while it runs: playback causes no load at all, and systemd and the bus take 0.2 % of a core at most each (measured on the same box, playing, one screen open or none). The service-down watch reads its services the same way, and the check that decides at start whether to connect to Roon no longer makes systemd load a unit either.

  • [core] With the app closed, playback no longer rebuilds the audio chain every second. The player's watch read every source and laid out the whole chain — devices, links, the other machines probed over the network — every second of playback, for the DSD protection alone once no screen was open. With nobody following the player, it now reads the sources only while a DSD track plays, while HQPlayer can play on the box (its NAA or HQPlayer Embedded running) or while Roon Bridge or Roon Server runs — the sources that report none of their changes — and never lays the chain out then. MPD reports its own: its DSD is caught the moment it starts. Measured on the x86 test box, radio playing, app closed: the core from 11.8 to 2.2 ms of CPU per second. With a screen open, or a client of GET /sse/player, nothing changes.

  • [core] A sound card without a volume control is no longer asked for one again and again. A DAC whose volume is fixed — an HDMI or S/PDIF output, many USB DACs — has no volume for Audiogravity to read, and it asked again every 5 seconds while the music played: whenever the app is open, and with it closed whenever HQPlayer plays through the box (its NAA or HQPlayer Embedded), Roon Bridge runs or a DSD track plays. Each time, amixer was started up to nine times: 27 in eleven seconds of playback on the x86 test box's HDMI output (measured). The card is now asked once, and again only when its controls change, when it is unplugged or replaced, or when a DSD track starts; the HDMI outputs of the ARM test box, which have no control at all, are asked once too. While DSD plays, the protection that keeps the volume at 100 read the whole card every second: it now reads the volume only — nothing from a card without one, one amixer instead of eight; one instead of two from a DAC already at 100 (measured on both boxes). A card the core cannot open is not asked at all, and a volume read at the very moment it changes is no longer kept for a minute.

  • [core] A Roon Server declared in the topology is probed every 30 seconds, like the other machines. Its address was tried at every rebuild of the audio chain — every second while the music plays with a screen open — and each rebuild waited half a second when the machine did not answer: 510 ms to build the chain instead of 8 on the x86 test box, whose topology names a Roon Server that is not there (measured). Its state is now kept 30 seconds, as those of a NAS, MinimServer or a remote HQPlayer are, and the chain is rebuilt once a second again, as intended. When Roon Bridge on the box is connected to the Core, the Core itself is still asked at every rebuild, so Roon stopping shows at the next one.

  • [core] Keeping the profiles up to date no longer starts programs. Every 30 seconds from the box's start, app open or not, the watch behind the profiles read their services through systemctl: one read of 36 properties where it needs three answers per service, then one more systemctl for each stopped service — asking what the first read had just answered — and one for each service not installed, which started a script of the old init system on top and was asked again every time. Seven programs every 30 seconds on the x86 test box: 110 ms of CPU, 161 ms on the ARM test box (measured). Two questions to systemd over D-Bus answer now, starting nothing and making it load nothing: about 8 ms on the x86 box, 9.5 ms on the ARM box, and systemd's own share on the x86 box drops from 1.3 to 0.3 ms per second (measured, app open). When D-Bus does not answer, a single systemctl with the few properties needed takes over. A service's details no longer start systemctl for the control group of a stopped service, nor to ask whether a missing service starts at boot, and a disabled service read through that fallback reads disabled, no longer unknown.

  • [core][ops] The network view no longer starts programs, and asks the WiFi without sudo wherever it can. While a screen follows the audio chain, the box's network interfaces were read again every minute: ip was started for each of them, and sudo iw for the WiFi, whose three lines a minute went to the system journal and the core's log screen. The addresses are now read from the kernel, a WiFi that is not connected is not asked at all, and iw is first asked without sudo: on both Raspberry Pi test boxes it answers in full that way, for 2 ms of CPU where sudo took 24 (measured). Only a WiFi chip that refuses — the x86 test box's — gets the sudo call, and that call no longer writes to the journal. Screens and requests arriving together share one reading, and the support report reads the WiFi the same way.

  • [site][manual][ops] The requirements say what Audiogravity runs on, and the core installer turns down the rest. The landing, the README, EDITIONS and the manual promised Debian / Ubuntu — any recent kernel. The core is built against Debian 13's C library and needs glibc 2.38 or later — Debian 12 has 2.36, Ubuntu 22.04 2.35 — so the install went to its end there, and the core did not start (measured on the 0.9.63 binaries, x86_64 and aarch64). And installing audio software from the app looks the vendors' builds up by Debian release, which an Ubuntu box is not. They now say Debian 13 (Trixie) or later, DietPi and Raspberry Pi OS included in their Trixie release. The core installer checks both before it changes anything, and says why it stops; a box already installed on another distribution is warned, and keeps its updates.

  • [core] The Services tab's figures no longer start a program, nor make systemd load the stopped services. While a screen is open, the figures of the services Audiogravity follows — CPU, memory, disk, network — were read every 10 seconds, every 2 while one of them works hard, by starting systemctl show for the eight of them, and systemd loaded every stopped one to answer: 46.7 ms of CPU each time on the x86 test box, 72.8 ms on the ARM test box (measured). They are now asked over D-Bus — the states in one question that loads nothing, then the figures of the running services only: 9.6 and 23.4 ms (measured). The tab shows what it showed — same states, same figures, same dashes, compared service by service on both boxes. When D-Bus does not answer, systemctl takes over, as before.

  • [core] AirPlay and UPnP are no longer asked to describe themselves before every question. Each question to AirPlay or to a player that speaks MPRIS — what it plays, every second while the music plays with the audio chain on screen, or a command such as play, pause or volume — started by asking it for the description of all its interfaces, and reading it: 4.3 ms of CPU per question to a running AirPlay on the ARM test box. The description is now kept, and asked again only when the player stops answering: 1.45 ms (measured). And a player of the box that systemd says is stopped — AirPlay, MPD — is no longer asked what it plays: a stopped AirPlay answered every question with an error, 0.65 ms each on both test boxes (measured).

  • [core] The topology editor refuses what the box would leave out. Its check let anything through in a device's services, connection, network interfaces and input ports, and in the outputs Audiogravity pins for each service: a number among a service's controls was saved, then stopped the drawing of the whole chain. The editor and the box now read the file through one model: a value of another type anywhere in it is an error that blocks the save, as it already was in a device's outputs. Numbers and yes/no are read as before — "1" is the number 1, "true" a yes.

Fixed

  • [ui] The latency test's result said milliseconds for microseconds. Its four figures — minimum, average and maximum delay, standard deviation — were labelled (µs) in capitals, and the capital of µ is the Greek Μ, which reads as an M: MAX LATENCY (MS) over a value in microseconds made a 2 ms spike read as 2 seconds. The unit now follows the value, as it does for the percentiles below, and a test fails on a stat box that puts µ in its label.

  • [core] What runs in the terminal ends with it, a command ignoring the hang-up included. Closing the terminal sent SIGHUP to its shell and to everything it ran, and stopped there as soon as the shell had left: a command started with nohup, or under trap '' HUP, kept running after the terminal was gone. Whatever is left after the grace period is now killed, whether the shell has gone or not, and a close cut short — the core stopping with a terminal open — kills what is left at once instead of leaving it. Still out of reach: a command that leaves the terminal's process group (setsid, a daemon, tmux, or sudo running its command in a session of its own).

  • [core][ops] The secrets kept on the box are readable by Audiogravity's account alone. The Tidal, Qobuz and HIGHRESAUDIO accounts, the Roon token, the licence and the push notifications' private key were written readable by every account on the box (measured on the x86 box and the ARM box), and so was the accounts file the installer creates. The core now writes them for its own account only, each through a temporary file of its own: the one writer of secrets used a fixed temporary name, and a leftover readable by all passed its mode on to the secret. The installer creates the push key and the accounts file that way, and at every update narrows the copies earlier versions left, whether their module is enabled or not — each change made as the service account, so that a file swapped for a link cannot lead it to a system file. And the core narrows them again each time it starts: the manual's backup is restored with tar, which puts each file back with the mode it was archived with, and an archive made before this version brought the accounts file back readable by every account until the next update.

  • [core] The Roon token and the licence are replaced whole. Both were written straight over the previous file, so a write that failed half-way left a truncated one — a licence then read as unreadable, a token as none. They now replace the previous file in one step, as the other settings do, and the log no longer says token saved when it was not.

  • [core] Backups of a configuration no longer pile up. Every save of the audio topology left one more dated copy beside it, and nothing removed them — nineteen on a test box. The ten most recent are kept, the one just taken always among them, even on a box whose clock runs behind until it reaches the network.

  • [core] Each HQPlayer keeps its own DSP settings. The filter, noise shaper and mode kept for one HQPlayer were sent to the next one the HQPlayer card chose, as positions in HQPlayer's lists — which differ from one HQPlayer to another: 54 of the 70 filter numbers an HQPlayer Desktop 5 and an HQPlayer Embedded 6.1 share name different filters (measured), so the new one played a filter nobody had picked, with no word. Each HQPlayer now keeps its own: one met for the first time starts from the settings it has, one chosen again gets its own back, and a setting that answers after the card has changed HQPlayer is kept for the one that took it. Reset to HQPlayer defaults concerns the chosen one only; the last ten HQPlayers are remembered. API: DELETE /hqplayer/dsp forgets the chosen HQPlayer's settings only.

  • [core] Service-down and critical-temperature notifications reach a device with the app closed. Both were raised by the watch that feeds an open screen, which runs only while the app is open somewhere: closed everywhere — the moment a notification is for — a service that failed or a CPU that overheated went unannounced. A watch of their own now checks the services every 30 s and the temperature every minute, whatever the screens do, and only while a device is subscribed to notifications: a box nobody subscribed from reads nothing. A service counts as down as the Services tab shows it, one that systemd keeps restarting after a crash included, even when a reading caught it between two restarts; a service stopped on purpose does not count, nor services stopped by a shutdown or a reboot of the box. A service that stays down is announced once, one that keeps failing once per quarter of an hour at most, and a failure that happened while nobody could be told is not announced as new. The critical-temperature alert could not go out in the first quarter of an hour after the box started. The services' states are read without starting any program, in one call to systemd that makes it load none of them: 0.6 ms for 8 services on the x86 box (measured).

  • [ui] An unknown figure of the top bar is a plain dash. Before the first reading, and for good on a machine without a temperature sensor, the temperature showed --°C on the red, pulsing background of a critical one.

  • [ui] The tabs, the logs and the Audio Software badges follow the Animations switch. Switched back on, the tabs' bell, the logs' LIVE badge and Audio Software's update badges stayed still until the page was reloaded: the tabs and the logs waited for a notice nothing had sent since 0.9.4, and the Audio Software page read the setting only when it redrew.

  • [core] The DSD volume protection keeps running once the app is closed. It is started with the box so that a DSD track gets its volume set to 100 even with no screen open, and it stopped with the last screen that closed the player — MPD's own listener, which catches a DSD file the moment it starts, included. From the first time the app had been opened and closed, nothing set the volume of a DSD track to 100 any more. It now runs for good. Tried on the x86 test box: after a screen had opened and closed, a DSD128 file started from MPD's queue took the volume from 24 to 100 within a second, and the radio played after it got its 24 back.

  • [core] The sleep timer checks what plays when it fires. It trusted the last state the player had published, which could be hours old once the app was closed: a toggle on a stale playing would start music paused since, instead of pausing it. It now reads the player's state at that moment, and falls back on the last one only when that read fails.

  • [core] A service that crashes and is restarted at once is announced. The service-down notification read the services every 30 seconds, and systemd restarts a crashed audio service within 0.1 to 5 seconds (restart settings read on the x86 and the ARM test box): both readings found it running, and the crash went unannounced. Audiogravity now listens to systemd, which tells each change as it happens: on a service crashing every four seconds, the notification went out at the moment of the first crash, on both boxes (measured), and once per quarter of an hour after that, as before. Listening costs nothing while nothing changes, and happens only while a device is subscribed to notifications; the 30-second check no longer asks systemd anything.

  • [core] A WiFi network's name shows as it is written. iw prints every character beyond plain ASCII as a code, so a network named Café showed as Caf\xc3\xa9 in the network view and the support report — and the report also cut every name at its first space. Both show the whole name now.

  • [core] A topology edited by hand no longer loses the whole audio chain. One value of a type Audiogravity does not expect — a device or a service written as a bare name, a number among a service's controls, a text where a connection is expected, a value left empty (null) such as a service's label — stopped the drawing of the whole chain, and the player lost its signal path, with nothing on screen to say why. The file is now checked against the topology editor's own model when it is read: each such value is left out, the core's log names it once, and the rest is drawn. The output switching and the guided setup read the file the same way.

  • [core] Choosing an output no longer wipes a topology that cannot be read. When audio-topology.json had been broken by a hand edit — a comma left out, in a terminal —, choosing an output in the guided setup or in Guided mode wrote it over with the outputs alone: the chain described in it was gone, kept only in a dated copy beside it that no screen brings back. The file is now left as it is: the service changes output all the same, the choice is not recorded in the topology, and the core's log says why. A file holding no JSON object — a list, say — made the save fail after the output had changed, so the screen called a change that had happened a failure; it is left as it is too.

  • [core] A topology saved from the editor reaches the output switching at once. The output switching read the topology when the core started, and only then: an output renamed or added in the editor stayed unknown to it until the next restart. The drawing of the audio chain caught up with a save from the editor, but not with a file edited by hand. The core now holds one copy of the file for all its readers, read again as soon as the file changes, whoever changed it; a file caught half written leaves the last good copy in use. Asking the file whether it changed takes 2.6 µs on the x86 test box (measured): about 34 µs a second while music plays with the app open.

Removed

  • [ui] The radio card and the player's cover no longer answer the keyboard. Audiogravity is used by touch, mostly on a phone, and neither can become a button: the card holds three buttons of its own, and the swipe that removes a station does not start on a button; the cover holds the album details. The keyboard code that stood in for a button is gone. Nothing changes by touch or with a mouse, and a screen reader still announces them as buttons.

[0.9.63] - 2026-09-28

Changed

  • [core][ui][manual] Restore Backup and Remove Override apply the settings they put back. They rewrote a service's settings and stopped there: the service went on running on the settings being undone — those someone reaching for these buttons most wants gone — until its next start. They now apply them as a save does: a running service is restarted on them and watched; if it does not start, it gets back the settings it was running on and is restarted on those, and the message says so; a service already restarting in a loop keeps them and is stopped; a stopped service stays stopped. Remove Override now says when systemd refuses to reload, where it answered success. Like a save, both are refused on Audiogravity's own service while software is being installed. The confirmation says a running service is restarted, and the message shown afterwards is the core's. API: both routes answer { success, message, service, service_restarted }, 400 with the reason when the service did not start, and 409 on the core's own unit during an install.
  • [ui][site] The pipeline's green badge reads FULL RES. It read BIT-PERFECT, but what it checks is that the connection can carry the format sent to the card, not that nothing upstream altered the audio. The graph's bubble and legend, the Editions window, EDITIONS.md and the manual say FULL RES too.
  • [ui] The INFO buttons are gone. Thirteen tabs opened a window of their own explanations, most unchanged since the first version: of about 115 sentences, some 55 were at least partly wrong and 9 plainly false (checked against the code on 2026-09-28), and nine windows repeated the manual — errors included. The manual, corrected, is the one explanation.
  • [ui] The interface shows no tooltips. They are removed, every one: a button with an icon alone is named for screen readers instead, and what needs reading is on the screen — the reason the role and the enabled state of an account cannot change is written under them, in the account window. A test fails when a title comes back. Our components that had a title (windows, panels, tiles, rows) read it from a heading attribute: from title, the browser showed it as a tooltip over the whole component. A window is still named by its title for screen readers.

Fixed

  • [core][ui][ops] Sessions end when the account behind them changes. A token was checked for its signature and expiry alone: a new password, a new role, a disabled or deleted account left every open session working for up to 12 hours. The account is now read back at every request that carries a token, whatever the route — a stat of the users file, and 0.1 to 0.2 ms (measured on the x86 box and the ARM build box) — and the token refused as soon as the account no longer stands behind it: most routes asked for the API key alone, so an ended session went on working from its tab, as no one in particular; an account deleted and created again under the same name no longer takes back the old account's tokens, nor a token the role it was issued under — the built-in admin included, when the manual's recovery has the installer create it again. An open terminal checks its session every 15 s and closes once it has ended; whenever it closes, what runs in it now stops too — an interactive shell ignored the signal it was sent, and a command running in it went on after the terminal was gone. An admin who changes their own password stays signed in on the device they changed it from; the other sessions end. In the interface, a refused session sends to the sign-in page at its next request — for the page's requests, its uploads and its passkey registration alike — but only when the token refused is still the one the tab keeps: a tab whose token was replaced meanwhile, by itself or by another tab, carries on with the new one, its role as it now is, and its live updates reconnect with it; a tab never takes up, overwrites or erases the session of another account signed in the same browser. The terminal says its session ended, rather than asking to sign in. Tokens issued before the update keep working until something changes on their account. API: PATCH /auth/users/{username} returns access_token when the change ends the caller's own sessions; an ended session answers 401 with WWW-Authenticate: Bearer.
  • [core][ui] Nobody can disable their own account or change their own role, and the built-in admin keeps its role. The core refused only to disable a protected account: an admin could disable or demote themselves by the account window's own save, and demote the built-in admin. Both are refused now, with the reason; the account window greys the fields out and says why, and the status badge of your own card is no longer a button. API: PATCH /auth/users/{username} answers 400 for them.
  • [core] A guest controls what is playing, and nothing else. Many actions were open to a guest account: HQPlayer's filter, shaper, mode, connection and output, signing in to and out of Qobuz, Tidal and HIGHRESAUDIO, the radio library and its favourites, playing a station, switching the MPD output, restoring a service's systemd settings, the licence's upload, check and activation, starting the metrics and log streams. They are closed to guests; what is playing stays theirs to control — play and pause, skip, seek, volume, repeat, shuffle and the sleep timer, on the box, a network speaker (which it could not, before) or HQPlayer. A test now reads every route of every module and fails on a write route that says nothing about who may call it.
  • [core] An idle terminal no longer sends twenty empty messages a second. The terminal forwarded each empty read of its shell's output, one every 50 ms (measured on 2026-09-28): it now sends only what the shell writes.
  • [ui] Restart Core and Reboot OS reconnect by themselves again. Waiting for the core to come back, the page asked /health at the interface's address — a page the interface does not serve, answered 404 on an installed box — so it waited to the end of its countdown. It now asks the core.
  • [ui] A service's detail window stays live. It showed the service as it was when opened, while its metrics moved on around it. It now follows the service's current state.
  • [core][ui] A failed install, update or removal shows as such on its card. The core recorded the failure and then published the package's state read from the system, so the card never showed ERROR, and its REPAIR button — offered to a required package that failed while installed — could not appear. A failure now stays shown until the next operation, or until the version on the box changes (the package was put right from a terminal); the card keeps saying a service is not configured, or has no web password, when the previous version is still installed, and the INSTALLED filter keeps it. API: status can now be error in /packages/* and in the package_state event.
  • [core][ui] The THROTTLED badge of the Performance tab appears. Each CPU card compared the throttle counters with its own previous reading, and the tab rebuilds its cards on every reload: there never was a previous reading. The core now compares them at each tick of its monitoring, forgetting the last tick of an earlier run, and hands the answer out with the metrics. A Raspberry Pi exposes no counter; its under-voltage alarm marks every core instead (tried on the ARM build box under load on 2026-09-28: alarm raised with the kernel's Undervoltage detected!), and a quiet alarm says nothing, since the Pi also slows down for heat without reporting it. API: the sysinfo event carries cpu_throttled, null when nothing was measured.
  • [core][ui] The latency test runs where it is told, and its figures are measured. The CPU affinity was sent as cpu_affinity while the core reads affinity: it was dropped, and the test never ran on the cores asked for. The percentiles are now read from the histogram rather than estimated, and the progress counts every thread. The form stays within what the core accepts (1–16 threads, 1–99 priority, at least 1 000 loops) and names a setting out of range before anything is sent. The test's title read LATENCE TEST.
  • [ui] Apply All applies the governor last chosen. It applied the one CPU 0 had at the last reload of the tab.
  • [ui] Three colours were missing. The NOT INSTALLED badge of Audio Software, the state badges of the Config tiles and the added and removed lines of the raw diff had no colour: their CSS rules named classes the interface did not write.
  • [ui] The licence panel's portal links accept only web addresses. The License portal and Download .lic links took the address the licence server gave as it came, a javascript: one included.
  • [ui][site] The Editions window says what each edition holds. It classed the configuration editor in Pro, where the code and EDITIONS.md put it in Starter, and it now matches EDITIONS.md, itself corrected: push notifications are sent for four events, not for "playback events"; choosing the output belongs to Pro; the configuration check is a basic syntax check; the DSD volume protection, which every edition runs, is listed in Starter.
  • [ui] Buttons that are not buttons answer the keyboard. A clickable badge, the icons of the Audio Software and Services titles, the API icon of the Settings panel, the radio card, the player's cover and source dots and the pull tab could only be clicked: they are now reached with Tab and pressed with Enter or Space — on the element itself, so a key pressed on a radio card's own buttons still does what that button does. A test fails on an element announced as a button that does not answer the keyboard.
  • [manual] The manual describes the app as it is. Every sentence was checked against the code on 2026-09-28. Among the corrections: what each role can do, and that changing an account's password, role or state signs it out everywhere; that a session lasts 12 hours, the installed app included; the Config editor, whose guided mode and CONFIGURED badges are for admins; Audio Software; a rewritten Performance chapter, with the latency and network tests; Systemd, whose Validate button does not exist; the licence; the Pipeline graph, its colours and commands; the profiles, whose switch is not undone when a service fails, and starts nothing when an essential one cannot be stopped; and a Roon queue, which is changed from the Roon app. The sentence saying every tab has an INFO badge is gone with the badges, and the seven figures that showed one are retaken without it: the Audio Software, Config, Services and Audio Pipeline tabs, the network test and both licence panels — the Config tiles with their state in colour.

Removed

  • [ui][manual] The cockpit's LOW POWER badge and Device Battery tile are gone. LOW POWER turned the app's animations off, refreshed its live figures half as often and slowed two of its checks, on the device showing the app — the box never depended on it. It came on by itself from the battery level in Chromium browsers only, and after a switch by hand it took three taps to turn off. The Animations setting already turns animations off. The battery tile existed for it, and the app no longer reads the battery at all. The cockpit's timer table loses its Effective column, which only differed from the interval under LOW POWER.

[0.9.62] - 2026-09-27

Added

  • [core][ops][manual] A command applies a change made by hand to the audio services and profiles. Between two versions, support may still have to change a box's profiles — to try one, or to fix one at a customer's. Over SSH, copy /etc/audiogravity/audio-config.json, edit the copy, then run sudo /opt/audiogravity/ag-apply-audio-config <copy>: the core checks it with the very rules it loads the file with, and a file it would not load is refused, entry by entry with the reason, with nothing changed; a file it would load takes the file's place in one step, the core is restarted, and the command waits until it runs the profiles. --check checks without applying. A copy, as visudo works: parts of the core re-read the file as soon as it changes, before any check; should the core not start on the copy, the previous file is put back. On a box, the command only reads and replaces /etc/audiogravity/audio-config.json, whatever the core's .env names. The change lasts until the installer runs again — the next update, or a reinstall — which puts the shipped file back and keeps the replaced one as audio-config.json.previous. The manual's troubleshooting chapter has a new entry, Changing the profiles between two versions. The command runs as root and reads files the audio service's account may write: it refuses a symbolic link, and never goes through a proxy — its requests carry the API key.
  • [site][ui] The manual's commands are coloured, and a button copies each one. In the manual — on audiogravity.app and in the app's Manual window — the command, its options, quoted text, variables and comments each take a colour, as JSON keys and values do, and every code block carries a button that copies it, turning to a check mark once done. The copy leaves without its final newline, so the last line waits for Enter. A block holding values to replace — an address, a password, the example NAS share — has no button, and the manual says so; the NAS mount is now one step per block, and one variant per block. The site and the app colour a block identically, byte for byte, which a test checks on the whole manual.
  • [site] Who I am shows the system it names. A photo of the author's listening system, its Naim logo lit green, stands beside the note, taking a third of its width, the words centred against it in the other two; on a phone it heads the note. The file carries no metadata, so no location.

Changed

  • [core] One set of rules reads audio-config.json, and it refuses what used to slip through. The core loaded the file with one model and checked it with another, stricter one, read by another parser: the standard one reads NaN, Infinity and a lone surrogate, which the loader refuses (measured on 2026-09-26), so a file could be approved, then dropped at the core's start with an empty Profiles page. Both now share one model and one parser. A key the model does not know is refused — stpo for stop used to leave a profile stopping nothing — except the keys of earlier versions, accepted and read by nothing (appconfigfile, depends_on, topology_file, added_by_upgrade); so is a service a profile both starts and stops, which left the profile partial for good. A missing configuration file of an installed service is now a warning: it says something about the box, not the file. API: POST /config_validation/validate reads its body as the core reads the file — what it cannot read answers valid: false with a json_invalid error — and reports the warnings above.

  • [core][ops] The audio services and profiles are delivered with each version. audio-config.json was copied once, at the first install, and afterwards only received what an update knew how to add: a box installed before a change to the shipped file never got it — the ARM build box kept an MPD profile that needed the UPnP bridge, disabled at boot, so its MPD tile could not be active after a restart. The installer now replaces the file with the one its version ships, at every install and update, as it does for the software registry. The file it replaces is kept beside it as audio-config.json.previous, with its date; a file already identical is left alone, and a self-update that rolls back puts both back as they were. A change made by hand therefore lasts until the installer runs again — the next update, or a reinstall. The copies the retired import and export left on the box are removed (/etc/audiogravity/audio-config.json.backup.*, /opt/audiogravity/core/audio-config-export-*.json). API: GET /profiles/configuration no longer returns added_by_upgrade. The manual's backup and restore no longer counts these files as yours, and leaves Audiogravity's own files out of a restore.

  • [core][manual] The profiles are named after what they run, and a profile is critical when it starts a critical service. Each name joins the software it starts — Roon Bridge + HQPlayer NAA, MPD + HQPlayer Embedded — and each description says what the profile is for: Roon plays directly to this box, Local library through HQPlayer, or MPD directly — one at a time; HQPlayer-Signalyst becomes HQPlayer NAA. A profile's criticality — the orange left edge of its tile, and the Critical Profile confirmation, This will stop other audio services — is derived from the services it starts, the NAA and HQPlayer Embedded being critical, instead of being written in the file, where it had drifted: Roon Bridge + MPD + HQPlayer NAA, which starts the NAA, now has it, and Stop All, which starts nothing, no longer does — it keeps an ordinary confirmation. A critical written for a profile in audio-config.json is ignored. The file opens with a line saying Audiogravity manages it, and how to change it: on a copy, applied by the support command.

  • [site] The landing's text comes in six sizes, and nothing on it is drawn under 11px. The stylesheet set 38 different font sizes, five of them under 9px, and 1,457 characters of the page were drawn at 9px — the tier badges, the chips, the section labels (measured on 2026-09-26, at 1440px). Text now takes one of six steps — 11, 13, 15, 17, 20 and 28px — and the big titles and figures four display sizes of their own, the fluid ones following the width of the screen; only the brand's ty, a command named in a sentence and the currency sign are sized from the word beside them. Labels and badges move from monospace to Inter; monospace is kept for what is read character by character, the install command and the configuration file. The open-source line under the pricing lead, an inline 0.9em at 70 % opacity, read at 3.1:1 on its grey (4.2:1 in the dark theme); it now takes a step of the scale and the lead's colour, at 6.1:1 (7.5:1). A test holds the scale, the fonts and the use of monospace: a new Landing workflow runs it on each push that touches the landing, and the Manual workflow now runs its own tests only, so a stylesheet regression no longer holds the manual back.

  • [site] The site serves its own fonts, and the online manual is drawn in Inter. Every visit fetched Inter and JetBrains Mono from Google Fonts, handing the visitor's IP address to Google. They now come from audiogravity.app itself — the same files, byte for byte, as the interface's, plus Inter's italic for the landing's titles — and the two faces of the first screen are preloaded. The manual's pages named Inter without ever loading it, so they were drawn in the system font; they now show Inter, like the landing. The licence texts travel with the fonts, and the third-party notices no longer say they come from a CDN.

  • [site] The pricing shows two editions, with the trial above them. Three tiers drew the 30-day trial as a third card at €0, beside the Starter Edition's €0, and left the Pro card last in the row, though the trial is a period every install goes through, not an edition to choose. The section now reads Two editions. Lifetime license.: a band over the Starter and Pro cards says every install starts with thirty days of Pro, and Editions / Plans opens on the same idea.

  • [site] The hero's kicker reads Now playing. The audiophile streamer · Now playing, where it said Live.

  • [site] The landing's carousel opens on HQPlayer. The first screen every visitor sees now shows UPnP feeding HQPlayer Embedded in DSD128, and the DAC receiving it, where it showed HIGHRESAUDIO; HIGHRESAUDIO, Qobuz, Tidal and the radio follow, each with its caption and annotations.

  • [manual] Fourteen figures show the interface as it is now. Most were taken on 2026-09-16, and the interface had moved since. Retaken on 2026-09-27, at the same iPhone 13 mini size: the Audio Software tab, where MPD has no Uninstall; the fullscreen player, the queue, the cast to a network speaker and the Qobuz grid, with the add-to-playlist button; the playback bar without repeat and shuffle, in the queue and the Pipeline; the stream-origin badge, no longer a pill; both HQPlayer cards; the HQPlayer Embedded install window, with its 30-minute notice; the Services charts, drawn from what was measured. The first-run panel and the network test are whole again, where the playback bar covered their end — the packet loss the text points to among them. Every playing figure is real playback, the busy output included: another player held the DAC, where that figure used to be staged. Five alternative texts name what their picture now shows.

  • [ui][ops][site] The app carries its manual, and reads it from the box. The Manual window fetched every chapter and figure from audiogravity.app: a box without internet access had no manual, and every box showed the website's latest edition rather than the one matching its version. Each build now copies the manual of the site repository into the app (ui/scripts/sync-manual.js) and the box serves it: the window asks audiogravity.app for nothing, and the page no longer accepts images from it. A chapter read once opens from the device's cache until the next version, offline included. A build stops when the manual is missing, and release.sh ui refuses a landing repository that is dirty or behind origin/main, so the copy is the release's. A fix to the manual published on the website reaches the app with the next version: the contents page, the FAQ and the landing's FAQ say so, where they promised a manual "always up to date" in the app.

  • [ui][site] The trademark notice appears once wherever the manual is read. It closed each of the thirteen Markdown files, and read as part of every chapter. It now sits in the footer of each chapter page on the website, under each chapter in the app, and in the contents page GitHub shows first. The app and the site's generator read its words from that page; the generator refuses a chapter that repeats it, and an app build refuses a manual whose contents page lost it.

  • [ui] The manual's figures are retaken from the live interface, beside the published ones. tools/manual-shots/ replaces tools/shoot-manual.cjs, which drew the figures from Storybook stories at 390×844 and double density and wrote them straight over the manual's images: none of the thirty published figures came from it any more, and running it would have replaced twenty of them. The new tool takes each figure from the dev instance in the manual's own format, an iPhone 13 mini at 375×812 and density 3, and writes nothing into the manual: a review page sets each new figure beside the published one, with the share of pixels that changed. What a box cannot be put into without consequence, a trial licence or an update on offer, is staged in the capturing browser only, and the licence panel's Device ID and Order ID are blurred. The figures taken from what plays now wait for --playback. A test fails when a figure of the manual has no recipe; the five recipes lost with the session that wrote them are written again.

Fixed

  • [ui][site] The manual's paragraphs and rules have their spacing back. In the app's Manual window and on the website, two paragraphs read as one and a rule sat on the text above it, since the manual was first published: both pages reset every margin, and the manual's styles gave paragraphs and rules none back. They now have their spacing, and the contents page's version line keeps its own.
  • [core] A mistake saved in audio-config.json no longer takes the audio services away from the Services tab. The list of units the tab may drive is read from that file, and re-read as soon as it changes: a file saved with a mistake left only Audiogravity's own units in it, so starting or stopping an audio service — and switching profiles — was refused until the file was corrected. It now keeps the last good read, as the services' monitoring already did.
  • [ops] An install cut short no longer risks leaving the software registry half-written. The installer copied packages-registry.json straight over the one in place, so a full disk or a power cut during the copy could leave a truncated file for the core to read. It is now written beside it, then renamed over it in one step — as audio-config.json is.
  • [site] The comparison table and Five layers drew every word at 16px. Both asked for sizes named in the interface's vocabulary, which this stylesheet never declared, so every word fell back to 16px, ever since the two blocks were added. On a phone, the table's header broke AUDIOGRAVITY over two lines. They now take the sizes they were written for, and the header holds on one line: 38px tall on a 390px phone, where it was 67.
  • [ui] Syntax colours are readable in every theme. The colours of the configuration editor, the event viewer of the System tab and now the manual's code blocks fell as low as 2.2:1 against their ground — numbers and options at 2.6:1 in Gravity's light theme — under the 4.5:1 the interface holds its text to; every theme and mode had some. Sixteen values are raised, and the check that guards the text colours now guards these too.
  • [ui] The HQPlayer card names a network HQPlayer Desktop, and its Forget button stays inside the card on a phone. The card of an HQPlayer on the network read HQPlayer 5.28.1, while the card of the one running on the box read HQPlayer Embedded 6.2.3 and the manual promised Desktop or Embedded: it now reads HQPlayer Desktop 5.28.1. While HQPlayer Embedded runs, the card offered Forget 10.0.4.200:4321 beside DSP and Web interface, and on a phone the button ran out of the card, which cut it off — from 320 to 430 px wide, 420 px of buttons for a card of 270 to 380 (measured on 2026-09-27), since Web interface arrived in 0.9.59. The button now reads Forget, the address staying in the sentence above it, and the buttons go to a second line when the card is too narrow. Web interface is renamed Web UI: with it the three buttons hold one row from 375 px wide, and the manual names it WEB UI.
  • [ui] Copying no longer says Copied when nothing was copied. On a box served over plain HTTP, the fallback that copies text reported a refusal by returning a value nobody read: the device ID in the licence panel, the logs and the support report showed their success message with the clipboard untouched. They now say the copy failed.
  • [manual] The fix for the memory dash on a Raspberry Pi works as written. Memory shows a dash for every service told the reader to remove cgroup_disable=memory from cmdline.txt — where it is not: the Pi adds it ahead of the file's contents — and offered cgroup_memory=1, which the kernel does not know. The entry now gives the procedure tried on a Pi on 2026-09-26: one command adds cgroup_enable=memory, which takes over, the file is checked, and only then does the box restart, in a block of its own. It also says how to undo it.
  • [site] The third-party notices gave ffmpeg the wrong licence. They listed it under the LGPL. The build Debian and Raspberry Pi OS ship — the one installed on the box — enables FFmpeg's optional GPL parts, and FFmpeg's licence then becomes the GPL 2.0 or later: the notice now says so, carries the GPL text, and names both of its uses, Tidal streaming and the conversion to FLAC for HQPlayer.

Removed

  • [core][ui][manual] Settings no longer exports or imports the audio configuration. audio-config.json — the audio services and the profiles built on them, behind the Services and Profiles tabs — becomes Audiogravity's own: it defines the profiles it covers. An import would overwrite that file, and the export only existed to be imported again, so Export Configuration and Import Configuration are gone from the Settings panel. API: GET /profiles/configuration/export-file and POST /profiles/configuration/import-file are removed; GET /profiles/configuration stays. The manual's Settings figure is retaken without them.

[0.9.61] - 2026-09-25

Added

  • [core][ui] Your Qobuz playlists are filled, opened, played from any track and edited from the box, as your HIGHRESAUDIO ones are. The Add to playlist button now shows for the Qobuz track playing in the full-screen player, on Qobuz album covers and on the album rows of a search. Every Qobuz playlist has an open button and a page of its own, and a tap on one of its tracks plays the playlist from there. On a playlist the account owns, a track comes out with one tap, the pencil renames the playlist and edits its description, the bin deletes it; New playlist, the first tile of Mine, creates one. Only the account's own playlists are written — the box checks their owner again before every write — and one the account merely follows stays read-only, like the selections Qobuz publishes. A playlist is created private: Qobuz makes it public unless told otherwise (measured on 2026-09-24), which shows it on the account's public profile. Nothing goes in twice: Qobuz keeps a track added twice, and its own guard refuses a whole album as soon as one of its tracks is already there, so the box reads the playlist first and sends only what it lacks. Removing a track takes out every copy of it. Qobuz lists a new playlist about half a second after creating it (measured on 2026-09-24); the box shows the list as it will be meanwhile, the new playlist at its head, where Qobuz puts it. Played from a track on HQPlayer Embedded and on MPD on the x86 box: the queue starts on the track tapped, with the rest of the playlist behind it. API: /library/playlists* and /library/playlist-tracks take src_qobuz; in /library/qobuz-playlists?type=mine, a playlist the account owns comes back as mine:<id>; start_index and start_item_id of POST /library/queue work for Qobuz albums and playlists.

Fixed

  • [core][ui] Connecting the Qobuz account works on a slow line, and says why when it cannot. To sign in, the box downloads Qobuz's web player — 9.2 MB, which Qobuz serves uncompressed even to a browser asking for compression (measured on 2026-09-25) — to read the app credentials in it. It gave up after 15 s, which failed any line under about 5 Mbit/s; it now waits as long as the download keeps moving — up to 2 minutes, giving up after 30 s without a byte. The sign-in window now opens on the click and is pointed at Qobuz once the address arrives: opened after the box's answer, it was blocked by the browser beyond a few seconds — Chromium blocked it, and the new-tab fallback with it, 6 s after the click (measured on 2026-09-25). A failure is shown on screen with its reason — Qobuz too slow, the connection cut — where it went to the browser's console only, and the reason itself was empty for a timeout. An answer without an address no longer leaves the button on Waiting for login….
  • [core] Signing in to Qobuz takes a quarter of the memory, and no longer holds the core. The web player was decoded into text to be searched — 36.5 MB of text for its 62,116 non-ASCII characters — and the core grew by 46 to 50 MB during a sign-in. It is now searched as the bytes it arrives in, read once: 11 MB (measured on the x86 box). The search also held the core's event loop — playback streams included — for 160 to 240 ms, one of its patterns being tried at every position of the file; about 30 ms now. An answer over 64 MB is refused as it grows.
  • [core] A Qobuz playlist longer than 500 tracks is read whole. Qobuz hands out a playlist's tracks 500 at a time, whatever the limit asked for (measured on 2026-09-24), and the box read a single page: the playlist, and a play of it, stopped at the 500th track without a word. The tracks are now read page by page, to the end.
  • [core] A Qobuz playlist deleted elsewhere is said to be gone, not Qobuz to be failing. Playing it answered Qobuz API error (404) as a service failure (503); playing or opening it now answers that the playlist no longer exists (400).
  • [core][ops] Restarting Audiogravity, updating it or restarting the box no longer cuts an install short. A stop of the core killed an install still running — apt and dpkg with it, since they run inside the core's service — and left the package manager interrupted: nothing more could be installed until sudo dpkg --configure -a had been typed on the box. Measured on 2026-09-24 on a service set up like the core's: stopped 2 s into a package's install script, it was gone in 38 ms and dpkg was left interrupted. While software is being installed, Restart Core, Reboot OS and the update of Audiogravity are now refused with a message naming what is being installed, and so are a stop, restart or reload of the core's own service from Services. A restart of the core counts what the core itself installs; a reboot or an update counts any installation on the box — from the Terminal tab, over SSH, or the network-share and USB-library set-up. What decides is dpkg's own lock, so a download, a simulated install or a vendor script stuck elsewhere never blocks a restart. A stop that comes anyway — systemctl, a shutdown — lets the installation finish, for up to four minutes (HQPlayer Embedded takes 51 s on a Raspberry Pi 4, measured on 2026-09-22), then goes on; an operation that has not yet reached the package manager — a download, a request waiting for another — gives up instead of starting, and says so. The service's stop limit is 20 s instead of 10, and the core asks systemd for more time only while an installation is finishing: measured on 2026-09-24, an install whose script ran 40 s went through a stop with a 20 s limit and ended installed; without that request, the same stop was killed at 20 s. A stop with nothing to finish no longer waits on an open page either: 5 s, where the service used to be killed when its limit ran out.
  • [core][ops] An install no longer waits on a screen nobody can see. Audiogravity installs packages with apt's no-questions settings (DEBIAN_FRONTEND=noninteractive and its companions), but sudo removed them on the way, so apt never received them (measured on 2026-09-24). Where the core had a terminal type in its environment — a core started from a terminal, as in development — a package that asks a question at install time drew its screen on a terminal nobody was looking at and waited for a key forever: HQPlayer Embedded froze on its licence screen, and the core's next restart killed it and left the package manager interrupted. A box's own service has no terminal type, and its installs went through (measured on the same box, with the same package). apt now keeps those settings — a sudoers env_keep rule for apt-get, installed with the core — and commands run as root never receive the terminal type. The vendor's licence is still shown by Audiogravity before the install, as it has been since 0.9.59.
  • [core][ops][manual] An interrupted package manager is named before anything is downloaded, with the command that repairs it. When an earlier operation had been cut short, an install downloaded its package — 121 MB for HQPlayer Embedded — and then failed with Installation failed, apt's reason visible only in the log window, and worded for someone at a terminal. The core now looks first: an install, update or removal is refused before anything happens, with the repair to run — sudo dpkg --configure -a — which the Terminal tab (System, admin) may now run as root: that exact command, no other use of dpkg. If another installation is running at that moment instead, the message says which and asks to try again once it has finished, rather than calling it an interruption: dpkg keeps the same journal while it works as after it was cut short, and only its lock tells the two apart. An operation is also refused while Audiogravity is stopping or updating itself. The manual's troubleshooting chapter has a new entry, Audio Software: "the system's package manager was interrupted".
  • [core] The version offered for HQPlayer Embedded and the NAA is the vendor's own, and it is asked again after each install. The Audio Software page could offer an older version than the one just installed: Older version offered, HQPlayer Embedded 6.1.0-4 installed and 6.0.2-3 offered (seen on 2026-09-24). For a package downloaded from its vendor, the list asked apt as well as the vendor, and apt answers such a package with the version already installed — an answer the list could keep until the next check. The list now asks these packages' vendor only, as the package's details already did, and the version on offer is asked again after every install, update or removal, in the background, instead of at the next daily check.
  • [core][ui] A service restarting in a loop no longer holds every profile on PENDING, and Stop all stops it. On 2026-09-24, HQPlayer Embedded saved from the Systemd tab with real-time scheduling at priority 0 failed at every start — the kernel refuses that pair — and systemd restarted it every 2 s for two hours. Between two attempts such a unit reads activating: every profile naming it — all but AirPlay — read as switching, and its tile stayed on PENDING for good, since the page only lets go of a switch on a final state. A unit that systemd keeps restarting after a failure now reads as failed — on the services, the profiles and the metrics alike — while one restarting after a clean exit keeps systemd's own state. A profile switch now stops a service that is starting, reloading or failing, where it stopped only a running one: even Stop all left the loop running. A profile tile gives up waiting after 75 s and shows the profile as it stands, and a switch whose end reaches the page before the answer to the click is settled at once. The three restart sub-states systemd 254 added are known, where they read unknown.
  • [core][ui] Settings a service cannot start with are refused before they are saved. systemd-analyze verify, which a save relied on, passes every one of them (measured on 2026-09-24): the service then failed at start while systemctl answered success. The Systemd tab now refuses, before anything is written and with the reason: a real-time policy (fifo, rr) with priority 0, and a priority above 0 without one (the kernel's status 214); a CPU affinity naming a CPU the box does not have (215) — a CPU isolated at boot is the box's own, and pinning a player there stays allowed; a memory limit (MemoryMax, MemoryHigh) under 16 MB; an open-file limit under 1024, the default every service gets (203). A limit of 1 KB is not a slow service but a frozen box: measured on 2026-09-25 on the x86 box (systemd 257), MemoryMax=1K hung systemd itself for two minutes, every service with it — systemd starts a service through a copy of itself born inside the service's limit, which the kernel can then neither feed nor stop. Every service is started that same way, so one saved with such a limit would do it at each of its starts, at boot too. Validate gives the same answers. The kernel's refusals were measured on the x86 and the ARM box alike.
  • [core] A save that stops a service from starting is undone, and a stopped service stays stopped. A running service is restarted on its new settings and watched for 3 s; if it does not stay up — or systemd refuses to start it at all, as with a task limit of 1, where HQPlayer Embedded could not even be launched (measured on 2026-09-25) — the previous settings are put back, byte for byte, and it is restarted; the answer says so, and whether it runs again. It used to answer Properties updated successfully over a dead service. A service that is not running — stopped by a profile, or dead after a failure — is no longer started by a save, which overrode the profile that had stopped it: the settings apply at its next start. One already restarting in a loop keeps the new settings, since the old ones failed too. A service that does not start in the end is stopped, so that systemd stops restarting it. systemd is given the time it takes to stop and start a service — HQPlayer Embedded may take 15 s to stop — instead of the 5 s after which a slow stop read as a failed start. Checked live on HQPlayer Embedded: it was back on its previous settings, running, 4.5 s after the save.
  • [core] Restore Backup works on a box, and brings back what the last change replaced. Backing up and restoring a service's settings used sudo cp and sudo mv inside the systemd folder, which the account Audiogravity runs under on a box is not allowed (checked on 2026-09-24): no backup was ever made on a box, and Restore Backup always failed there — only a development box, whose account may do anything, ever saw them work. Both now go through /tmp, like the settings themselves. The backup is written once the new settings stay: a save refused or undone no longer replaces it.
  • [core][ui] FAILED shows in red, and only on the profile in use. A profile one of whose services had failed read IDLE, like a stopped one, with a small badge as the only sign, and a failed service read FAILED in grey; both read FAILED in red, and the dot of a failed service in a profile's details, which had no colour at all, is red. A profile reads FAILED only when it is the one in effect — a service it starts has failed, or one it stops keeps being restarted, and everything else is as it describes: any profile naming a failed service used to, even one never activated, even one that only stops it — killing mpd on the x86 box turned ten tiles of eleven red, for one profile in use. For the others a dead service counts as not running, and shows in their failed count only. Activate on the red tile starts again what failed.
  • [core] A service no longer reads enabled as its state once a minute. The reader of a service's details reused one variable for its running state and for whether it starts at boot: each time its 60 s cache ran out, a service read enabled or disabled where its state belongs — HQPlayer Embedded did on 2026-09-24.
  • [ui][lic][site][manual] After reinstalling the operating system or moving to another machine, you are told to write to [email protected], not sent back to the licence portal. A reinstalled system gives the box a new Device ID, and a licence belongs to the Device ID it was activated on: the portal hands a .lic back only to that one (/ls/portal/download), a kept .lic is refused, and a key already activated answers No license found for this key on this device. The manual, the licence panel's About Licensing and the emails delivering a key or a licence file said otherwise. They now keep the key and the portal for a reinstall of Audiogravity on the same system, where they work, and send a new system or a new machine to [email protected], with the Order ID and the new Device ID. The manual's troubleshooting chapter has a new entry, Licence not recognised after a reinstall or a new machine, with the three messages the customer meets. The manual's backup restore leaves the licence file out: brought back onto a reinstalled system it was refused, and kept even the 30-day trial from starting (read in the code, not measured on a box). The emails give one address, support@ — the purchase email also gave contact@ — and no longer speak of a support team; the licence portal's support link and the site's FAQ point there too. The licence panel takes the address from the licence server's Contact email setting, set to support@ on 2026-09-25, and leaves the line out when there is none.

[0.9.60] - 2026-09-23

Added

  • [core][ui][manual] The track playing now, or a whole album, goes into your HIGHRESAUDIO playlists from the box. An Add to playlist button sits beside the title in the full-screen player, and next to the star on HIGHRESAUDIO album covers and album rows in search. It opens the account's own playlists with New playlist at the top: pick one, or name a new one and the item goes into it. The box reads the playlist before writing and sends only what it lacks — HIGHRESAUDIO keeps a track added twice, and a later removal then takes out every copy at once (measured on 2026-09-23) — and the message says what happened: added, already there, or an album added in part. An album goes in as the list of its tracks, never through HIGHRESAUDIO's album route, which inserts the album's last track twice (two albums out of two). A read of the playlist that fails stops the add rather than passing for an empty playlist, and a playlist that no longer exists is refused: HIGHRESAUDIO itself accepts a write to a playlist number that does not exist. A playlist just created is listed at once, although HIGHRESAUDIO's own list shows it only seconds later. Purchases (the Vault) and playlists are not offered the button; Qobuz and Tidal are not written yet. New routes GET|POST /library/playlists and POST /library/playlists/add; neither write is retried by the interface, so a lost answer cannot make a second playlist.
  • [core][ui][manual] A HIGHRESAUDIO playlist opens on a page of its own, and the account's own playlists are edited there. An open button sits at the bottom left of every HIGHRESAUDIO playlist cover and at the end of its row — the cover itself still plays it. The page lists the tracks with their durations and counts them itself (HIGHRESAUDIO's own count lags behind the content: 0 for a playlist of 3, measured on 2026-09-23); Play and Queue take the whole playlist, and a tap on a track plays it from there, the tracks before it left out, on every output. On the account's own playlists a track is removed in one tap — every copy of it, as HIGHRESAUDIO does — the pencil renames the playlist and sets its description, the bin deletes it after a confirmation; on HIGHRESAUDIO's selections each track carries Add to playlist instead. New playlist, the first tile of Playlists › Mine, creates an empty one. A playlist just created, renamed or deleted is listed as it now is at once, everywhere, while HIGHRESAUDIO's list catches up. The account's playlists are read strictly for the page and the queue as they were for the writes: one deleted — whose tracks HIGHRESAUDIO keeps readable — or unknown is refused, and a failed read is an error, never an empty playlist. The dialogs of the page are drawn above the whole app, not under its top bar and player. New routes GET /library/playlist-tracks (open to guests, like the rest of the browse), POST /library/playlists/remove, PUT /library/playlists (a description left out is kept, an empty one clears it) and DELETE /library/playlists; POST /library/queue takes start_index and start_item_id — the track is looked for again when it plays, and a track that left is refused — for HIGHRESAUDIO albums and playlists, and refuses them anywhere else. Playlist ids must be numbers before anything reaches HIGHRESAUDIO. None of the writes is retried by the interface.
  • [core] The player names the track it plays, not only the source it comes from. content_item_id, beside content_source_id in the player state, is the track's id on its streaming service — read off the address the box handed MPD, HQPlayer or a network renderer. Measured on MPD: it reports the box's own address, not the service's after the redirect, so the id is there. null for a local file, a station, a media server, or a playback started outside Audiogravity.

Fixed

  • [ui] On a phone, the last lines of a list no longer stay under the mini-player; in a low window, the tabs can be reached again. Two layout heights were declared as a bare 0 for phones and low windows, and calc(44px + 0) is invalid: every declaration adding them was dropped without a word. Measured on 2026-09-23 in Chromium: a 1000×450 window with the tabs as a bar put the bar at the top of the screen, under the top bar, none of its eleven tabs reachable; the library of a phone kept 0px for the mini-player. They are 0px now, guarded by a test that refuses a bare 0 for any layout height, width or offset. Measured the same day, and fixed with them: the card grids of the other tabs kept only 12px at the bottom in portrait; the library reserved nothing at all, top or bottom, with the phone on its side; and in a low window the content of a tab started under a tab bar. Content that runs to the bottom of the screen now keeps --bottom-clearance free — the mini-player's height while it shows, and once it is folded the 44px touch area of its pull tab (--pull-tab-height, which the pull tab now reads too), which used to cover the lower half of the last row. A normal computer window is unchanged.

[0.9.59] - 2026-09-23

Added

  • [core][ui] HQPlayer plays what it cannot decode, because Audiogravity converts it on the way out. HQPlayer reads a short list of formats, and an Apple-encoded library is not on it: it answers Play result="Error" on an ALAC file without so much as opening it, whether the file is served over HTTP or handed to it by its path (measured on HQPlayer Embedded 6.0.4). A library ripped with iTunes or Apple Music is almost entirely M4A — the media server on the lab network answers with 185 M4A files out of 200 — so choosing HQPlayer as the output meant losing nearly all of it, and the manual could only advise converting the lot by hand. Audiogravity now converts each track to FLAC as it hands it over, losslessly, leaving your files untouched: a CD-rate track takes about two seconds and the result is the same size as the original. The conversion is a complete file, not a stream, and that is what lets you move inside the track — asked for 180 seconds in, HQPlayer is there. Three paths are covered: a file in your own library, a track a media server publishes (MinimServer, Plex — the same album was accepted from the box and refused from the NAS until now), and an internet radio station, which is transcoded as it arrives for 0.8 % of a processor core. AAC is what many of the best-sounding stations broadcast in, with no equivalent elsewhere, and they were refused outright. DSD is never converted: turning it into PCM behind your back is not something a music player should decide, so those formats stay refused with the reason, as does an address Audiogravity only forwards. An album whose tracks must be converted starts on its first one and fills in behind it, because HQPlayer inspects every entry before accepting it and that inspection is what triggers the conversion — queued in one go, a sixteen-track album would have been half a minute of silence. The conversions are kept in a bounded space on disk, sized from what the box actually has free, emptied at every start, and remade on demand.
  • [core][ui] HQPlayer Embedded installs from the Audio Software page, on a line this machine can run. Signalyst publishes it for Debian 13 on Intel and on 64-bit ARM, and Audiogravity now offers it beside the adapter it has always installed. The vendor keeps two major lines side by side and only you know which one your licence covers, so the box does not pick for you: it lists the newest build of each line — and checks each against what this machine can actually install. On Debian 13 that leaves one: the 5 line depends on libgmpris, a library neither Debian nor Signalyst publishes, so it is named with the reason instead of being offered as a choice that would end in a half-installed package. The line you install is then held: updates stay on it, and moving to another major stays something you ask for. Signalyst's licence provides for a time-limited trial. A machine the vendor publishes nothing for is told so by name — Debian 12 gets no build of this package. Checking a line costs a few seconds of low-priority processor time the first time it is asked, and nothing for the next ten minutes.
  • [core][ui] The vendor's licence is shown before the install, and the install waits until you accept it. Signalyst ships its end-user licence agreement as a screen its own installer displays. Audiogravity installs without that screen, so the agreement was being accepted by someone who had never seen it — for the adapter too, on every box, not just for the new package. Both licences are now read from the package itself, each time, so the text is always the vendor's current one, and shown before anything is downloaded. The acceptance is held by Audiogravity itself, not only by the screen: an install asked for without it is refused. If the licence cannot be read — the vendor out of reach, a proxy in the way — you are told so, and still asked to accept, rather than shown nothing and let through.
  • [core][ui][manual] The 30-minute trial of HQPlayer Embedded is announced before it can surprise anyone. Signalyst lets it run without a licence key, but only 30 minutes at a time — and the clock runs from the moment HQPlayer Embedded starts, not from the first note: measured twice on a box, once playing nothing at all, the control port refused at 32 min 11 s of service uptime after accepting a command at 31 min 51 s. So a box whose player has been up half an hour has no music left in it, which the vendor's own wording ("plays for 30 minutes at a time") does not suggest. What happens then looks like a breakdown: HQPlayer keeps running and keeps the DAC, its web page still answers, and only its commands are refused — it writes nothing about the expiry in any log, so Audiogravity, which can read that state nowhere, showed HQPlayer unavailable and the music stopped for no stated reason. The limit is now named where it can still be acted on: in the install window before anything is downloaded, on the package's card once it is installed, and in the manual, which says to restart it from Services — and to start the music again, since the restart leaves HQPlayer with an empty queue. Declared by the package itself (trial_notice), like the password its web interface needs, so no screen holds a rule about a particular package; and worded as a condition, with no badge, because the box never learns whether a licence was entered. Audiogravity does not restart it by itself: that would work around a limit the vendor set, and cut the sound without warning.
  • [core][ui][manual] HQPlayer Embedded's settings page is one click away, from the HQPlayer card. Its filters, its modulators, its DSD rate and its licence key are set on a web page of its own, which the manual sends you to and which you had to reach by typing the box's address followed by :8088. The HQPlayer card in Library → Sources now offers Web interface beside its DSP controls, opening in a new tab. Only for the HQPlayer running on this box — an HQPlayer on your network serves no such page — and also when it does not answer: that port keeps answering while its control port refuses, which is exactly what an expired vendor trial looks like, and that page is where the licence key is entered. The address is composed by the interface from the host you reached the box on — a name, a LAN address, a tunnel all work — and the port comes from the package's own declaration, where it already stood for that page's password. It is served over plain HTTP, so your browser will say so about HQPlayer's page: HQPlayer serves no HTTPS of its own.
  • [core][ui] HQPlayer Embedded gets a password for its web interface when it is installed, and keeps it out of the logs. Its Debian package installs its web interface with no credentials at all: the settings pages refuse everybody, and every sign-in attempt writes a line to a log kept in memory — nine million lines in two minutes. The install dialog now asks for a password, prefilled with a random one shown in clear so you can note it down; Audiogravity sets it with the vendor's own tool right after the install, then restarts HQPlayer once, when it has finished starting, so it reads it. Credentials already on the box — a reinstall keeps them — are never overwritten, and are not asked for again. If the password could not be set, the card of the installed package offers Set web password: no need to uninstall and reinstall. The password is never logged, returned or shown in a message; the system journal does not keep it either — sudo records the command it runs twice, in its own line and in the lines of the session it opens, and both are turned off for this one command (measured: a marker passed the same way appeared twice before, nowhere after).
  • [core] HQPlayer Embedded's log lives in 10 MB of memory, with a hard limit. The package writes its log to /tmp — memory, with no limit, never emptied — and a loop in its UPnP server wrote 600 MB a minute there on a test box. The log now goes to a memory space of its own, where the kernel refuses to write past 10 MB: measured with that space full, HQPlayer keeps playing, restarts normally and uses no more processor time. Nothing is written to disk, which spares the SD card of an ARM box, and the log is emptied when HQPlayer starts once it is past half the space. Only the log's location changes in HQPlayer's configuration, written as its own account and replaced in one step, so a full disk can never leave it cut short.
  • [core][ui] HQPlayer Embedded can be started, stopped and switched to like the rest of the stack. It is declared on every box — a new one from its template, an existing one when it updates — with two profiles of its own: HQPlayer Embedded, and MPD HQPlayer Embedded, which keeps MPD running so that your own library plays through it. Every profile that starts or stops the HQPlayer adapter stops HQPlayer Embedded too, since the two cannot share the DAC. Declared once: if you take it out of your configuration, an update or an import does not put it back, and a box where it was already declared by hand, under any name, is left as it is.
  • [core][ui] Music plays through HQPlayer Embedded while it runs on the box. When a profile starts it, what you start from the Library goes to HQPlayer Embedded, and from it straight to your DAC — no NAA in between, so nothing asks for one. The HQPlayer card shows This box with its name, and Use as output on and locked for as long as it runs. An HQPlayer you had chosen on your network is kept, with its own Use as output and its saved DSP settings, which nothing on the card can reset meanwhile; the card goes back to it when HQPlayer Embedded stops. The rule follows the service itself, not the name of a profile. The player shows what HQPlayer Embedded plays, the signal path starts from it, and the output switcher moves its output. The network search no longer lists the box itself, and the box's own address is refused as the card's choice: its HQPlayer is used on its own while it runs. A format HQPlayer cannot decode is converted to FLAC on the way out (see above); what stays refused is refused with advice that works — a profile that stops HQPlayer Embedded, since its switch is locked.
  • [core][ui] HQPlayer Embedded's output is chosen in Audiogravity, like any player's. It has none after installation. Its tile in Config now has a Guided view that offers the box's outputs, and choosing one changes that setting alone: the rest of the file is HQPlayer's own, set in its web interface — its DSD rate included — so the tile has no Reset to default and no form, and the Expert view edits the whole file as text, comments kept. The file stays HQPlayer's, written as its own account and replaced in one step. Saving it restarts HQPlayer only if it is running: it takes the DAC as soon as it starts, so a save never starts it — and the answer no longer says restarted when nothing was.
  • [manual] HQPlayer Embedded has its own section, and the Audio Software page is described as it now behaves. The Outputs & engines chapter walks through installing it — the version your licence covers, Signalyst's licence, the password to note down — choosing its output, setting its DSD rate on its web page before the first play, switching to it with its two profiles, playing music through it, and what to do when a step did not go through; it says that HQPlayer Embedded needs a licence from Signalyst. Troubleshooting says how to set a new password when it is lost, and what to do when the DAC hisses instead of playing. An Apple-encoded (ALAC) library needs nothing done to it: the box converts each track as it hands it over. The Audio Software part of Administration now describes the install window, an uninstall that keeps a package's settings, and Install Incomplete. The install window is pictured as a phone shows it.

Changed

  • [core][ui] The box's HQPlayer adapter is now called naa, and stops borrowing HQPlayer's own name. Audiogravity used the one word hqplayer for two different machines: the Signalyst Network Audio Adaptor, the daemon running on the box, and the HQPlayer instance it receives sound from, which runs somewhere else. The collision had reached the interface it offers — /packages/hqplayer meant the adapter, /hqplayer/status meant the player — and ran through the service identifiers, the configuration file it edits and the output switching. The adapter is naa from now on, everywhere it is named by identifier; the player keeps hqplayer and every one of its routes. Nothing changes on screen: each screen shows the label, HQPlayer NAA, which has not moved. A box that already holds an audio-config.json has the name changed in it when it updates — profiles included, so no chain loses a step — and keeps the file's owner and permissions. The configuration editor's saved copies of the adapter's file move with it, so none of them disappears from the editor, and a failed update that rolls back restores the previous file and still finds them. A configuration exported before this release is brought over when it is imported back. For API clients: the package id and the service id hqplayer become naa, in /packages/*, /audio_app_config/*, /audio-stack/* and the service-metrics events; /hqplayer/* is untouched.
  • [core][ui] Removing a package keeps its settings; erasing them is a separate choice. Debian distinguishes removing a package from erasing what it leaves behind, and Audiogravity only ever did the first — so a box that tried HQPlayer once kept its system user, its privileged grant and its device rules for ever. Uninstalling now offers to delete all of it, as a distinct and explicit action, because the same sweep also takes the DSP settings with it: every filter and modulator choice, which are the hours nobody wants to lose to a misclick.
  • [core][ui] An install whose next step could not be done says "Incomplete", not "Failed". The package is on the box, and the card says so; what is left to do — its password, its log space, its restart, or other software apt had to remove on the way — is what you read, as a warning. It used to read Install Failed about a package that was installed, and an unexpected error in one of those steps — a full temporary folder, a command that cannot start — was reported as the install's own failure and skipped the steps after it. Update all now says it too, where it used to announce All Updates Complete.
  • [core] HQPlayer Embedded is turned on after any install, a reinstall included. Its removal script disables the service, so a reinstall left it off, now and at every boot, where a first install turned it on.
  • [core] A stop of HQPlayer Embedded ends within 15 seconds. Stopped within about a second of its start, HQPlayer ignores the request and systemd waited 90 seconds before ending it — after a quick second restart, two profile switches in a row, or while it was being removed. A normal stop takes under half a second.
  • [core][ui] The restart badge only shows when a restart is owed. After installing or updating HQPlayer Embedded, the card asked for a restart Audiogravity had just done itself, and a click cut the music for nothing. It now shows only when that restart failed.
  • [core] An install that removes other software says so. apt takes out, without asking, a package that conflicts with what it installs, and the screen said Successfully installed. Audiogravity now compares what the system holds before and after, and names what went.
  • [core] Audiogravity no longer holds the right to install any package file with dpkg. Nothing has used it since packages from a vendor install through apt; a box loses it at its next update.

Fixed

  • [core] A successful install no longer shows errors, and is only called successful once the package is really there. A package downloaded from its vendor was installed with dpkg, which never installs what a package depends on: whenever something was missing it stopped, the log showed dpkg failed and dependency problems in red, and apt repaired afterwards — so a normal install of HQPlayer Embedded read as a string of errors before its success. It is now installed through apt, which brings its dependencies along in the same step, as the upmpdcli bundle always did; switching the adapter back to an older line still works. The repair step is gone with it, and so is its worst case: apt mends a package whose dependency cannot be met by removing it, and that used to be announced as "installed successfully" for a package absent from the box. The verdict now comes from dpkg itself, after the fact.
  • [core] A box set to another language than English is offered updates again. apt writes its answers in the language of the system, and Audiogravity looked for an English word in them: on a box set to French or German, no update was ever offered for the software installed from the system's repositories. apt is now always asked in English, and an answer that cannot be read is "could not check", never "nothing available" — which the new check of what a vendor's package needs relies on too. That check also looks first at what is already installed, which takes a few milliseconds instead of seconds.
  • [core] System files written by Audiogravity are handed to the administrator, and the ones left behind are repaired. A file Audiogravity writes into the system — a mount unit, a service setting — is prepared under its own account, then handed to root. That last step was never checked, and the permission it needs for mount units never existed: every USB library and network share mount written so far stayed owned by Audiogravity's own account, although the system runs it with full rights. Every step is now checked, and the update hands those files back.
  • [core] A command that overruns its time limit is stopped, even when it runs with administrator rights. A deadline ended only the sudo in front of the command and then waited for the command itself — the refresh of the package lists could keep running, holding apt's lock, and a support report probe could wait far past its budget. The command is now asked to stop through sudo, which passes it on, and the wait is bounded.
  • [core][ui] Restarting a service no longer reports a failure while the restart goes on. The Services tab's restart went through a command given 5 seconds, and a service that took longer to stop was reported as failed although it restarted. It now goes through the system's own interface, as start and stop always did. The restart badge of the Audio Software page also says so when a restart really fails, where it announced success whatever the answer.
  • [core] Importing a configuration no longer stops on a service that is not installed. A service declared before its package is installed is an ordinary state — its profiles read as unavailable — yet the import refused it, so a box without every optional package could not restore its own export. It is now a warning, and the import goes on after you confirm.
  • [core] A network share whose name has an accent is recognised as mounted. Its mount point was read with the wrong decoding, so /mnt/Musique été was never found. The support report also reads a mount point with a space correctly.
  • [ui] Log lines show >= as >=. The install log and the system log viewer escaped each line twice, so a comparison in what a tool printed — libomp5-19 (>= 1:19.1.6) — appeared as (&gt;= 1:19.1.6). A line containing markup is still shown as text.
  • [core][ui] A refused install says why, on screen. A package that cannot go on this machine — a dependency nothing provides, not enough room — is now turned down before anything is downloaded or changed, and the reason is what you read, instead of Failed to install.
  • [core] A package is no longer downloaded into memory. The download went to /tmp, which on these boxes is a RAM disk — so installing HQPlayer Embedded would have held 144 MB of memory on a machine whose whole job is to play audio without a glitch. It now goes to disk, chosen when the install runs rather than written down, since which folder is in RAM differs between machines. A TMPDIR set by the operator still wins, and a download left behind by an install that never finished is cleared the next time.
  • [core] An install that cannot fit is refused before it starts, not halfway through. Only the download was checked for room, and not even from the right point of view; nothing looked at what the package unpacks to — 257 MB for HQPlayer Embedded against 144 MB downloaded. The unpack is now measured with the space dpkg really has as root (5 % more on ext4 than a regular user sees), counts the download when both land on the same disk, and is checked before a byte is fetched. A package that declares no size is let through rather than refused on a guess.
  • [ui] Notes in dialogs keep the space, colour and line height their style gives them. A rule shared by every dialog reset its paragraphs with more weight than any style of their own, so none of those applied — measured at 0 px between a note and the next block in the Audio Software dialogs. It now only sets a default for paragraphs that have no style of their own: in the install window the version a box cannot install sits with the version choice, the web password and uninstall windows get their spacing back, the No activity this session line of the service and profile details takes the lighter grey it was given, and the support report's introduction its spacing.
  • [core][ui] The support report describes the HQPlayer you actually use. Its HQPlayer section read an address from the settings file, which only seeds the card: on a box whose card had chosen an HQPlayer, it said not configured. It now reports the instance chosen in the card, whether the box's own HQPlayer is the one played through, whether library playback goes to HQPlayer, and which HQPlayer answers, by product and version. An HQPlayer that does not answer is asked once, not twice.
  • [core] HQPlayer as the output no longer refuses every play after a start-up where its adapter started at the same moment. Audiogravity read the adapter's state, then began listening for its changes: a start completing in between went unseen, and every play was refused as NAA not running until the adapter next restarted. It now listens first.
  • [core] A UPnP renderer is recognised as this box on any loopback address, and a renderer address with no host never is.
  • [ui] A box no longer publishes the interface's own source code. The build asks for hidden source maps, which only keeps the reference comment out of the bundle: the .map files were still written, and copied into the package — so every box served the whole original source at an address anyone could guess. They are now left out of the package, and stay on the build machine for debugging. The comment that presented the option as a protection said the opposite of what it did; it is corrected.
  • [manual] The certificate entry carries the box's name, not its address. In First run, trusting the local authority on an iPhone or an iPad: the entry reads Audiogravity Local CA followed by the box's name in brackets. The manual announced its address, which is not what the certificate carries.
  • [core] Adding tracks to what HQPlayer is playing no longer blanks the album on screen. Audiogravity follows the album HQPlayer plays by counting the entries it pushed, and lets that list go when the count stops matching — which is how a playlist changed from HQPlayer's own remote stops wearing your album's titles. But HQPlayer only acknowledges a longer playlist on a timer of its own: measured on HQPlayer Embedded, about five seconds after the tracks were added, whether one was added or ten. So every append blanked the title, the artist and the cover about three seconds later, and for good. That wait is now recognised for what it is — and only it: the exact count from before the append, and only for as long as an acknowledgement can still be owed. A playlist that really did change is still let go at once, a stopped HQPlayer cannot hold the exception open, and a list too long to follow takes it with it.
  • [core] A command HQPlayer carried out is no longer reported as failed. Every command opens its own connection, and HQPlayer answers the first one on a connection only after a delay of its own — measured across the network on HQPlayer Desktop, with that instance stopped: 2.37 s for a status, and four pauses at 2.60, 4.55, 3.43 and 2.59 s. Audiogravity waited three seconds, so next, previous, pause and a jump within the track went through and were announced as failures. A command now waits eight seconds. What Audiogravity reads over and over keeps the shorter wait: the playback state, and the volume, which a slider sends at every step of a drag.
  • [core] An HQPlayer named rather than numbered is told where to fetch from. Audiogravity hands HQPlayer the address of the box on the route toward HQPlayer — the only right one when the box has two networks, sits behind a VPN, or is on an audio LAN with no gateway. That route can only be worked out from an address and not from a name, since resolving one there could hold up everything else the box is doing, so an HQPlayer configured as musicscore.local fell back without a word on the default route — wrong on exactly the networks this exists for. The address is now read from the connection Audiogravity already holds open to it, where the system has done the resolving and the routing itself. An HQPlayer reached over IPv6, and a name that resolves to the box's own loopback, still get the box's ordinary address: that is where its files are served.
  • [ui] The tracks HQPlayer could not open are named on screen. HQPlayer accepts an address and then drops it without an error of its own — a file it cannot open, a server it cannot reach — and Audiogravity has been reporting which ones since the previous release, with nothing showing them: the album played short, and nothing was said. A notice now names them, without calling the playback a failure, since what was kept is playing.
  • [ui] Starting an album says it is on its way. The tracks leave as one exchange, and HQPlayer answers the first command of a connection only after that same delay of its own: measured on the box's own HQPlayer, 0.47 s for a single track but 2.39 s for sixteen, and longer over the network. Nothing on screen said anything during that wait, so the album was clicked a second time. A notice now appears when a start takes more than a second and a half — never for a single track, which is the common click.
  • [ui] The filter you pick after changing HQPlayer's mode is the one that applies. HQPlayer offers different filters and shapers in PCM and in DSD, and what it takes is a position in that list: the card kept the previous mode's lists, so a pick after a mode change silently applied another filter. The lists are now re-read when the mode changes.

[0.9.58] - 2026-09-20

Changed

  • [ui] The Compact Mode switch is gone; the interface keeps the density it has always shown. The switch was on from the very first launch and the only thing turning it off did was make everything larger — a second layout nobody asked for, carried through 31 rules and every screen of the interface. Those rules are now simply what the interface is: the spacing steps read 4/6/12/18/24 px and the type steps 12/13/14 px where three names were being substituted underneath. Nothing moves on screen, except in two places. The settings drawer gives each switch a full-width row — light/dark mode, notifications, animations, portrait lock and Face ID / Touch ID were paired two to a line, each label squeezed into half the width beside its switch. And a service tile that starts, or a profile that becomes active, no longer shifts its contents by a pixel as its border thickens: the second border pixel is taken out of the padding, as the processor tile beside it already did. The sign-in page is deliberately left untouched: it is the one screen the setting never reached, so it had been drawing the larger layout alone, and following the interface would have shrunk it by itself (its card 558 → 460 px on a computer, 536 → 445 px on a phone). It now states the steps it used to inherit, and is pixel-identical — checked over every element it holds, at both sizes. Verified the same way across ten tabs on a computer and a phone: no other difference.
  • [manual] The Administration chapter stops offering a switch that no longer exists. Compact Mode was listed among the settings and the drawer's picture showed it; the picture has been retaken and the list now describes what the drawer holds.
  • [ui] Repeat and shuffle leave the playback bar for the fullscreen player. They sat on a second line under the track title, in a bar whose job is to show what is playing and offer the transport — the same two controls are in the fullscreen player, where the rest of the playback settings already live. The bar keeps the height it has always had: the space they leave is not taken back, so nothing above the footer moves. As a side effect of pinning that height, the interface now reserves the right amount of room for the bar — the figure it went by said 61 px for a bar that measured 83.
  • [manual] The Listening chapter says how to open the fullscreen player, and stops saying something that never worked. It read "Tap it … to open the fullscreen player"; tapping the bar does not open it, and tapping the cover art opens the album's details and tracklist instead. The three ways in are now listed — the tab at the top-right corner on a computer, a swipe up from the track title on a phone, and by itself when you start something playing from the Library — which matters more now that repeat and shuffle are only in there. The bar's picture has been retaken on the same track it always showed.

Added

  • [core][ui] Audiogravity now reads which HQPlayer it is talking to, and says when the box cannot work with it. Connecting an instance reads its identity once — the product, so a Desktop is told from an Embedded, and the engine version, now shown beside the name on the card and against each result of a network scan. From it comes the one thing that decides whether the chain carries sound: HQPlayer and the box's network audio adapter have to be on the same major line. When they are not, the card says so and names both versions, and Use as output is refused rather than routing playback into silence — the same treatment an adapter that is not running already got. An unknown pairing is not a broken one and blocks nothing. The refusal also applies when a track starts, not only when the setting is switched on — the setting is remembered, and the pairing can break under it. Nothing is polled: the identity is read when the instance answers and forgotten when it goes away, so an upgrade on the other machine is picked up by itself.
  • [core][ui] The adapter now follows the HQPlayer it serves, which is what repairs a box that drifted. Holding it to the line already installed protected a healthy machine but left a broken one broken. It now takes its line from the HQPlayer in front of it when that one can be reached, and falls back to the installed line otherwise. A box running adapter 6 with HQPlayer 5 is therefore offered the 5 line — an older version than the one installed, on purpose. The Audio Software page no longer calls that an update: it says an older version is offered, and asks to switch rather than to update. Which way a version goes is now decided by the core, because ordering Debian versions is not a comparison of two strings — and the update notification follows the same rule, so a repair no longer arrives on your phone announced as an update.

Fixed

  • [core] The HQPlayer NAA update no longer offers a version that cannot work with your HQPlayer. Signalyst publishes the 5 and the 6 line of networkaudiod in the same directory, and the update took the last link on the page — 6.1.4-71 on a machine running 5.1.8-68, although NAA 6 does not work with HQPlayer 5. The choice is now made by comparing versions instead of reading the listing in order, and the NAA is held to the major line the machine already runs: that machine is told it is up to date, and one on 5.1.5-67 is finally offered the 5.1.8-68 that was hidden underneath a version it could not use. A machine with nothing installed still gets the newest published. Checked against the vendor's own listing on amd64 and arm64. Holding a package to its line is declared for that package rather than applied to every downloaded .deb, and where the line cannot be established the update is refused instead of falling back to the newest — a Debian epoch or a rebuild suffix in the installed version used to make the constraint vanish silently. The package name is also read from the registry rather than matched as the literal networkaudiod.
  • [core] A package is no longer reported as missing while it is running, nor as present while its installation is broken. Whether a package is installed was decided by looking for the word "installed" anywhere in what dpkg reports. half-installed — an installation interrupted midway — contains it and passed; triggers-pending, triggers-awaited and half-configured, which describe a package that works, do not contain it and failed. The answer now comes from the status word itself, in the one place all three readers share.
  • [core] Finishing an install no longer sends Audiogravity out to the vendor's website, twice. The notification pushed after every install, update or removal asks for the package's details without its available version — and that request was being honoured everywhere except where it mattered, so each operation ended with two lookups over the network, plus an apt-cache query for packages apt cannot answer for at all. That query costs seconds of processor time on DietPi, whose apt cache is disabled. The last known version is still shown, from memory.

[0.9.57] - 2026-09-19

Changed

  • [ui] The default theme's corners are softened, to one 2 px radius everywhere. Minimal drew every corner square, which read as hard. Cards, panels, buttons, filters, badges, the top bar's figures, the player's artwork, the pipeline's controls and the notifications now share a single 2 px radius — graded values were tried and dropped, because 6 px on the large panel of each tab changed the theme's character more than it softened it. Elements whose corners were written into their own style, and so ignored every theme, now take them from the theme; slate and gravity keep exactly the corners they had, except the offline banner's Retry button (2 px in slate) and the reference box of the help windows (8 px in gravity), which now follow each theme's own scale.
  • [ui] The Services and System charts show what was measured, and nothing else. On Services, each box draws a thin line over a light area with a dot on the latest value; on System, each tile draws one bar per measurement, the newest in full, with the highest value of the window and the time it covers written above — measured from when the readings arrived, since the box sends them every 2, 10 or 30 seconds depending on activity. Both start empty and fill from the right, leave a gap where a reading is missing or the stream was paused (the app hidden, the box offline), and start their scale at zero. The second curve of a box (outgoing, writes) is grey instead of the warning orange it fell back to.
  • [ui] In the mini player, the stream-origin badge is the twin of the output badge beside it. QOBUZ next to USB: same height, type, greys and border, drawn by one shared rule so they cannot drift apart. The HIGHRESAUDIO logo is left out of that badge — wider than tall, it stretched the badge to 26 px beside an 18 px one.
  • [ui] Status dots no longer glow. The green dot (up, online, active) and the orange one (a service starting or stopping) are flat discs, in every theme; they still blink.
  • [core] With HQPlayer as the output, playing or queuing now says what HQPlayer actually kept. HQPlayer accepts an address and may then drop it without an error of its own — a file it cannot open, an address it cannot reach — so the answer was "OK" whatever happened, and the only trace was a track missing from its playlist. Audiogravity now reads the playlist back in the same exchange and answers with the number of tracks HQPlayer kept, naming the ones it dropped. A play goes on with what was kept. A play that keeps nothing is refused at once, with the track names, where it used to wait 45 seconds and then blame the sound card. An enqueue that loses a track says so. A dropped file no longer wipes the title and cover of the rest of the album either, because the list Audiogravity follows is now the one HQPlayer really holds. A local track the library does not know is refused before anything is sent (400, formerly 503). For API clients: /library/queue answers routed to HQPlayer carry refused and refused_titles, media-server albums included.
  • [manual] The manual says what HQPlayer kept, and what Services shows when memory is not counted. Outputs & engines now explains that Audiogravity reads HQPlayer's playlist back and names the tracks it dropped, and Troubleshooting lists the new message — "HQPlayer could not open this track" — beside the others, since a message nobody can look up leaves the reader where they started. The troubleshooting entry for a box that does not count memory described 0 MB and a flat graph; the cards have shown a dash and no graph since 0.9.42, which the Administration chapter already said. The entry now describes what the screen shows, under a title that matches it.

Fixed

  • [core] The Systemd tab shows the real-time settings each audio service runs with, and saving keeps them. systemd reports a service's scheduling policy as a number — 2 for round-robin — and Audiogravity only understood names, so every service read "default", the five audio engines of the lab included, all of them running in real time. The editor starts from what the tab shows, so saving any other setting rewrote the service's settings file without its policy: a service put in real time through Audiogravity, with Audio Optimized for instance, would have lost it at its next edit (read from the code, not tried on a live service). The nice level was never read either — it was asked for under a name systemd does not have — so HQPlayer NAA, running at −10, read "default" too. Both are read now. A service that sets neither still reads "default", and saving writes nothing for them, so systemd's own defaults are not frozen into Audiogravity's settings above the package's.
  • [core] For API clients, a service's details agree with systemd. In GET /services/{name}, an unlimited memory limit read 17179869183G — systemd writes "unlimited" as the largest possible number when asked over D-Bus, and that number was converted as a size — and now reads infinity, as /services/{name}/properties already did. The network totals (io.ip_*) were always null because they were never asked for; they now carry the counts whenever IP accounting is on for the service. Nothing on screen showed either.
  • [ui] The Library button in the top bar now looks like the two buttons beside it. Between the menu and the settings wheel, it was the only thing in the bar drawn inside a framed box — taller than its neighbours, filled with a background, its icon small in the middle. It borrowed the generic style meant for buttons elsewhere in the interface instead of the one the top bar uses, and had done so from the start without anything saying why. It is now the same size, frameless, with its icon filling the button like the other two. The button at the foot of the page, which uses that generic style on purpose, keeps its frame.
  • [ui] Offline, the tabs stay within reach in a narrow window. With the tabs laid out as a bar along the top, in a window 768 px wide or less — a computer's browser window made narrow, a large tablet sharing its screen — the offline banner was given exactly the place of the bar and drawn over it: a 10 px strip of the bar showed below the orange, and a click on any tab landed on the banner, so changing tab was impossible until the box came back. Measured in a browser window 700 px wide. The bar now steps down below the banner and the page follows it; online, nothing moves. Phones were not affected: their tabs always sit in the side panel.
  • [ui] The Services charts follow the measurements. Each box began with thirty invented zeros — a flat line at zero, then a cliff on the right, for the first five to fifteen minutes (the box reports a service every 10 seconds, every 30 at rest; measured: two real points out of thirty after one minute). The charts were not redrawn at each new measurement either: they only moved when something else redrew the page. A missing reading was drawn as a zero.
  • [ui] The System charts no longer invent readings. An update carrying only the uptime added a point to every chart, copied from the previous one or at zero, and a machine without a temperature sensor drew a row of zeros headed "max 0.0 °C". A reading of exactly 0 was, on the other hand, taken as missing.
  • [ui] The expanded view of a Services chart opens on its first measurement. A single point was drawn as nothing (and computed as a division by zero), and a missing reading as a zero; the figure above the chart now shows a dash when the latest reading is missing, rather than an older value presented as current.
  • [ui] The Services chart history survives a reload. It was meant to be saved every 30 seconds, but the delay restarted with every measurement, so nothing was written while the tab was in use.
  • [core + ops] Your own library plays through HQPlayer on every box, not only on ours. To play a local file, HQPlayer is given an address on your network to fetch it from — and that address was written into the code as the one of the development machine in our lab. Every box handed it out. On a box anywhere else, HQPlayer was sent to fetch your files from a machine that does not exist on your network, and the local library never played through HQPlayer; in the lab it worked because the development machine happened to be running with the same library, which is precisely what hid it. Measured on the lab's ARM box: with the development machine stopped, adding a track answered "OK" and nothing reached HQPlayer's playlist. The address is now the box's own, taken on the route toward HQPlayer, which is also the right one on a box with two network connections or behind a VPN. The files are fetched from the same signed address network speakers use, with the file's extension at the very end: HQPlayer picks its decoder from the end of the address, and a DSD or WavPack file whose address ended with its signature was dropped without a word. HQPLAYER_STREAM_HOST stays for a box HQPlayer reaches through an address it does not own (NAT, a proxy): empty means detected, it is now present in the configuration of new and upgraded boxes alike, and a value that is not an http(s):// address is ignored with a warning instead of breaking every source sent to HQPlayer.
  • [core] MP3 files from your library play through HQPlayer. HQPlayer drops an MP3 when it is given its size — measured one response header at a time, the size alone decides — and one sent without a size, the way a radio is, never started when the file opened with its tag and embedded cover. MP3 files now reach HQPlayer without a size and from their first audio frame. Network speakers still receive the file as it is, so they can seek in it.
  • [core] An MP3 whose length HQPlayer cannot measure shows the library's length, and can be scrubbed. HQPlayer works out the length of nearly everything, but not of an MP3 without an Xing/Info header: the progress bar had no scale and could not be dragged, although the library knew the length. The library's figure is now shown for such a track — HQPlayer's own measurement still wins whenever it has one — and it is tied to the track it belongs to, so a radio queued behind never inherits it, even when the queue changes at the moment the track does. For these files the library's figure can run a little long: 2.4 s on each of the files measured, which matches the size of the cover embedded in them, so the bar stops just short of the end.
  • [core] Starting an album on HQPlayer while a radio plays no longer fails in silence. The first thing a push asks HQPlayer is to stop what is playing, and while it plays it takes about four seconds to answer — longer than the three Audiogravity waited. Its late answer was then read as the answer to the next command, so every answer that followed was shifted by one; and when two answers arrived together, the push stopped before Play, leaving HQPlayer stopped with an emptied playlist while the request answered 200. Each answer is now matched to its command by name, as HQPlayer names it. Verified on the lab's HQPlayer while music was playing: the album starts, and a dropped track is still named.
  • [core] A track from your library played through HQPlayer shows the library's own cover. It was shown through an address with no fallback: a file whose picture the music server does not find — a folder picture it does not look for — showed no cover at all, where an internet search used to find one; and on a network without a gateway the address pointed at the box itself and was refused. The cover now comes the way the library shows it: the file's own picture, then MusicBrainz. Verified identical, byte for byte, to the picture embedded in the file.
  • [core] Sending albums to HQPlayer no longer erases what the local queue knows. Every track sent to HQPlayer was also written into a list of 512 entries kept on the SD card, which nothing ever read for HQPlayer. It pushed out, oldest first, the durations and covers of the streams queued on the local output, which then showed --:--, and rewrote the file each time. Nothing is written there for HQPlayer any more.
  • [core] A play HQPlayer refused no longer leaves an old error behind. When HQPlayer kept nothing of a play, the check started by the previous play went on running, and 45 seconds later published its own "never started" about a playback that had already been replaced.
  • [core] A library on a network share that stops answering no longer freezes the box when HQPlayer or a speaker fetches a file. The file was looked up on the core's main loop, and a share that stops answering blocks without a time limit: volume, transport and live updates stopped for the whole box. The lookup now runs aside.
  • [core] A network speaker is told the right format for files in folders whose name reads like a web address. A path such as AC:DC/Back #1/… was parsed as an address, which lost the file's extension, and an M4A, OGG or Opus file inside it was announced as plain bytes. The formats Audiogravity names itself — FLAC, WAV, AIFF, MP3, DSF, DFF and WavPack — were never affected.
  • [core] Looking for HQPlayer on the network follows a box that has changed networks. The search scanned the network the box was on when the core started, and after a move found nothing until a restart.

Security

  • [core] Your library files are no longer served to anyone on the network without a signature. HQPlayer fetched local files through a route of its own, open to every device on your network and signed by nothing: anyone who could guess a path could download any file of the library. That route is gone. HQPlayer now uses the signed route network speakers already use.

[0.9.56] - 2026-09-17

Fixed

  • [ui] Offline, the interface now shows the data it says it is showing. A banner reading OFFLINE MODE — VIEWING CACHED DATA stretched across the screen while the panel underneath it said "Unable to connect to server. Please check your connection." Both were telling the truth, which is what made it worse: only the strip of figures along the top was ever restored, because it was the only surface listening for the state that was being kept. The panels each fetch their own data, and nothing handed them back what they had last seen — they simply asked the box again, got nothing, and reported it. Six now keep their last reading and serve it when the box cannot be reached: profiles, services, the system dashboard, audio software, accounts and the processor description. The reading is only served when nothing answered at all — a box that replies and refuses, say after a key is changed, still says so rather than showing hours-old state as though it were current. Three surfaces deliberately keep the error: the configuration and systemd editors, because a stale copy on a screen with a Save button offers to overwrite the live file with it, and the log viewer, because frozen logs presented as live mislead on exactly the subject they are opened for.
  • [ui] The offline banner no longer covers the top of the page on a phone. It is fixed below the top bar, so whatever scrolls underneath has to leave room for it — and the rule that did so was shut inside a desktop-only block. On a phone, which is where the banner is seen most, the first section title was cut in half and the chip beside it disappeared behind the orange.
  • [ui] The installed app starts offline after an update instead of starting locked. The app keeps a copy of everything it needs, and that copy is rebuilt at every release. Two files that the app cannot start without were missing from it: the one carrying the address and key of your box, and the one that holds the connection for live updates. So the first launch after an update, made without a network, opened an interface with no credentials — unable to ask the box for anything, and staying that way until it was reloaded with the network back — and no live updates, silently, since that failure is not one the app can see. Both are kept now.
  • [ui] Parts of the interface were fetched again on every single page load. The rule deciding which files never change spelled out the shape of the fingerprint in their names, and the tool that generates them may put a dash in it — so whichever files happened to draw one that day fell through to being treated as changeable: re-requested, and rewritten into the app's own copy, on each load. Which files were affected was redrawn at every build, so the cost moved around instead of being noticed. The images, icons and launch screens had a related problem of their own: they were re-checked against the box on every request, cache hit included, which is one round trip and one write per picture per page load — for files that cannot change until the next release replaces all of them at once.
  • [ui] Tapping a notification raises the window that is already open. It compared a full address against a bare path, then looked for the product's name inside the address — false on a box reached at a number, which is what the installer prints. Neither ever matched, so a second window was opened beside the one already showing the interface, every time.
  • [ui] The offline page is in English again, and names the product. The page the box actually serves had been translated to French at some point and called the product Gravity, while the fallback copy of the same page inside the app stayed in English — two copies of one screen, drifted apart. There is one text now, and the fallback has been reduced to the bare minimum so there is nothing left to keep in step.
  • [ui] Keeping a copy of the box's state no longer interrupts the main thread every few seconds. Saving it writes to browser storage, which blocks while it works, and it was done on every incoming measurement — several times a minute, for as long as the interface stayed open, on a machine whose spare capacity belongs to the audio. Only the last value is ever read back, so all the writes before it bought nothing. They are now grouped, and written out at once when the app is put away so nothing is lost.
  • [ui] Cold start is lighter. On every page load the interface quietly loaded nine of its heaviest screens in the background, so they would be available offline. They already were — the app's own copy covers them — so this did nothing but run nine modules nobody had asked for.

[0.9.55] - 2026-09-16

Changed

  • [manual] The manual shows the interface you are actually running. Its pictures were taken between eight and ten weeks before this release and the interface had moved underneath them: they offered MPD 0.23 where the box installs 0.24, an engine that is no longer listed and none of the two that now are, three outputs where a plain box lists six, a user card without the passkeys button beside it, a settings drawer without the Face ID switch, and an update banner announcing v0.9.11. Every picture in the manual has been retaken against a running box, at the size of a phone screen and at three times its resolution, so they stay sharp on the device most people read them on. Four captions were saying something the screen no longer says — the button is Restart Core, not "Restart Backend" — and were corrected with the pictures they sit under. The licence panel is shown twice, on trial and once activated, because it holds different things in each; the device identifier and the order number are blurred, an order number being the licence key itself.
  • [manual] Seven things the manual described and never showed. The stream-origin badge shipped in 0.9.54 was explained in words and had no picture at all — it is now shown, MPD beside the name of the station it is carrying, next to a paragraph saying why the engine alone tells you nothing. The bar at the bottom of every screen, the list of sources that the whole Library chapter is organised around, the radio collection, the signal chain drawn on a phone, and the tiles of the configuration editor were all in the same position: described, never pictured. Six new figures, placed where the text describes them.

Added

  • [manual] After an update, an app that shows an old version and says the core is offline. A device that was asleep or closed while the box was updated can keep serving the copy of the interface it already had. The old copy asks the box for things the new version moved, gets nothing back, and reports the only thing it can see — no answer — as CORE · OFFLINE. So the screen blames the box, and the box is fine: the address opens normally from any other browser. The troubleshooting chapter now says so, and gives the way out, including in the installed app, which has no reload button to press.

Fixed

  • [ui] The Library no longer offers to switch you to what you are already listening to. Playing a station and standing on the Radio screen, a banner announced "Radio Choco HD is now playing" and offered to switch you to the radio — the screen you were looking at. The two halves of that screen were not answering the same question: a station, a Qobuz album and a stream from a media server all travel over the same engine, so the field naming the engine says the same thing for all three, and the Library was reading the engine on one side and the actual source on the other. It now reads the source on both. The radio has a further quirk of its own — its screen is reached from the tab bar without changing the source you are browsing — so the banner now compares against what the screen is showing as well, and starting a station makes the radio the source you are browsing, as picking it from the sources screen would. That last step waits for the station to actually start: a station that fails to play no longer rearranges the tab bar underneath the error message. The banner also stops naming things that are not sources at all: with HQPlayer driven from its own remote, it offered "HQPlayer is now playing" and its Switch button sent the Library to browse a routing handle.
  • [ui] Dismiss now dismisses. The banner's Dismiss button cleared the banner and remembered nothing, while the box republishes what is playing about every three seconds — so the banner came back within three seconds, every time, and the button did nothing at all. The refusal is now remembered for as long as that source is the one playing, and lifts on its own once something else plays. It survives Roon naming itself in either of the two ways it does, and glancing at the Radio screen no longer counts as changing your mind.
  • [ui] Connecting a streaming account no longer moves you off the screen you were on. Signing in to Qobuz, Tidal or HIGHRESAUDIO refreshes the list of sources, and that refresh had come to adopt whatever was playing: it could drop you onto the radio screen, or replace the media server you were browsing with the local library. A refresh now refreshes the list of sources and nothing else. Arriving on the Library still opens on what is playing, which is the one moment that is a useful answer.
  • [ui] Opening the media server a stream is coming from. The Library asks the box for the address of a known UPnP server by a field name the box does not answer with — it publishes location, and the interface was reading the name that field carries inside the box rather than on the way out. An address that comes back empty is indistinguishable from a server that is no longer known, so this failed silently and always had. It is read under the right name now, which is what lets arriving on the Library open the server a stream is coming from rather than stopping at the list of sources.
  • [core] Sending music to a network speaker now stops what was playing in the room. Casting an album to a UPnP speaker left the box playing whatever it already had on its own DAC, so two different pieces of music came out of two outputs at once — and the player kept following the local one. The screen named the station, with its cover and its format, while the speaker played the album you had just chosen, over a signal path reading Radio → the speaker, which describes nothing. Nothing had gone wrong in what the box knew: it had the speaker's track, its artist and its album all along. The local output simply never stopped, and the player picks the source that is actually playing — so it went on picking that one, correctly, from a situation that should not have existed. HQPlayer hides this by accident, its network adaptor holding the sound card exclusively so the local player is stopped by being refused; a speaker on the network takes nothing away from your DAC, and there was nothing to hide behind. Handing over is now part of sending: the local player is stopped just before the speaker is given its music — late enough that a speaker gone from the network, or a streaming link that expires as it is fetched, leaves your music playing rather than stopping it for nothing, and once per album rather than once per track. A box whose local player is deliberately stopped, as the Roon and HQPlayer profiles do, is recognised as having nothing to hand over instead of being reported as a failure; a refusal, which the player answers rather than raises, is no longer dropped — that silence is what let the whole thing go unnoticed.
  • [ops] Installing the interface no longer runs a command by accident, nor ships a truncated comment. Every single install of the interface printed install.sh: line 347: -: command not found in the middle of its success report — alarming, and meaningless to anyone reading it. The unit file the installer writes is assembled by a shell block that fills in the paths of your box as it goes, which also means it acts on anything written between backticks; a comment explaining the leading - in front of the command systemd runs had its - written that way, so the shell ran - as a command and wrote the comment to your box with the one character it is about missing. The service itself was never affected — the - in front of the command is outside the quotes and always arrived intact, which is why this survived so long. Both the stray command and the mangled comment are gone, and a test now refuses any backtick left unescaped in a block of either installer, including the blocks nested one level down, which are plain text while the package is built and become live code when the install script runs.

[0.9.54] - 2026-09-12

Added

  • [ui] The players say where the music comes from, not only what carries it. Qobuz, Tidal, HIGHRESAUDIO, a station, a media server down the hall and the files on the box all travel over the same engine, so the Audio Pipeline tab's Now Playing card could name nothing but that engine: a Qobuz album and a webradio both read "MPD". Where it comes from was in the box's answer all along — the origin, and the name of the server or the station when there is one — and no screen read it. The card now carries it beside the engine, and leaves it out when it would only say the same word twice. Three things the box published and nothing showed go with it: a stream from a media server reads MinimServer rather than "UPnP", a recognised station reads its own name rather than "Radio", and a playback driven from HQPlayer's own remote reads HQPlayer rather than "External", which named nobody. All of it comes from the live player feed the interface already holds open — the box is asked nothing it was not already answering. The pipeline's own data never carried the origin, and putting it there would have meant asking the music engine a second time for what the player had already worked out, then a third time for HQPlayer and a fourth for a network renderer. The queue is deliberately left as it was: what it publishes about a track names the kind of source, never the particular server or station, so it has nothing specific to show.
  • [ui] The support report gives the address to type. Every half was in it already and no line held them together: the port the interface is served on under WEB INTERFACE, the name the box announces under NETWORK, the box's own address on the network quoted only inside a certificate warning. Joining them by hand is what sends a reader to port 80, where nothing listens, and makes a box answering perfectly on 8080 read as unreachable — reported from the field, off a report exact on every count. WEB INTERFACE now carries an Address line: http://10.0.4.254:8080 · also http://musics.local:8080. The LAN address leads because it does not need mDNS to reach the reader, and a box whose announcement is broken is precisely the box a support report is opened for; the .local name follows as the one that survives the lease changing. Written only where the interface answered — a TLS handshake alone is enough for the core to record a scheme, so a laxer test would have printed an address two lines above the report's own "NOT REACHABLE on port 8443" — and the announced name is left out when the daemon that would answer it is known to be stopped, the two probes behind that line being independent and able to disagree. The port is dropped when it is the scheme's default, a :443 reading as a second, different address. Neither address is a promise: both rest on a port measured from the box itself, and the name needs mDNS to cross the reader's network.

Fixed

  • [ui] The Admin page could have lost the marks saying who is online. Its two halves disagreed on the name of the field carrying that list — one wrote it under one name, the other read another — on the path taken when the viewer is not an administrator. While the tab counter existed the mismatch threw an error loudly enough to be seen; removing the counter is what would have made it silent, leaving nothing where the list belongs and the browser failing on the first user drawn. Found in review of that removal, and fixed with it.
  • [ops] A box that serves only the interface now answers to its own name. The .local name is the one address that survives the router handing out a new lease, and the certificate has always promised it — but the small service that answers it, avahi-daemon, was installed by the core package. On the standard box that is invisible, because core and interface share one machine. On a split install it put the promise on one machine and the answer on the other: the machine you actually open in a browser is the one serving the interface, and nothing there installed the daemon, so the name worked only if that machine happened to have it for some unrelated reason — the same accident, moved one package along, that having it depend on AirPlay used to be. It is now installed, enabled and started by the interface package, which is also the one that issues the certificate making the promise, so the two can no longer be separated. Placed before the HTTPS branch rather than beside the certificate it serves: a name is typed the same way behind a reverse proxy as in front of one, and gating it on the mode would have dropped it silently for every plain-HTTP install. The core no longer installs it, since nothing addresses an audio-only machine by name and AirPlay carries the daemon itself. Failure stays non-fatal both ways — a box that cannot announce its name still works perfectly on its address — but it is no longer silent: the box is probed again after the install rather than apt's exit code being believed, because apt can report success without the command becoming available, and that was the one case that produced no message at all. The two package-list refreshes such an install could run seconds apart are now one.

Changed

  • [manual] The manual describes what you do, not how it was built. A user manual had been carrying the specification alongside it: how many genres and sub-genres each streaming catalogue publishes, how many playlists a Tidal shelf held the day it was counted, why six of Qobuz's fifteen shelves are left out, which arrangement HIGHRESAUDIO asked for and approved, and what we had or had not verified on an iPhone. None of it tells a reader what to press, all of it dates the moment a provider changes their catalogue, and a count taken once and published as a fact is the kind of sentence nothing signals when it stops being true. It also editorialised on the services it browses — naming nine Qobuz shelves "the ones worth browsing" is not the manual's call. The browsing sections now say what each shelf holds and what a tap does. Internals went the same way: the systemd mount unit behind a network share, the RPC daemons an NFS mount would need, MPD's control socket, why the certificate authority is fetched over plain HTTP. One correction is a matter of fact, not of tone: the manual said all three streaming services are reached "through their official APIs". That is true of HIGHRESAUDIO alone — Qobuz's credentials are read out of its web player and Tidal's are the reverse-engineered ones of its Android client — so the claim is gone rather than softened. Four passages were rewritten rather than cut: the External badge now says what actually produces it (another app casting to your box, or HQPlayer's own remote) instead of alluding to it; HTTPS is stated as the recommended mode and why; a note recommends reserving a fixed address on your router, next to the one about the .local name; and a scan for UPnP media servers now says it only reaches the box's own network. Landing FAQ regenerated from the manual, which had drifted from it.
  • [manual] Two things the manual claimed that the code does not do. It listed an AirPlay receiver among the destinations you can send music to. Your box receives AirPlay; it never sends to one, and no output in the selector does. And it promised that when Tidal produces no sound, Audiogravity "checks what Tidal actually serves and names the cause instead of leaving you with silence". That check exists only when HQPlayer is your output — on your own DAC there is no message, only the silence, which is the hardest symptom to report and the one the manual was quietly telling you would never happen. The entry now says which case gets a message and which does not, so a reader whose Tidal is mute knows to look there rather than at their sound card. Both were found by reading the code rather than the comments around it: one of those comments asserts the protection is general, and it is not. Elsewhere the same pass dropped what a reader cannot act on — how the support report folds repeated journal errors, why MPD's control port is closed, the handling of a certificate you supplied yourself — and stopped claiming a bit-perfect DAC "commonly" has no volume control, advice that cannot be checked and ended by suggesting you pick a different output.
  • [core] [ui] Switching the order of the local library no longer makes the box read the whole library again. All, Recent and A–Z are answered by the box, over the complete collection, so touching one asks for the first page — and asking for the first page was the only way the box had of knowing that somebody wanted fresh data. It could not tell "the same shelf, in another order" from "go and look again", so every change of order re-enumerated the collection: one question to list the albums, then one round trip per album to find the file its cover is read from. Measured on a box holding 475 albums: 0.19 s of that work against 0.002 s to answer from the list already in hand, and the ordering never needed the fresh copy — the box holds the whole collection and can turn it around on its own. The intention is now said rather than guessed: only the ↻ Refresh button asks for a new reading. Nothing else changes on screen, and nothing is traded away: the box now asks its own music index whether the collection has changed — one question, measured at 0.17 ms against the 190 ms of the reading it can spare — instead of throwing its list away every minute on principle. So a collection that grew underneath appears at once, wherever the change came from, where it used to need the shelf to be opened again at the right moment; and a collection that has not changed in a fortnight is no longer re-read once a minute for nothing. The reading itself is also no longer done twice at the same time: two open tabs, or a second tap on ↻ before the first finished, each started their own walk through the collection, and they now share one — the same guard the streaming catalogues were given in 0.9.53, on the one enumeration that reads the box's own daemon.

Removed

  • [core] [ui] The Admin tab no longer counts who is connected. A figure beside Admin in the tab list read "people connected / accounts on the box" — shown whenever the tabs run down the side of the screen, which on a phone is always. It answers a question nobody asks of a box that plays music at home, and every load of the interface asked the box for it. Who is online is still shown where it means something: on the Admin page itself, each user carries a green mark for as long as they have the interface open. The interface now also ignores that figure if a box still sends it, so an interface updated ahead of its core stops showing it straight away rather than waiting for both halves to be upgraded — worth saying because admin is the name of a real tab, so a rule as broad as "keep the counters that name a tab" would have kept this one.

[0.9.53] - 2026-09-07

Added

  • [core] [ui] The radio is a source, like Qobuz or the local library. It streams over the same engine as the three streaming services, so it held no entry of its own and had no card on the sources screen — it existed only as a tab. It is now listed, it can carry the playing badge, and its queue is the shared MPD queue instead of an empty one. Its catalogue keeps its own screen with its own filters (country, genre, codec), so it is deliberately left out of the common search picker, and the tab bar drops the two tabs it cannot serve rather than offering an empty grid and an error.
  • [core] [ui] Every source now says what it is — sources[].kind: library, radio, streaming, roon or input. The sources screen filters on that instead of a hardcoded list of six ids, which is why the radio had no card in the first place: every source added to the backend had to be remembered in the interface too, and one was not.
  • [ui] Inputs are listed as sources. AirPlay and the UPnP bridge do put out sound, so they belong on the screen — but an input hides the source behind it (what a phone pushes over AirPlay has an identity the box cannot see), so it holds no catalogue: it is listed, it lights up while it plays, and touching it raises the player rather than a grid.

Fixed

  • [core] The public audio proxy no longer hands out what went wrong upstream, and no longer says nothing when it matters. The two endpoints renderers fetch audio from take no API key — they cannot, since a renderer sends no headers — so anything they return is readable by anything on the network. When a track could not be resolved they returned the streaming service's own error verbatim: its status, and a slice of its response body carrying identifiers of ours. Meanwhile the box's journal recorded nothing at all, so the one place the reason belonged was the one place it was missing. The caller is now told only that the track is unavailable, and the reason is written down where it is useful. Two neighbouring gaps went with it: a CDN that refuses a signed URL — a stale token, a mangled signature — left no trace either, which is exactly the failure this machinery exists to catch; and a track id chosen by the caller went onto a log line unescaped, so a crafted address could append a line of its own and have it read as if the box had written it. Ids and third-party error text are now stripped of anything that could end a line before being logged. A bare /highresaudio/stream/ with no track at all used to become an outbound request to HIGHRESAUDIO for whoever asked; it is answered locally.
  • [ui] The sources screen named the engine, not the source. A webradio playing showed "Local Library" as the active source, because Qobuz, Tidal, HIGHRESAUDIO, the radio and local files all travel over the same engine and the badge was keyed on that. It is now resolved from the content, so a station lights the Radio card and a Qobuz album the Qobuz card. A UPnP server lights its own card too, instead of the local library claiming a track served by a machine down the hall.
  • [ui] Several sources playing at once are all shown as playing. Casting AirPlay to one output while Qobuz plays on another lit only one card — whichever had started first — because the screen read the single source the player follows. It reads each entry's own state now.
  • [core] A second stream's card no longer waits for a page reload. The event that tells the interface something moved was raised only for the source the player follows, so a stream pausing, switching from a station to an album, or simply advancing a track published nothing at all. Verified not to add traffic: with two streams playing, twenty-two seconds of the live stream carry four events.
  • [ui] The "… is now playing" banner's Switch button lands where the banner says. It named the station correctly and then opened the local album grid.
  • [core] /library/albums and /library/search answer the same way for the same source. One returned an empty list for a mistyped source id — which reads as "this source is empty" and hides the typo — while the other returned an error for a perfectly real source. An unknown id is now a 400 on both; a real source that simply holds no local catalogue is a 200 with an empty result on both.
  • [core] Tidal now really says when a subscription has ended — 0.9.52 shipped that unable to work. The card read the plan with the stored access token as it stood, and that token lives a week with nothing renewing it unless something plays. On a box whose owner listens to anything else, Tidal answered 401, the answer came back "unknown", and unknown reads as subscribed — so the card announced Hi-Res on an account that had none, which is the very thing the feature was added to stop. Measured on a box the morning after the release: the token had expired 100 hours earlier, and one renewal through another path was enough to make the same lookup answer correctly. The lookup now renews like every other caller.

Changed

  • [core] The two addresses a streaming service serves audio from now behave the same way. One of them accepted shapes the other refused — an address with nothing after it, or one spanning several segments — and each of those became a request to HIGHRESAUDIO, on an endpoint that takes no key because the players on the network cannot send one. Nothing recorded why it had been built that way, so rather than guess it was measured: 940 track identifiers from the shop, the editorial playlists and the purchases, in two shapes, all of them made of the same kind of identifier joined by underscores. Not one contains the character the permissive form existed to allow, and that kind of identifier cannot contain it. Both addresses now answer the same, and anything that is not a track is refused before the service is asked.
  • [core] [ui] The box names the music on it "Local Library", everywhere, and once. It used to answer "MPD" — the name of the engine that plays the files, correct for the engine and wrong for the shelf of music it carries. The interface corrected it on the way to the screen, with a table of its own, so the two disagreed and only one of them was ever seen. The name now comes from the box, and the interface shows what it is given: the same words on the source card, on the now-playing badge and in the queue header, where the badge used to read the shorter "Library". Two more translation tables went with it — one existed only to recover the word "Local", the other rebuilt, in the interface, an answer the box already had.
  • [core] [ui] A UPnP media server is now identified, not merely named. When a track played from one of them, the box said which server by name only, so the sources screen had to match names to know which card to light — and two servers sharing a name lit both, with nothing to say which one was actually playing. The server's own identifier now travels with the track it serves, from the moment it is queued. On a network with one media server nothing changes; with two named alike, the right card lights. The same identity travels on the two paths where it was missing: a track cast to a network renderer, and one pushed to HQPlayer — where a local file now lights the Local Library card instead of naming the processor playing it.
  • [core] Opening a streaming shelf no longer asks the same question several times at once. Each catalogue list — genres, categories, labels, search filters, playlist groupings — is fetched once and remembered, but between the moment a screen finds nothing remembered and the moment the answer is stored there is a wait, and every other reader arriving in that window used to start the same fetch again. Measured on HIGHRESAUDIO: five readers arriving together sent five identical requests; they now send one. Qobuz and Tidal already had the guard, written by hand in each, and its easily-forgotten half — checking a second time after waiting your turn — is now written once for all of them, together with the ceiling that stops a catalogue walk from crowding the connection pool the audio itself uses.
  • [core] Qobuz and HIGHRESAUDIO now share the one route their audio actually travels through. The two services hand the box a signed address and nothing else to do with it, so their proxy endpoints had been written twice — the second created by copying the first, down to the wording of its documentation. Nothing was wrong with either, but the only change they had seen since had to be made in both files at once, and their tests had drifted apart: the byte relay was covered on one side, the purchased-track address on the other, and neither had the other's case. One envelope now serves both, and one test suite runs every case against both, so a behaviour can no longer be proven for one service and merely assumed for the other. Tidal keeps its own route on purpose: it rebuilds a stream rather than passing one through.
  • [core] Reading a configuration file is written once. Every service carried its own copy of the same few lines — does the file exist, decode it, log and carry on when it does not read — and the copies had already drifted: one decoded with a different library, another said nothing when it succeeded. The shared reader keeps the distinction that matters (a file that is not there yet is normal, a file that is there and unreadable is not) and restores the traceback that pointed at the failing service, which the copies had between them started to lose.
  • [core] The signature Qobuz requires on every playback request is computed in one place. It was written twice — once to play a track, once to find which credential works — and a wrong signature is indistinguishable from a wrong credential in Qobuz's answer, so a drift between the two would have been read as an expired account.

[0.9.52] - 2026-09-06

Fixed

  • [ui] A box that answers, and says what went wrong, is no longer reported as unreachable. The interface held that the core never returns 502 or 504 — only a proxy standing in for a stopped box does — so it replaced both with "Unable to connect to server. Please check your connection." The core returns both, and always has: 502 when the licence server is unreachable or answers nonsense, when a Qobuz sign-in cannot start, when a streaming favourite could not be written; 504 when the HIGHRESAUDIO catalogue is answering but slowly. On those screens the reader was sent to check their network while the box was fine, replying, and naming the real cause — with the badge beside the message reading CONNECTED. One rule now covers all three statuses: a 502, 503 or 504 carrying a message is the core speaking and the message is shown; one carrying none came from a proxy, since a proxy answers with HTML and leaves no message behind.

  • [core] [ui] A slow HIGHRESAUDIO no longer reads as a broken box. When their catalogue took its time, every shelf but one answered a flat "HIGHRESAUDIO unavailable" — the box owning a failure that was not its own. The honest answer existed already, and said the useful thing (their cold filtered search takes tens of seconds and then serves the same one instantly), but it was written on one route out of sixteen: the advanced search. It was never a decision, only a clause copied by hand forty times and completed once. The wording is now shared by every shelf, and says what it can honestly say everywhere — "HIGHRESAUDIO took too long to answer. It is not the box: the service is slow to reply." — without carrying over a promise about retrying that was measured on the search alone. The interface no longer classifies that answer as "something replied instead of your box": it never did so on the shelves, which show the core's own sentence, but the rule that said so was false and would have caught a licence or passkey screen the day one of them answered the same way.

  • [core] A HIGHRESAUDIO playlist grouping numbered zero disappeared from the strip. The row was kept on truthiness, and 0 is falsy, so a grouping HRA published simply was not there — with nothing anywhere to say one had gone missing. The same defect sat on the path that opens a grouping, one function away, and survived only because the ids are stored as text where "0" reads as true. Both are fixed, and a grouping with an unusable id (an empty one) is now dropped on purpose rather than listed as a shelf nobody can open. Two groupings sharing a title are no longer swallowed in silence: HRA's own filter addresses a grouping by its title, so two of them cannot be told apart by HRA either and nothing on our side recovers the second — what was worth fixing is that it vanished quietly, and it is now said once, with both ids.

Changed

  • [core] 281 lines of code nothing called are gone. Twenty-seven definitions with a single reference in the whole repository — their own: eight request/response models never sent or returned, nine client and infrastructure methods, three catalogue methods never wired to a route, and five helpers. An inventory had listed them a fortnight earlier; it was re-run before deleting rather than trusted, and that mattered — one entry had since gained a caller and was kept, another had already been deleted. Each definition was cut at the span its own syntax tree gives, not where a search suggested a block might end, and the published API of the library's fifty routes is byte-identical afterwards.

  • [core] The answer a streaming shelf gives when it cannot serve is written once. Forty copies of the same three clauses, differing only by the service named in the last one; thirty-two routes now share it. Three routes deliberately keep their own — an unknown Qobuz shelf and a malformed Tidal page or section answer 400, where the shared rule would make the same class of failure a 500 — and both the code and API.md name them, so a client is not told to expect one mapping for the whole catalogue. An unexpected failure is finally logged with its traceback: centralising the handler is what made that a one-line change instead of twenty-seven.

Added

  • [core] [ui] A streaming account whose subscription has ended now says so, instead of announcing a quality nothing plays. The sources screen showed Hi-Res · FR for Tidal and Studio · Hi-Res 24/192 for Qobuz — but both of those are what the box asks for, not what the service grants, and the two part company the day a plan ends. Measured on a real lapsed Tidal account: four tracks of 1'54, 7'10, 10'07 and 23'33 all come back as a 29.9-second manifest, at any quality setting, while the connection stays green and the catalogue keeps browsing. Qobuz does the same, serving a 30-second MP3 for every track it is asked for. Nothing in the playback path can tell that from a short track, so the music simply stopped after half a minute with no explanation anywhere. Each card now says what will actually be heard — No subscription · 30-second previews — and the state is re-read rather than stored once at sign-in, which is what makes it still true the day the plan lapses. The three services publish the same two fields for it, and a client reads them the same way whichever answered. What is not the same, and is documented as such, is what a lapsed account may still play: Qobuz and Tidal keep the whole catalogue browsable in excerpt form, while HIGHRESAUDIO refuses it outright and leaves only the purchases.

Fixed

  • [core] A Qobuz account whose subscription has ended could not connect at all — and was told the wrong reason. The sign-in refused any account Qobuz answers without a streaming-rights block, which is exactly what it removes when a plan ends. The browser was shown "Failed to exchange the authorization code. Please try again." — naming a code that was perfectly good, inviting a retry that could never succeed, while the real cause never left the journal. So the account was locked out of Qobuz entirely, including anything it had bought. Measured on such an account: Qobuz signs it in, answers its purchases, favourites, playlists, search and new releases with 200, and hands out a playable URL for every track. It now connects like any other, plays what Qobuz gives it, and the screen says why the music is short. HIGHRESAUDIO made and undid this same mistake a fortnight earlier; the reasoning is now written down in one place rather than learned twice.

Changed

  • [core] Asking Tidal what format it serves no longer costs a request when the box already holds the track. The format probe that runs before a push to HQPlayer called Tidal even for a track sitting complete in the seek cache — where the relay, one line further on, serves that file without calling Tidal at all. The cached file is the answer: the remux runs -c:a copy -f flac and is only kept on a clean exit, which a non-FLAC stream never gets. Replaying a track through HQPlayer now costs nothing where it used to cost one round trip.

Changed

  • [ui] The application's icon stands beside its name on every screen that shows the name. Two surfaces still showed the wordmark on its own. One is the passkey panel: it replaces the sign-in card's contents entirely, so the lockup somebody had been looking at a second earlier vanished at the exact moment they were asked to authenticate. The other is the start-up screen the app paints while it loads. Both now carry the same row as the sign-in header, from one shared rule rather than three copies of it. The start-up screen keeps a hairline of its own instead of the theme's, and that is not a detail: it paints a fixed ground — #12141c, the colour the installed app declares for itself — whatever appearance the interface is set to, because it is drawn before any theme exists. The icon is a near-black square, so on that ground it needs its edges given back in both appearances. Measured against that ground, the theme's border was worth 1.21:1 under Gravity, 1.28:1 under Minimal and 1.78:1 under Slate, against 15.28:1 for the screen's own — in a dark appearance the outline was there in the stylesheet and invisible on the screen.

[0.9.51] - 2026-09-01

Added

  • [ops] [core] A box now answers to its own name, and its certificate carries that name. Every certificate a box issues already promised <hostname>.local — the one address that keeps working when the router hands the box a different IP, which is the day every bookmark and every home-screen icon breaks at once, an installed app reopening for ever the address it was added from with no address bar and no error page to say why. Until now that name only answered by accident, on boxes whose owner happened to install AirPlay: its Debian package pulls in the service that announces names, and nothing else did. A box used as a local output or an HQPlayer endpoint had no reason to have it, so its certificate named an address nothing on the network answered to. That service is now installed, enabled and started with the rest — and if it cannot start, the install says so and carries on, because a box that cannot announce its name is still a box that works on its address. Two measured corrections came with it: the name is built from the first label of the hostname, which is what the announcing service actually publishes (a box called musics.1 answers to musics.local and never to musics.1.local), and a certificate that does not carry that name is now reissued. Only the address was ever checked before, so a certificate naming the wrong one was served for its full term — up to 825 days — and nothing on a running box would have replaced it. The authority every device was told to trust is not reissued in the process, which is what keeps the phones and computers of the house from all having to be set up again.

  • [core] [ui] The support report answers what a certificate incident actually asks. Five facts were missing, and a real case turned on every one of them: when the served certificate was issued — the day it changes is the day devices trusting it stop, and it was computable only by subtracting the issuing policy from the expiry, by hand; when the local authority was created and its fingerprint — an upgrade never reissues it, so those two tell "this authority changed, and every device stopped at once" from "this one device was never set up"; whether an interface's address is leased or fixed, which is whether it can still move; the name the box announces, asked of the announcing service itself rather than guessed from the hostname; and the date of the last self-update, since "done — 0.9.49 → 0.9.50" says nothing about whether that was an hour ago or last month. The report also flags an announced name its certificate does not carry — both facts were already in it, two sections apart, and nobody diffs them by eye.

  • [core] [ui] Qobuz browses like HIGHRESAUDIO does: five shelves, each opening a strip of its own. Qobuz had four buttons — Favorites, New Releases, Selection, Playlists — two of which were a single hard-coded shelf apiece out of the many its catalogue holds, which is the same defect HIGHRESAUDIO's bar had before it was restructured. The bar is now Favorites, Purchases, Shelves, Playlists, Genres, using the strip mechanism the HIGHRESAUDIO menu already introduced; nothing new was built for it. Shelves lists nine: New Releases, All New Releases, Selection, Qobuzissims, Press Awards, Best Sellers, Most Streamed, Ideal Discography and Harmonia Mundi. Qobuz accepts fifteen names there and states them itself in the body of its own error; six are left out, each for a measured reason — recent-releases answers nothing at all, most-featured returns the same fifteen albums as most-streamed (two names, one shelf), and universal-classic, universal-jazz, universal-jeunesse and universal-chanson are one undifferentiated shelf whose content matches none of the four names, so it could not be labelled truthfully. Purchases is new and is Qobuz's answer to the HIGHRESAUDIO vault, but a simpler one: a bought album is read and played through the ordinary path, ids and all, so nothing about the queue had to change. Playlists gains the account's own beside the ones Qobuz publishes — they were simply never listed, though the route that reads their tracks already worked on them. Qobuz has no third tree: its two published names, editor-picks and last-created, answer the same playlists id for id. Asking for a shelf outside the nine now answers 400 rather than 503 — a request that can never succeed is not a service that is down — which also makes the six excluded ones genuinely unreachable rather than merely undocumented. Genres brings thirteen genres and seventy-eight sub-genres, none of which were reachable before. A genre is not a shelf on Qobuz — it only ever narrows one, which is why the genre grid browses the Selection: measured across eight sub-genres and four genres, a sub-genre returns nothing at all on the new releases and between one and thirty albums on the Selection. Two genres publish one sub-genre or none, which is the catalogue and not a failure.

  • [core] Qobuz search asks once instead of three times. The same words were sent to the same endpoint three times over, once per kind of result, when leaving the kind out fills every one of them in a single answer — measured, 1.08 / 0.49 / 0.29 s for the one call against 2.87 / 0.64 / 0.70 s for the three. Album results also carry their year now, since every listing reads albums through one shared piece of code. There is no advanced search for Qobuz and there will not be one on their current API: the genre filter it accepts on a search is disregarded — four queries crossed with four genres each returned the same ten albums, filtered or not — and so are hires, release_type, order and sort. Filtering on our side would narrow the page of fifty rather than the catalogue, which is the very defect the local library's old Recent button had.

  • [core] [ui] Tidal browses the way the other two services do, and two of its shelves start working again. Tidal had five buttons above the album grid, and two of them were empty: Charts returned nothing at all, while Editorial showed the chart playlists mixed in with its own. The cause was the same for both — the box found their content by matching the word "chart" inside the headings Tidal writes on its own home page, and Tidal has since renamed that row "The Hits". Nothing reads a heading any more: the charts come from Tidal's own Top page, and the editorial shelf now excludes them instead of swallowing them. The bar is now five shelves — Favorites, Shelves, Playlists, Genres, Moods — each opening a strip, as HIGHRESAUDIO and Qobuz already do. Shelves lists the catalogue's own, Playlists gathers its three surfaces on one strip (the editors' selections, the account's own, and the charts), Genres brings twenty that were unreachable before, and Moods is new here — the shelf HIGHRESAUDIO has and Qobuz publishes nothing for. A shelf that holds nothing is left out, and that is decided by asking rather than by believing Tidal: one shelf declares it has albums and answers an error, while another declares the same and holds two hundred and eighty-one playlists, so trusting the declaration would have shown one dead button and hidden a real shelf. Two things Tidal does not have, and that are therefore absent: purchases, since it sells nothing, and sub-genres — its genres are one level deep, so the strip does not drill. One Tidal shelf could not be played at all: Exclusif is two hundred and eighty-one playlists and no album, and every card on it was taken for an album — pressing play asked for the tracks of an album that does not exist, and the favourite star would have filed a playlist among the albums. What a shelf holds is now stated by the box rather than guessed from the shape of a row, on every source. A sixth shelf, Explore, opens Tidal's own navigation: its 42 entries, which include HiRes, seven decades from the 1950s to the 2010s, 52 record labels and its rich genre pages. A page there holds either further pages — Record Labels leads to 52 — or sections of content, so the strip drills the first in place, the way the genres do, and shows the second on a strip of its own. Each section pages properly, through the handle Tidal gives it: asked for without the two parameters it insists on, that handle answers an error that reads exactly like "this list cannot be paged", which is how it came within an inch of being written off as a dozen frozen thumbnails.

Fixed

  • [core] An accented genre no longer lands at the end of a strip that claims to be alphabetical. Seen on screen, on Qobuz's classical sub-genres: Mélodies & Lieder sat between Musique vocale sacrée and Opéra, and Électronique ou concrète was exiled to the far end. The sort only folded case, which leaves an accent at its own code point — é is 233, past every plain letter — so every accented name was pushed behind every unaccented one. It now folds accents too, which is what puts the strip in the order a reader looks for. Not a French curiosity: Qobuz answers its genre names in the language of the account's country, and that country is not something the listener can change from their own account page, so an accented alphabet is the ordinary case there rather than the exception — a German or a Spanish account met the same jumble. A handful of letters carry no accent to fold — Ø, Æ, Œ, Ł, Đ, Þ — and Unicode offers nothing to do with them, so they were stranding Ørjan Nilsen, Æther, Łódź and Þórir behind Zappa: the same symptom, and not a hypothetical on a catalogue carrying labels like Norway's 2L. They are now mapped to the letters a reader files them under. Ligatures and typographic forms come along for the ride, so flute sorts with flute and № 5 with no 5. The sort is shared with HIGHRESAUDIO, whose own lists are unchanged by the fix, and it is deliberately not a full alphabetical collation: languages disagree on where a folded letter belongs — Danish files Ø after Z on purpose — and settling that needs a locale the box has no reason to trust.

Changed

  • [ui] The sign-in card and the footer carry the application's own icon, drawn the same way. Beside the wordmark, the way the licence server's own sign-in sets it — the two now open on the same lockup. It is apple-touch-180.png, not the 512px icon: the same artwork, 61 kB instead of 465, and already in the service worker's precache, so on a box where the app is installed it costs no download at all. Sized at 36px against the mark's 42, which puts it a little taller than the capitals — the proportion that makes an icon and a name read as one block rather than a stamp set beside a word. On a dark card the icon needed help: its artwork is a near-black square, so it lost its edges against a card measured at #0a0a0a under Minimal and #12141c under Gravity, leaving the glyph floating with no bounds. A hairline in the theme's own border colour gives the square back its shape, rather than inventing a second, lighter icon — what the owner sees here stays what they see on their home screen. The footer already drew the same artwork, and drew it differently: 50px filling the bar's whole height, corners at 4 % of the side against the card's 22 %, so one square read as two marks depending on the screen. Both now name one shared rule, and the preview that opens when the footer icon is clicked joins them — its corners were flatter still. The radius is written as a proportion rather than a length, which is what lets the same rule serve a 36px icon and a 192px preview. And that footer icon is finally reachable: it opens a preview, but it was a bare image with a click handler, so no amount of tabbing ever landed on it and nothing announced it as something to press.

Fixed

  • [site] Links to a section of the manual land on the section. The manual's top bar stays on screen as you scroll, and nothing offset the headings for it: following a link to a section scrolled the page so the heading arrived exactly underneath that bar, out of sight. The reader landed mid-paragraph with no title to say where, which reads as a broken link. It affected every anchor in the manual, including the ones chapters use to point at each other and the FAQ on the landing page.
  • [core] Two lines of the support report said something they did not mean. no_license reads as a verdict from the licence server; it is what a box reports when it holds no licence file at all and therefore sent nothing — the ordinary state of a trial, and a hunt for a problem that does not exist. And MPD's last error was printed bare, as if it had just happened: MPD keeps it until something clears it and never dates it, so a stale line was there to be blamed for tonight's symptom. Both now say what they are.

[0.9.50] - 2026-08-31

Added

  • [core] [ui] HIGHRESAUDIO's menu, laid out the way its publisher asked for it. Their own audit put it plainly: eighteen pills ran end to end in a single bar, in the order their API happens to publish them, with the fourteen shop categories spilling off the edge — and the remark that started this was that the list needed organising, not lengthening. The bar is now seven shelves — Favorites, Vault, Categories, Charts, Playlists, Labels, Genres — each opening a strip of its own below it, which is the mechanism the genres already used; nothing new was built for it. Three shelves are new. Charts comes from an endpoint absent from their documentation and found in their own JavaScript client, as does Labels, which lists the seven they named, 2L and audite included — five of those seven serve the same albums as the … Highlights categories, which is their catalogue and not a fault, so nothing is de-duplicated. And Playlists gains Genre and Theme beside Editorial and Mine: not a third tree but two of HRA's own groupings of the selections they publish, so a grouped browse asks for the editorial tree and names the grouping beside it. Their specification presents those two as filters of the playlist endpoint; measured, they are not — that endpoint answers no data found for a genre, and only the dedicated routes reply, which is why the code takes the longer path their document says is unnecessary. Nothing is hidden and nothing is re-arranged beyond the order they asked for: the six categories they named come first, in their sequence, and the other eight follow in HRA's. Three of the fourteen appear nowhere in their plan — Bestsellers, UNAMAS, Featured Album — and all three serve real albums (50, 50 and 26, and the first two share only nine titles, so they are not one list served twice), so removing them would have put catalogue out of reach. The one shelf they asked for that does not exist is Explore: two of its four entries are already categories, and the other two match nothing their API serves — not a category, a genre, a theme, one of the thirty-seven moods, or a label — so it is not built, and the question is with them.
  • [core] The genres and themes of HIGHRESAUDIO's playlists are listed alphabetically. They arrive in an order that follows no rule anyone can read: the identifiers form a complete series yet come back as 4, 7, 5, 8, 1, 6, 3, 2, 9 — neither the identifier, nor the alphabet, nor the number of playlists, and no field in the answer explains it. Stable across repeated calls, so it may well be deliberate on their side; unreadable on ours either way, and a strip whose order cannot be predicted is worse for a listener than one that can. The same treatment the catalogue's own genres already had, at HIGHRESAUDIO's explicit request, and with the same key, so R&B still sits beside Rock. They are being told, so they can say if that order was a ranking after all — restoring it is one line.
  • [core] [ui] HIGHRESAUDIO's advanced search, copied from their own application, defects included. Searching Queen returned a list with little to do with Queen, and filling in a filter with the search box left empty did nothing at all — no result, no message. Both causes were ours. Their advanced endpoint requires no search term (label=ECM, composer=Bach, artist=Queen, genre=Jazz and format=fl192 each fill a page on their own, measured through the box), yet the box refused to ask below three characters — a rule that belongs to their quick search and was carried onto this one by mistake; and the interface only ran a search when the box held something, so the explanation it could have shown was unreachable. The form is now theirs, criterion for criterion: free text — the search box already on screen — artist, composer, label, year, format, mood and the order. Two fields are deliberately absent, and for different reasons. album belongs to their shop's nine-field form — a different back-end — and not to their application's, so copying the application leaves it out; the measurement only confirms nothing was lost, since the streaming API accepts it and disregards it (artist=Queen&release=2026 returned the same thirteen albums id for id with no album, with Innuendo, and with a nonsense value). Genre is absent because it has a shelf of its own on the browse. Year is HRA's own name for the year an album went online, not the year it was recorded — Innuendo is production year 1991 and reads 2026 — and it is named as they name it. The order offers their nine entries; + sorts ascending and - descending, measured on all four criteria, and their own web client labels them the other way round, so the measurement is what the list follows. An order they do not know does not get ignored by them, it returns nothing at all, so the list is closed and anything outside it is refused rather than forwarded. One defect of theirs is passed through untouched, on purpose: setting a format discards the words typed — queen and london in FLAC 192 both return the same fifty albums, the whole 192 catalogue — which is how their application behaves, and a form that quietly behaved differently from the one people already use would be the worse surprise. A second one is passed through the same way: mood is accepted and disregarded — measured on a result small enough not to be capped, the same thirteen albums came back id for id with no mood, with Amused and with Melancholic, two moods of opposite families, and a mood on its own answers nothing at all. It stays because the field is theirs and because that class of defect is not a property of this one control: album and the direction of importDate are the other two. All three are reported to them. What the form does not do is ask a question that cannot be answered: an order with nothing to arrange returns an empty list, so the interface says an order arranges results — add something to search for rather than spending their slowest call to come back empty. The form is applied by its Search button rather than on every keystroke: this is the one slow endpoint they expose (measured on the box, three different cold searches each ran past thirty seconds and came back in about a second on the retry, their side having cached them), and the box runs two of them at a time, so seven controls each firing their own would queue behind one another.

Fixed

  • [ui] HIGHRESAUDIO's playlist covers are shown whole. Their audit said the artwork was a rectangle displayed as a square; measured on thirty-two of them, taken from the oldest to the newest of their catalogue, every single one is 410 × 205 — a banner, not a sleeve. Shown in a square cell, only the middle half survived, and these are compositions laid out across their width: the title and the label's mark are exactly what the crop threw away. The cell now takes the shape of the artwork. Each place keeps whichever dimension its layout owns — a card fixes its width, so the banner is half as tall; a list row fixes its height, so the thumbnail is twice as wide and the rhythm of a list mixing both shapes never changes from one row to the next. One defect came out of that change and was caught on a real screen rather than in a test: the row's cover column named a fixed width, so the wider thumbnail overflowed it and printed itself over the first third of every title. Nothing could have caught it in the test suite — the test environment lays nothing out — so the column now takes its width from the cell, and a guard reads the stylesheet rather than the DOM to refuse a hard-coded one.
  • [ui] The favourite star and the add button no longer sit on two lines in a search result. A result row is a three-column layout — sleeve, text, one place at the end — and an album row on a streaming source carries two controls there, the star and the +. That is a fourth item for three places: the + wrapped onto a second line and came back at the far left, under the sleeve, with the star stranded alone on the right. It has been that way since the star was put on search rows on 2026-07-13, on Qobuz, Tidal and HIGHRESAUDIO alike, and it is the one screen in the application where a row carries both. The two now share a single place. The test environment lays nothing out, which is exactly why nothing caught it, so the guard counts the row's children instead — and refuses the other possible repair, a fourth column, which would leave a gap at the end of every row that carries only one control.
  • [core] A malformed row from HIGHRESAUDIO now costs one album, not a whole shelf. Their answers are not typed, and two of the shelves come from endpoints missing from their documentation altogether. A row that was not an object raised where the album is built and turned the shelf into a server error, rather than into a grid one entry short — the three other places that read their lists already guarded against it.

Changed

  • [ui] A streaming service is now shown by its own mark, and the library stopped pretending to report a state. HIGHRESAUDIO already appeared as its logo above the album grid, because that logo is its name written out; Qobuz, Tidal and Roon were still spelled in capitals beside it. All four now carry their own artwork, in a light and a dark variant apiece, each set at its own size, because these four images do not hold the same thing: their proportions run from 1.4:1 to 7.5:1, the share of the picture the name itself occupies runs from 11 % to 99 %, and the ink each one paints on screen differed by a factor of two. So no single rule — one height, one width, one size of name — makes them comparable; the sizes are a judgement, and the stylesheet carries the measurements it was made from. HIGHRESAUDIO keeps the size its own composition needs, and Tidal stays the odd one: its logotype is spaced so widely that matching the others' weight would make it half again as long as any of them. Each stays an <img> with its name as alt text, never a CSS background: a background paints nothing at all when the file is missing or the browser is in forced-colours mode, and the line would then be empty with no way to tell which source is open. The green pulsing dot that opened that line is gone with them. It had exactly one value — there was no state it could report — so it announced "live" at the one moment that is true by construction: you have just chosen the source and its covers are already on screen. A status light that never changes is a word spent in advance, unavailable the day a source really is in difficulty; and it animated for as long as the page stayed open, on a machine whose CPU budget is a first-order constraint.

[0.9.49] - 2026-08-29

Added

  • [ui] Light or dark, decided before signing in. The appearance switch lived in the configuration panel, which is behind the sign-in form: someone starting the box at night got the light palette full in the face until they were through it, and a reader who prefers dark had no way to say so on the one screen they always meet. A button now sits in the corner of the sign-in card — a moon on a light page, a sun on a dark one, always naming where pressing it takes you. It switches the appearance, not the theme: which of Minimal, Slate or Gravity applies is still chosen in the panel, and the button only flips that theme's palette. The choice is remembered where the pre-paint script reads it, so the next cold load opens dark with no white flash, and the browser's own chrome — the status bar of an app launched from a phone's home screen — follows in the same instant. The sequence behind both controls is now written once and shared, which is what the panel's switch had that the new button was missing: the event that repaints that chrome inside the application. Two faults found on the way, in code that had only ever run before the first paint and was now being called again on every flip: two thirds of the Safari remedy were absent, so a home-screen launch would have kept the old status-bar tint; and an unknown theme name — themes are extensible — threw after the palette and the stored preference had changed, leaving the switch half applied.

Changed

  • [core] [ui] [ops] The interactive API reference is off unless a box asks for it. /docs, /redoc and /openapi.json answered without an API key — they have to, since Swagger fetches the schema from the browser with no header of its own — so every box served the full map of its API to anything that could reach it: measured on a deployed core, 282 kB describing 188 routes, authentication, passkeys, licensing, system and self-update among them. None of it is an access, every route keeps its guard, but it is the floor plan, handed to an unauthenticated caller on a network that may hold a guest's phone or a compromised appliance. A new API_DOCS_ENABLED setting now governs the three routes together — together, because hiding the page while leaving the schema mounted would have removed the convenience and kept the exposure — and it ships off. They are unmounted rather than protected, so a request gets a plain 404: a route that does not exist should not answer 401. Boxes already installed are switched off at their next upgrade, and a box whose owner turned the reference on keeps it. The interface follows without being told: the core's entry point lists the endpoints it actually serves, docs is simply absent when they are not mounted, and the two buttons that opened Swagger disappear — a button leading to a 404 is worse than no button. Development keeps it on, which is also what the tooling that compares API.md to the live schema needs. One caveat for anyone switching it on: the Swagger page loads its own JavaScript from a CDN, so on a box with no route to the internet it opens blank.
  • [ui] The Audiogravity wordmark is set in type, the way the landing page sets it. The interface drew its name from an SVG file, and that file was never a logo: it holds two text nodes asking for Helvetica, never converted to outlines. Every Linux and every Android the interface runs on lacks that face, so the mark was rendered in whatever the system substituted — different from device to device, while looking deliberate on the machine it was authored on — and dark mode had to repaint the image with a filter, because a picture cannot take a colour. The name is now text, in the typeface the interface already ships: same treatment as the landing page, down to the superscript ty, on the splash screen, the title bar, the sidebar, the sign-in card and the credit line of both footers. The splash screen was the last to keep its own version — capitals, wide tracking — and it had a reason the others did not: it paints before any theme exists, so it can read no token. Only its colour needed that exemption, which is why the shared class declares none. Sizes were matched to the capital height the drawn mark occupied, so nothing moved. The footers were a case of their own — they already wrote the brand correctly but without its styling, so it came out at the weight of the link around it with the browser's default superscript, more than twice the size the mark uses. Reaching the weight the landing uses required declaring the Inter faces over the range they actually carry: the files hold every weight from 100 to 900, the stylesheet claimed 400 to 700, and a declared range silently clamps the rest — the heaviest weight the interface asked for was being rendered one step lighter. The declaration costs no bytes: it is the same file.
  • [ui] The pipeline tab is named for what it shows. Its heading read AUDIO DSP PIPELINE, on the desktop layout and on the mobile bar. There is no DSP stage: the placeholder node it once had left the diagram when the three services the box never shipped were removed, and signal processing is not on the roadmap. A title announcing a stage the page cannot show is a promise the page has to break.
  • [ui] The sign-in badge names the core, not "API". API · CONNECTED above the form named a protocol nobody signing in has to think about; what the check actually reaches is the box's own core — the word the installer, the service unit and the rest of the interface already use for it. The footer of the application says CORE: before the address for the same reason; the button beside it keeps "API", since what it opens really is the API reference.
  • [ui] The configuration panel drops the word before its version number. The footer of the panel read Version: v0.9.49-dev; the number and the icon beside it say what they are.

Fixed

  • [ui] Five files the interface claimed to cache, and one it offered as an app icon, did not exist. None of it failed in a way anyone could see. The service worker's app-shell list named three icons under names the repository has never carried — apple-touch-icon.png for apple-touch-180.png, favicon-32x32.png for favicon-32.png, favicon-16x16.png for favicon-16.png — beside two logo_audiogravity_*.png that exist nowhere at all; cache.add() catches its own failure, so this was five 404s on every install and update of the worker, and icons that were never actually precached. The push handler asked for two more absent files, so every notification was drawn with the browser's default glyph instead of the product's icon. And the web manifest offered the retired wordmark SVG as an app icon at sizes: "any" — the entry a browser reaches for first on a scalable surface — which should never have been there in the first place: that file is a logotype five and a half times wider than tall, and an app icon is square. All of them are now asserted against the repository: the precache list, the images the worker names elsewhere, and every manifest icon, shortcuts and screenshots included.
  • [core] An album the cover catalogues have never heard of showed no cover at all — even holding the picture. Reported on HIGHRESAUDIO's Vault, where HRA Mixed Up Sampler 2015.1 played with a blank sleeve in the player while the very same artwork appeared, correctly, on the Up next row two centimetres below. The queue row asks for the cover with the bare image URL; the player asks with the same URL plus the artist and album as a fallback. That fallback was the whole difference. Covers are cached twice — once per track, once per album so one fetch serves a whole record — and the album entry was being consulted before the URL the source had handed over. Cached as a miss it answered "no cover" outright, so the picture was never fetched: measured on the box at under 2 ms per refusal, while the URL served 83 kB of JPEG the moment it was asked directly. A miss on that entry only ever meant MusicBrainz and iTunes know nothing about this album — true of a promotional sampler, and of any obscure release on any source, and no reason to ignore artwork the service itself provided. The album entry may now answer only when it holds a cover; a miss still spares the two slow catalogue searches, which is what it was cached for. And the first track that resolves its own artwork writes it back, repairing the album for every other track instead of leaving them to wait out the cache.

[0.9.48] - 2026-08-29

Added

  • [core] [ui] HIGHRESAUDIO's Vault — the albums you bought, next to the ones you stream. A Vault pill on the HIGHRESAUDIO source lists the account's purchases and plays them like any album: on MPD, on a UPnP renderer, through HQPlayer. HRA keeps purchases in a second tree with its own routes and its own ids — a purchased album is <album>_<transaction>, which the catalogue route refuses, and the catalogue search does not find it — so every Vault id carries a vault: prefix that travels back to the queue, the way a playlist id carries mine: or editorial:, and the proxy reads the prefix to resolve each track on the purchases route. It never reads the id's shape: three segments where the catalogue has two is what one purchased album showed, not something HRA documents, so an unprefixed three-segment id is logged and left alone. The two routes are not interchangeable, measured on the same session and the same purchased track — the purchases route answers, the catalogue route answers NO SUBSCRIPTION without a URL once the account has no subscription — and HRA's own §1.2 names the purchases route for exactly this. One trap on that route is guarded: handed an id of the wrong shape it answers OK with the metadata and no URL, which read as a success would enqueue silence. The purchases' own playlist tree stays unexposed: empty on both accounts available, so untested.
  • [core] [ui] An HIGHRESAUDIO account without a subscription can now sign in, and play what it owns. The backlog said HRA refused such an account at the door; measured on one, HRA answers the login NOK / NO_SUBSCRIPTION and ships a valid session in the same body — one the purchases routes accept, down to a CDN that serves the FLAC (HTTP 206, fLaC magic). It was Audiogravity's own login that refused, on the label alone, and it would then have recorded the account as subscribed. A session is now accepted whenever HRA issues one, the state is recorded in HRA's own words (SUBSCRIPTION / NO SUBSCRIPTION, the second absent from their documentation), and /highresaudio/connection carries has_subscription. The sources card says next to the account that it can play its purchases only, and the browse offers it the Vault alone — the favourites, the playlists and every shelf answer NO SUBSCRIPTION to that session, so a bar that offered them would offer buttons that fail. Search declines the same way — it covers the catalogue, so instead of running a query that can only answer NO SUBSCRIPTION it says so and points at the Vault. And signing a different account in on the sources card resets all of it at once: what the library offers, the cached shelves and the ★ state all follow the account that is actually connected, not the one that was. A core that predates the field is read as subscribed: that is the only state its login could ever have saved.
  • [site] The landing carousel says what each screen is. Fourteen captures went past with nothing but a row of dots to tell them apart: a visitor saw a phone, not a product. Each now carries a sentence beside the device and three or four labels along its left edge, each one at the height of what it names and joined to the screen by a hairline — the source badge, the chain from service to DAC, the format actually being played, the output that can be switched mid-track. The labels are anchored to the capture rather than to the column, so the phone keeps its size and position whatever is written beside it, and their widths follow the free space instead of being fixed, which is what used to cut a sentence mid-word below 1856 px. They do not cross-fade with one another: consecutive screens deliberately share heights, so the outgoing label is gone before the incoming one starts, and only the visible screen's text is reachable — a transparent label is still found by Ctrl+F. Below 1550 px the flanks are too narrow to read and the labels are simply not shown.
  • [site] A DSD-over-HQPlayer capture joins the carousel. UPnP feeding HQPlayer Embedded, DSD128 at 5.6 MHz, and the DAC receiving it — the one path the fourteen other screens never showed. The capture supplied was a full-resolution PNG at 1.5 MB; it ships as a 63 KB WebP at the 760×1497 of every other screen, its phone outline matching theirs to the pixel.
  • [site] The landing says what a finished streamer is made of. A new Board and software section, between Where it sits and How it compares: network streamers sell for anywhere between a few hundred and several tens of thousands of euros, and for the part that makes one a streamer — receiving, decoding, handing the result to a DAC — it is a board running Linux with a software layer on top, most often the very layer Audiogravity drives. The claim is checkable rather than asserted: those components are GPL-licensed, so the manufacturer is obliged to list them, and the reader can settle it on their own machine. It names no brand and concedes the analog side outright — the DAC, the supply, the clocks are the reader's ground, and nothing has to be built: a Raspberry Pi, a fanless mini-PC or an Intel Mac mini with Linux on it is enough. Reuses .rt-list rather than adding a class that would style the same thing.
  • [site] A roadmap that promises nothing. A What's next section between the FAQ and the call to action, where a reader who has just been told the product is in beta asks where it is going: a ready-to-flash image, Diretta, AirPlay sending, multiroom. The wording is deliberately open — the list is not closed, none of it is promised, none of it belongs to the V1 scope, and the order is decided by what owners ask for — so the section ends on a way to answer rather than on a commitment.
  • [manual] The FAQ answers what "Public beta" means. First question of chapter 11, and therefore first on the landing: the V1 scope is complete and every feature described runs today; what remains is correction and optimisation, not missing features. It went in the FAQ rather than the hero because a caveat in the pitch deflates it, while the FAQ is where a cautious reader looks before the price.
  • [core] [ui] HIGHRESAUDIO's editorial playlists arrive on shelves instead of in a heap. There are 1762 of them, and they came as one undifferentiated list in the order HIGHRESAUDIO happens to publish it: finding anything meant scrolling. They file them on four shelves of their own — New Releases, Recommended, Popular, Moods — and a strip under the Editorial tree now picks one, HIGHRESAUDIO doing the narrowing on its side rather than the box pulling 1762 entries to sort them (Popular is 92). The strip opens on All, and that is not a formality: walking the whole tree found fourteen selections filed under no shelf at all — Mr. Slowhand — Eric Clapton, In the Spirit of Jazz, The Sky is Crying — old entries from before the field existed, and among the best of the set. A bar of four shelves alone would have made them unreachable, so the shelves narrow a list rather than dividing one. The four are not enforced either: HIGHRESAUDIO publishes no list of them, so an unknown shelf answers "nothing here" instead of an error, and one they add later works without an update. Leaving the Editorial tree drops the shelf with it — coming back to a narrowed list with no strip on screen to widen it is a dead end. Fixed on the way: an answer HIGHRESAUDIO gives as plain text rather than a list, which every shelf that matched nothing would have turned into a server error, and which five other calls read the same unsafe way.

Changed

  • [core] [ui] HIGHRESAUDIO's shelves, genres and artists, the way HIGHRESAUDIO asked. Their first feedback on the integration, 2026-08-28, named three things. The four shop categories they publish in German were relabelled here with words of our own — Tips, Top Album — and two of the four were wrong; they now carry HIGHRESAUDIO's own English: Listening Tips, Top Albums, New Release, Recently Added. The 26 genres and their sub-genres came in the order their API publishes them, which is no order a reader can use (R&B, Soundtrack, Guitar, Electronic…); both levels are now alphabetical, case-folded so a lower-case title does not sink to the end. A genre is addressed by its path, never by its rank, so nothing else moves. And the artists a search returns were in no order either — measured, twenty of them scattered — and are now alphabetical too, after the cap: the best matches are kept, and shown in the alphabet. Albums keep HIGHRESAUDIO's relevance order; the request was about artists.
  • [ui] A playlist now says it is one. Every streaming service maps its playlists onto the album model so one grid renders both, and nothing on a card told them apart — identical cover, title and byline; the only clue was the shelf heading, far above. HIGHRESAUDIO underlined it: "It is very important that you differentiate between albums and playlists." A playlist card now carries the word in the slot an album gives its year (a playlist has none, so the slot is never contested), a list row says it ahead of the byline, and the glyph behind a missing cover is the list, not the disc. On Qobuz and Tidal as well as HIGHRESAUDIO: the grid already knew which pill it was on to queue the right item type, and that same answer now drives what it shows.
  • [site] The landing said Audiogravity conducts your streamer. It is one. "Your streamer is your source. Audiogravity is the conductor" — the hero, the footer, the kicker's "streamer interface" and the three description / og: / twitter: metadata all described a layer sitting on top of a streamer the reader already owned, which is the opposite of what installing it does. It now opens on "Audiogravity is a streamer. A conductor, not a soloist." — the fact first, then the image that does the explaining, since a conductor plays no instrument. The paragraph under it stops saying it coordinates each component of your stack, which presumed a stack: it installs, configures and conducts them.
  • [site] How it compares claimed not to compete, above a table of competitors. Its lede read "sits at a different layer than the rest — does not compete with Roon, MPD or HQPlayer", directly above a table that lists Roon, Volumio, Moode, HQPlayer and JPLAY iOS and ends on a row of prices. It also put Roon and MPD in one bag, when Roon is a column here and MPD an engine Audiogravity installs. The distinction the sentence has to carry is Roon versus Roon Bridge, and HQPlayer versus its NAA — one of each pair is a rival, the other is what gets installed to be compatible with it. The column header drops Embedded.
  • [site] The pipeline drawing credited the NAA with HQPlayer's work. One box read HQPlayer NAA · DSD256 · upsample, when the NAA receives audio already processed. There are now three nodes where there was one — Audiogravity resolving the stream URL, HQPlayer applying filters and conversion, the NAA handing the result to the DAC. The lit path starts from Qobuz, Tidal and HIGHRESAUDIO, spelled in full, instead of from Roon Core, which is one source among five. Internet radio, the NAS and both UPnP roles join the drawing — the control app driving the box on one side, the network renderer being driven on the other — and dashed frames mark what actually runs on the box, so the radio, the streaming clients and the two UPnP roles stop reading as outside boxes. There are two such frames, carrying the same label rather than one each: the NAA endpoint runs on the box as much as the rest, but HQPlayer sits between them and may run on another machine, so a single frame could not hold both without claiming HQPlayer too. The amplifier makes way: eight nodes did not fit across 1000 px, and the bit-perfect link moves to NAA → DAC, where the property is decided.
  • [site] Three more places still said the reader owned a streamer. The hero was corrected; the same presumption survived in Features — "every layer of your streamer", "every engine your streamer runs" — and in the title of Installation, "On your streamer in one command". They now read every layer of the stack, every engine the box runs, and "One command, and it is a streamer.", which makes the installation title demonstrate the hero instead of contradicting its premise. One occurrence is deliberately untouched: "expose your streamer as a UPnP renderer" does mean the Audiogravity machine.
  • [site] Five layers did not add up. The title claimed Audiogravity "covers three, manages two", over a table labelled 2 × Audiogravity, 2 × manages, 1 × tunes — the count only worked by lumping the two layers it is with the one it tunes, and a reader who counted the labels found 2/2/1 and stalled. It now reads "is two, drives three": two layers it is, three it drives, five in all. True quoted away from its table, which the old wording was not.
  • [site] The link preview showed an opening line the site no longer uses. og:image was a still of the page taken by hand months ago: it still advertised "Your streamer is your source. Audiogravity is the conductor" while the og:description beside it already carried the new one, so a pasted link contradicted itself. Regenerated — and, more to the point, regenerable: audiogravity.ops/scripts/gen-og-image.js drives the headless Chromium that audiogravity.ui already carries. Four settings decide whether the result is usable, and each was established by looking at the output: captured at 1800×945 and downscaled, since a direct 1200×630 shot crops the phone and the price row; the carousel frozen on slide 0, which otherwise advances every 3.5 s mid-capture; dark theme forced, a fresh profile having no stored preference; and the announce bar hidden, because it carries a version number and the card is cached for weeks.
  • [site] Where it is developed, and by whom. Who I am now ends on "I develop and maintain it myself, from France" — the badge in the footer stays Made in the EU, France being in it.
  • [site] The carousel opens on a different record. The opening slide is the only one every visitor is guaranteed to see, so which sleeve is on screen is an editorial choice, not whatever happened to be playing when the capture was taken. A reader pointed out that the one there read as the opposite of what a listener is after, which is a fair thing to say about a shop window. It now opens on the HIGHRESAUDIO screen, and the Qobuz screen that opened before moves one slide down. Captions and annotations travel with their capture and never with the slot — --y is measured against the capture itself — so the move changes nothing but the data-slide number and which caption carries active; checked in the markup, fourteen slides, contiguous indices, a single active caption with its three annotations, and the opening capture still loaded eagerly as the page's largest paint. Both it and the slide after it are light-theme captures, so the carousel now opens on two light screens in a row.

Removed

  • [site] The carousel no longer shows Qobuz twice. Two of its captures were the same Qobuz screen — same badge, same Qobuz → MPD → USB → DAC chain, same DAC — differing only by the theme, so a visitor letting it run saw the same thing twice, five screens apart. The dark copy is set aside rather than deleted; the light one stays, one slide in.

Fixed

  • [ui] A box that was switched off told its owner the password was wrong. On an iPhone or an iPad only, and for a reason nothing on screen could reveal: the sign-in screen recognised a failed request by reading the browser's error sentence, and that sentence differs per engine — Failed to fetch in Chromium, NetworkError when attempting to fetch resource. in Gecko, Load failed in WebKit. The list knew two of the three, and includes('failed') swept WebKit's into Invalid username or password. So a phone whose box had stopped invited its owner to doubt what they had typed, while the API · OFFLINE badge two centimetres above the message said the truth. It now reads the shape of the failure rather than its wording: the only line that can know a request never left — the one that called fetch — marks it as it throws, and every screen downstream asks instead of guessing. A TypeError raised three frames later, in code handling a perfectly successful response, stays what it is.
  • [ui] The sign-in screen no longer answers with a status line. Whatever goes wrong, it now says something a person can act on, and each sentence was checked against the response the server really sends. A box that cannot be reached — no answer at all, or the 502 its own front returns when the core is stopped — names the address that was tried, so the reader goes to look at the machine instead of writing to support. A box that is running and has just crashed says so, and points at the support report: an unhandled error reaches Starlette, which answers in plain text with no message of its own, and that is a different event from a machine that is off. Six wrong passwords in a minute, which is exactly what someone does when they think they are mistyping, said HTTP 429; it now names the wait. A refused field said [object Object], because the core describes those in an array and the array was handed straight to the screen.
  • [ui] The retry button on the passkey panel did nothing. The panel a passkey-registered phone meets before the form re-enabled its button after a failure, while the listener had been registered to fire once and was already gone — so the second attempt it invited was one it had made impossible, until the page was reloaded. Nothing looked wrong: the button came back exactly as it had been. Double submissions are prevented by disabling it for the length of the attempt, which is the guard that belongs there.
  • [ui] One fetch boundary, and every screen reads it. The sign-in fix tagged a transport failure where the request is made; it did so in three places, each written by hand, next to four different ways of turning a refused response into an Error — and the differences were visible. The licence upload sat on its own fetch call, so an iPad in front of a switched-off box read WebKit's raw Load failed there, the very sentence the sign-in fix had just retired; the passkey panel fabricated a detail from statusText, so it answered the same outage as the password form with a different sentence, over HTTP/1.1 only; the main client always set a detail, so the rule written on "no detail" could never fire for it. There is now one boundary in net-errors.js — fetchOrThrow, with throwForStatus giving every caller the same shape: the status, a string detail or null, a 422's field list joined into one line and kept raw underneath, slowapi's rate limit read from where slowapi writes it. The API client, sign-in, passkeys, uploads, downloads and the licence panel all go through it, and the message helper every toast uses reads the same predicates — a 502 no longer says Bad gateway on one screen and cannot reach the box on the next. The three screens that printed err.message on a passkey or upload failure now say what the box did — or that it could not be reached.
  • [ui] The installed app on a phone can say the box is off. The service worker answers a request that could not leave the device with a synthetic 503 {"error":"offline"} rather than letting the browser reject — an iOS quirk, and an old one. Read as an HTTP status it is a lie: nothing answered, and every screen of the installed app therefore said Service temporarily unavailable in front of a machine that was simply switched off, with no hint that the network was the matter. The boundary recognises that marker and turns it back into the transport failure it stands for, so the home-screen app reports an unreachable box exactly as a browser tab does.
  • [ui] A request is retried only when nothing answered — and not when the phone already knows. The retry layer used "no status" as its test for a transport failure, which also matched a 200 whose body was not JSON — replayed three times, seven seconds, for a response that would never parse. It now asks the boundary. Two refinements followed review: the service worker's offline answer is final, so it is not retried at all (the installed app on a phone was making three attempts per call, from every poller on the page, at a box it had already given up on); and a connection that drops while a body is being read — a core restarting under a long request — is a transport failure and gets its one retry back.
  • [ui] Six toasts in the Settings passkey section had their arguments reversed. showToast(type, title, message) was called as (message, type): the toast received "Passkey removed" as its type and "success" as its heading, and rendered unstyled under the word success. Nothing failed — a wrong string is still a string — which is why it lasted. Fixed, and a source guard now refuses any toast whose first argument is not one of the four types.
  • [ops] The landing's FAQ generator swept the end of the chapter into the last answer. sync-landing-faq.py splits 11-faq.md on ## and takes everything after the final heading as that question's body — so the horizontal rule closing the chapter and the trademark notice under it landed inside "How do I update Audiogravity?", published as a literal --- and a paragraph still wearing its Markdown asterisks. The fault was latent, not visible: it is the sync that publishes it, not the edit, so it surfaced the first time the FAQ was regenerated after that notice was added. A top-level rule now closes the answers.
  • [site] The manual's footer was a filled band, but only for readers whose system is dark. The manual is a reading page and its footer is meant to be a single rule with the way back below it; under dark it came out as a #1a1a1a block spanning the column, as if borrowed from the commercial page. The stylesheet said background: none all along — the landing paints every <footer> through two rules of unequal weight, and the manual's own rule beat one of them and lost to the other. Which one applied depended on how the reader reached dark: picking dark in the interface gave the correct footer, leaving the browser to follow the system gave the band. The same theme rendered two ways, which is why nobody caught it. The manual's rule is now restated at each of the two weights; the landing keeps its own filled footer, unchanged.

[0.9.47] - 2026-08-25

Added

  • [lic] A trial length, once served to a box, is never taken back. The trial every box starts with is set on the licence server and served to the fleet, signed, at each box's daily check-in — and the box keeps no memory of it: it recomputes "days remaining" from whatever length it is served that day. So raising the setting reached every box, which is what one wants, but lowering it did too: a box at day forty of a sixty-day trial was back to "expired" the morning after the setting went to thirty, with nothing said to anyone. The server now serves each box the longest length it has ever told it — the current setting the first time a box asks, the larger of the two from then on. Raising the setting still reaches every box, trials already running included; lowering it reaches only boxes the server has never seen. Recorded from a box's very first fetch, because a box asks for its configuration before it first reports itself, and a length served in that window is a length served. The administration console shows the length granted to each installation, finds the ones served a given figure, and can forget one — the way to undo a length typed by mistake, since the setting alone can no longer take it back; the setting itself is capped at the ten years the box would honour anyway. Nothing changes on the boxes: the rule lives on the server, and every box already deployed follows it at its next check-in.
  • [lic] The licence server backs up its database every night, and checks the copy. Everything but the signing key lives in one file — orders, the licence files that can be re-sent, revocations, the trial record of every box — and the only copy ever taken was the one made around an upgrade. A timer now runs at 03:30: it copies the live database with SQLite's own online backup, which is consistent even while the server is writing where a plain copy of the file is not, verifies the copy's integrity before keeping it — a backup that cannot be opened is not one, and the time to learn that is not the day it is needed — compresses it, and keeps the newest thirty. A night missed while the machine was off is caught up at the next boot. Fresh installs get it; existing servers get it at their next update, which also takes a snapshot before restarting into the new code, so a migration that goes wrong has a copy from before it. Losing this file stops no box from playing — a licence is verified on the box — but it loses the record of who bought what, and every revocation: a revoked licence's own file is still validly signed, so without the record it reads as valid again.

Removed

  • [core] [ui] Three services Audiogravity claimed to manage, and never shipped. audiogravity-pulse, audiogravity-camilladsp and audiogravity-visualizer were written into the code as units the box always manages — since the very first commit — while nothing in any installer ever created them: no unit file, no binary, no package, on either platform. They cost a widened systemd query every ten seconds, and they were why the support report printed a full realtime-tuning line for three pieces of software that are not on the machine. CamillaDSP also held a reserved node in the audio pipeline diagram, the last trace of a processing stage that was never built. DSP is not on the roadmap and would be designed from scratch if it ever were, so the placeholder is gone rather than kept as a promise the diagram could not keep.

Fixed

  • [ui] Three downloads that could hand you nothing, without a word. Exporting a configuration, saving a service's config file, and fetching your licence file after activation each built their own download by hand, and each got the same two details wrong. A browser cannot write to your disk: a page has to put the content in memory, make a temporary address for it, click an invisible link, then release the memory — and click() does not download, it asks the browser to, which the browser does a moment later. Releasing the memory on the next line races it to the data and wins often enough; the click reports success and no file is ever written. One of the three did not even put its link in the page, which several browsers ignore outright — and swallowed any error on top. Nothing failed loudly, which is why it survived: it works on a fast machine with a small file, and misses on a phone, or on a bigger one. All three now go through the one helper that gets it right, alongside the support report, which was the only caller already using it. Two of the tests that covered this were asserting the faulty behaviour — they required the memory to be released immediately — so a correct fix would have looked like a regression to whoever tried it; they now check what that screen actually owes: the right content, under the right name.
  • [ui] HIGHRESAUDIO's mark is theirs again, at their contrast. The artwork shipped since 0.9.44 was cut out by hand from a flat image, and the cut had washed out the RES at the centre of the mark — its strokes measured 141 against the original's 59, so the one word the whole logo is built around sat at half contrast, close to vanishing into its own white plate. HIGHRESAUDIO supplied their official cut-out file on 2026-08-25; the three images the interface uses are regenerated from it, at the source's own proportions, and the wordmark reads as it was drawn. The light-theme variant still has to be derived — they publish only the negative version, white on transparency, which is invisible on a light background — but it now derives from a clean original: white ink turned dark, the centre plate left white with its dark RES, and a dark outline of the thickness the logo already uses on its bottom square, so the cross stays legible on white without inverting a single colour. The small badge beside Qobuz and Tidal is redrawn from the same source at proper proportions, and joins the mask treatment those two already use: at 24 pixels the mark measures 24×18 and none of it is legible either way, so what a reader gets is a silhouette — and a silhouette that follows the text colour stays visible on a selected card, whose background inverts to black under the light theme, where a fixed dark image would not. Measured in all four combinations of theme and selection.
  • [core] [ui] The support report was answering some questions wrongly, rather than not at all. A report that says nothing sends you looking; a report that says something false sends you the wrong way, and five of its sections did the second. Its streaming section read not signed in for Qobuz, Tidal and HighResAudio on every box, whatever the truth — the probe called a property as though it were a method, the resulting error fell into a catch that wrote "no", and nothing anywhere said so; on the machine where this was found, all three accounts were connected. Its scheduling section described the realtime tuning of software the box does not have, because systemctl show answers for any unit name at all with the defaults it would apply, and nothing there could tell a stopped service from one that was never installed. Its list of detected outputs printed bare card names, so a machine with an onboard codec showed PCH twice and HDMI three times — five lines nobody can act on, and the analog/digital pair is precisely where an AirPlay receiver ends up on the wrong socket. It showed which output Audiogravity believes it selected and never which one each service is really playing on, so the one fault most worth reporting — a box playing out of a socket nobody chose — was the one it could not show. And it repeated an identical failure once per line, twenty-one lines for seven occurrences of a single denial, in the order journalctl stores them, which is not the order they happened. Now: a probe that cannot run says unknown and names its error instead of answering "no"; a unit that is not loaded says so, and only a genuinely absent one is called absent — a masked unit is present, and deliberately so; every output carries its full label and its hw: address; every service shows the device its own configuration names, flagged when that disagrees with the selection on record, and for MPD that is the output MPD has enabled rather than the first one written in the file; and repeated failures fold into a single line carrying their count and the last time they occurred, in chronological order. The report lost a tenth of its size on the way.
  • [core] The now-playing view survived a removed pipeline node type by luck, not design. Removing the unused processing node left one reference to it on the now-playing path, which would have raised on every refresh. Caught by the test suite before release; the whole enumeration is now checked by a scan that finds any reference to a member that no longer exists.
  • [lic] The licence console spoke French to pick a file. Restoring a database offered Choisir le fichier — aucun fichier sélectionné in the middle of an English console. That text was never ours: a file field is the one control a page cannot dress or translate, and the browser draws it in its own language, so the console read differently depending on who opened it. The field is now put out of sight and a real button stands in its place, with the chosen file's name written beside it — in the console's own words and its own style, where the field used to be a grey slab that matched nothing around it. Out of sight, not out of reach: it is clipped rather than removed, so it keeps its place in the tab order and the button still answers a keyboard, which is the trap in every recipe for hiding one.
  • [lic] Cancelling a database restore let go of the file — the field did not. Choosing a backup and then cancelling left its name showing next to a Restore database button that had gone grey, which reads as a file still selected and a button refusing to work. The console had in fact forgotten it; only the field had not, because the line meant to empty it did nothing at all: it looked for the field in a place this screen does not use, then called something a file field does not have, and the whole thing failed silently. It is emptied for real now, on a cancel and after a restore that went through — a file that has been applied must stop being on offer.

[0.9.46] - 2026-08-24

Added

  • [core] [ui] The support report now answers the questions a box could not be asked. It said what the box was — version, licence, services, configuration files — and nothing about the world it lives in, so half of every support exchange was still spent asking. It now measures, on demand: how the interface is really served (plain HTTP or TLS, on which port, which version of the interface is deployed, whether the certificate has expired, whether it still names the address the box answers on, and whether the authority phones were asked to trust is the one signing it — all asked of the running server, never read from a configuration file that may not match); the network the music crosses (gateway, DNS, which interface actually carries the traffic, the negotiated speed of a wired link — a gigabit port running at a hundred is a thing that happens — the WiFi rate and signal, whether the clock is synchronised at all, and whether the internet and the licence server actually answer, with the time they took); the audio path as it is at that instant rather than as it was configured — which DAC the system can see and at what USB speed, and the format and sample rate genuinely crossing ALSA, which is the bit-perfect claim, measured; how the audio services are scheduled, with what systemd was told to apply next to what the running process actually carries, since a tuning file that failed to apply is exactly the difference between those two; what else is on the network — UPnP renderers and servers, an HQPlayer or a Roon Core, whether or not this box was configured to use them, because "not configured" and "nowhere to be found" are different answers; free space, and any filesystem the kernel has quietly remounted read-only; whether the deployed settings file carries what the code expects, by name, never by value; whether a self-update could actually download, asked of the release repository itself; and the last seven days of failures — the services' own, and the kernel's, where a DAC dropping off the bus, a dying memory card, an undervolted supply or thermal throttling speak, and where nothing else in the report would ever show them. Also the boot history, because five boots in two days is a hardware conversation, not a software one. Secrets are removed as before, and every measurement that fails says so: a probe that could not run is never printed as a negative finding.
  • [ui] The support report window says what it is doing, and when it is done. Collecting now takes a few seconds — it probes the network — and the window said only "Collecting…", motionless, which reads as frozen. There is now a turning indicator and the sentence names what is being probed while it runs, then gives way to a plain "Report ready." in the affirmative colour. The Copy and Download buttons stay unavailable until there is something to copy.
  • [core] [site] OpenHome is a switch now. The landing has said it from the start: OpenHome apps such as Kazoo work too, once OpenHome is switched on — and that "once" meant finding the right key in the raw configuration file, which is Expert-mode territory for what is a yes/no question. The UPnP bridge's form now carries an OpenHome Renderer checkbox: tick it, save, and a second renderer appears for Linn Kazoo and its kind, named after your box with an -OH tail so the two renderers a box then publishes never wear the same name in an app that lists both — and switching it on or off never renames the UPnP renderer your other apps already know. Verified live: renderer names measured on the network in both positions of the switch.

Fixed

  • [ops] Audiogravity can create its own backup folder again. Every configuration change makes a timestamped copy first, and the folder those copies live in is created at startup — except the permission granted covered the sub-folders and not the folder itself, so the very first attempt was refused, silently, on every box. Nothing broke while the folder happened to exist from an earlier install; on a fresh box the backups depended on luck. Found by the new report's journal section on its first run, which is the job it was written for.
  • [core] [ui] The box announces itself under the name you chose — not a hostname with a protocol suffix. Every box provisioned by Audiogravity appeared in JPLAY, BubbleUPnP or mconnect as hostname-UPnP/AV — a name nobody picked, wearing a protocol badge nobody asked for, and for someone who drives their streamer from their usual control app, that name is the product. The configuration form made it worse: its one name field promised "Name displayed to UPnP clients" and did not deliver — upmpdcli keeps two name keys, the form only wrote the one the suffix gets glued to, so typing Salon produced Salon-UPnP/AV, and anyone who came looking for the suffix found nothing to fix. The form's Friendly Name now writes what control points actually display, exactly: type Salon, see Salon. A fresh box defaults to audiogravity-XXXX with a short token of its own, unique per box, because two boxes in one home must not publish the same name and UPnP — unlike AirPlay — has no mechanism to tell them apart; the hostname is not used, since on a fresh image it is the distribution's name, which nobody chose either. An existing box migrates the moment its form is saved — any field, any reason — and its suffix disappears; a name someone had deliberately written into the file, odd ending and all, is kept as it is. The Reset to default action says up front that the name returns to its default: a reset is a reset, and it announces its one visible consequence. And measured on the real network before and after, on both platforms — not deduced from documentation.
  • [core] Saving from the configuration form no longer flips a tile to NOT CONFIGURED. The form's save rewrote the file from its parsed values, which dropped every comment — including the one line saying Managed by Audiogravity. That line is how the box tells a config it wrote from a distro default someone may have edited: lose it and the tile's badge flipped to NOT CONFIGURED, the support report called the file hand-written, and the natural repair — regenerate — was exactly the gesture that would overwrite the edit just saved. The marker now survives every form save. It is preserved, never invented: a config Audiogravity did not provision does not start claiming otherwise, because the support report's word has to stay trustworthy — and if the box ever cannot re-read the file to decide, it says so in its log instead of deciding silently.
  • [core] Checkboxes in the configuration form tell the truth in both directions. A 0 in the file showed as a checked box — the display asked "is there a value?" where it meant "is it on?" — and saving a checkbox wrote True into a file whose program only reads 0 and 1. Nobody had hit it yet because no toggle was exposed; the OpenHome switch below would have shipped both bugs on day one. Boolean settings now arrive at the form as real booleans and go back in the file's own spelling, for every service's form at once.

[0.9.45] - 2026-08-23

Fixed

  • [ui] The Configuration tab shows the result of a first-time setup, without a reload. The setup reported success, the three services came back configured — and the tab behind it was exactly the one from before: the invitation banner still there, the tiles still NOT CONFIGURED, the UPnP bridge still given as stopped. It took reloading the page by hand to see what had actually happened, at the very moment an owner is looking for confirmation that something did. The refresh had been written, and read line by line, and it was correct; it simply never ran. The setup panel lives inside a dialog, and a listener written by one component but rendered by another is called on the one that renders it — so both handlers looked for methods on the dialog, found nothing, and threw where nobody could see it. The box was reporting itself configured the whole time. The same trap was closing the installation-log dialog by luck rather than by design: its own close button wrote to the wrong object, and the dialog only shut because a parent happened to be listening — on its own it stayed open. Both are bound to their owner now, and the tab asks the box for its state once per setup instead of twice.

  • [core] [ui] A box with no music of its own can now be set up. The first-time setup refused to finish until a music folder was named, and the box refused too — yet a box with no local music is an ordinary thing: HIGHRESAUDIO, Qobuz and Tidal need no file on disk, and neither does an AirPlay receiver or a UPnP bridge. Their owners were pushed into inventing a path to get past the button, and an invented path went through: the only check was that it held no double quote, which protects MPD's configuration file rather than the person typing. What followed looked like a complete success — configuration written, MPD running, tile marked CONFIGURED, indexing reporting "finished" because there was nothing to walk — for a library that was empty, with nothing anywhere saying why. The setup now offers No music library — online listening only as a choice of its own, and MPD's music folder line is left out entirely rather than written blank, which would stop MPD from starting. The reverse is possible too: a box that started without a library can be pointed at one later, and one that has a library can be released from it — with a confirmation first, since that takes something away. A path that is given is now checked: absolute, existing, a directory, and readable by the account MPD runs as — a folder you can read is not necessarily one MPD can, and that difference is invisible once MPD is running. Each refusal says which of those failed, and says it before anything on the box has been touched.

  • [core] A library that was never configured no longer reads as a library that is empty. MPD refuses every database question on a box with no music folder, and that refusal carries none of the fields Audiogravity was looking for — so it came back as "no albums" and "no results". On a box whose collection had simply never been set up, the Library tab said the collection was empty, and search said your music did not contain what you were looking for. Both now answer with the reason: No music library is configured on this box. Add one from the Config tab, or keep listening to your streaming services. A genuinely empty collection is still reported as empty, and an album whose title happens to contain those words is not mistaken for the refusal.

  • [core] [ui] A configuration tile now says when the software behind it is not installed. The Services and Profiles tabs have always been able to name absent software; the Configuration tab could not. It showed a tile for a package that is not on the box, with the path of a file that does not exist, a button offering to edit it, and a download button with nothing to fetch — no date, no state, no badge, and nothing saying why. An owner could as easily conclude that something was broken as that the software was there. Such a tile is now dashed and greyed, carries an UNAVAILABLE badge, says the package is not installed and links to the Audio Software tab where it is installed. What can still be done, still can: a package removed without purging leaves its configuration file behind, and that file remains downloadable, because a file that is on the box can always be taken off it. What cannot, is refused: editing is closed, since saving a configuration also restarts the service, and no restart can succeed for software that is absent. Backups are not lost either — they come back with the package. The reverse case is left alone: an installed service whose file is missing keeps its editor, which is what creates the file.

  • [core] [ui] A stopped service shows as stopped in the Configuration tab, instead of showing nothing. The tab read each service's state from systemd's list of loaded units, which leaves out units that are installed but idle — upmpdcli on an ordinary box among them. Those tiles carried no state badge at all, and nothing distinguished "stopped" from "not installed there". The state is now read per unit, so RUNNING, STOPPED and FAILED appear wherever they apply. Two side effects worth having: the tab asks the box one question instead of two, and the answer is read over the system bus rather than by running a command — measured at no process creation where the previous approach created six per refresh, on a path the audio pipeline also uses. Coming back to the tab after installing a package now re-reads the list, so the tile you were sent to fix stops being stale.

  • [ops] [core] A box that comes back from an install or an update can play. Installing the MPD package is not the same as having its daemon: Debian arms the service when it actually installs the package, and does nothing when the package is already there. So a box carrying MPD with its service stopped — or simply not set to start with the machine — went through the whole installation with nothing noticing, and came out with an interface in front of something that could make no sound. Every installation and every update now starts MPD and sets it to start with the box. It sits with the switch that turns off MPD's on-demand activation, because that is what the two decisions are: taking away the mechanism that used to start the daemon by itself is what makes arming it necessary. A box with no network, which reaches that point with no MPD at all, is told what is missing rather than having its installation aborted. If you run yours as a Roon endpoint or an HQPlayer NAA only, stopping MPD is not enough — turn its ENABLED switch off too, or it returns at the next boot; an update turns both back on.

  • [core] A stopped MPD says so, everywhere, instead of failing anonymously. Everything Audiogravity plays goes through MPD, and the box stops it itself when you choose the Roon or the HQPlayer profile — so a stopped daemon is a state an owner reaches on purpose, and the only failure they can fix in one click. It was the one failure the box could not name: pressing play answered Playback failed / Queue operation failed, which says nothing, and the library answered with no albums at all — a statement that the collection was gone, on a box that may hold fifty thousand tracks. Playing, queueing, removing a track, searching, casting to a network renderer and starting a radio station now all answer with the same sentence: MPD is not running. Start it from the Services tab. A refusal is confirmed a second later before being named, because Audiogravity restarts MPD itself when you change output and the port refuses during that second exactly like a stopped daemon — an instruction to go and start what is already coming back would be wrong. A daemon that is listening and not answering is a different failure and is not reported as this one.

Added

  • [core] [ui] A support report you can read, and send if you want to. Helping someone with their box meant asking what it looked like, one question per day: the version deduced from a timestamp, a service assumed to be off from a screenshot, a hand-edited configuration guessed at from a description — and some of those deductions were wrong, which sent the search after a fault that was not there. System › Actions › Support Report now gathers it all in one gesture: version and architecture, the licence verdict and the last check against the licence server, uptime and load, which audio packages are installed and at which versions, the state of every service and whether it starts with the box, the outputs detected and the one each service is pinned to, whether each configuration file is Audiogravity-managed or was written by hand, the declared music library and whether MPD has actually indexed it — and the configuration files themselves, which is what nobody could see. Any unit that has failed elsewhere on the box is listed too: a NAS mount that never came up explains an empty library, and never appears in the audio stack. Passwords, API keys, streaming tokens and share credentials are removed before the report is built — the name of a setting stays so you can still see it is configured, its value does not — and so is the content of your library. Comments are dropped, which takes upmpdcli's configuration from three hundred and eighty-five lines to thirteen and makes it readable; what was cut is always counted, never silently. Nothing is sent anywhere: the report is shown to you, with Copy and Download, and travels only if you attach it to a message yourself. A part the box cannot collect says so in its own place, rather than taking the whole report down with it.

  • [site] [ui] What the box tells the licence server is now written down. Once a day, and when a key is activated, a box contacts the licence server — that is how a Pro licence stays bound to one machine, and how revocations, announcements and update offers arrive. What travels was written nowhere: not in the licence agreement, which had seven clauses and none about data, not on the landing, not in the interface. The manual now says it, in the Licence section of the Administration chapter, and the licence agreement carries a clause of its own: the Device ID, the primary MAC address, the operating system and architecture, the Audiogravity version, the current licence status and, during the trial, its start date — plus the licence file itself when it is verified, and the hostname you type when activating. The server records the address the request arrives from. Your music, your library paths, your service credentials and what you listen to stay on the box. Also said, because it is the part that is easy to get wrong: whether Pro is unlocked is decided on the box, from the signature of its licence file — no round-trip decides it.

  • [core] [ui] HIGHRESAUDIO's catalogue can be browsed by genre. The service arranges the same music a second way — twenty-six genres, a hundred and eighty-six sub-genres — and none of it was reachable. A Genres button now opens a second row under the shelves: the genres, then, once one is chosen, that genre and its own subdivisions, with a way back to the list. Choosing the whole genre reads as All, because HIGHRESAUDIO gives two of its genres a sub-genre of the same name and two identical buttons side by side read as a mistake; the heading above the albums names where you are — Soundtrack · Original Score. A genre they publish with nothing in it says so rather than looking broken.

  • [core] [ui] HIGHRESAUDIO's playlists can be browsed and played — theirs and yours. The service keeps two separate collections: the selections its editors publish, two hundred of them, and the playlists belonging to your own account. Neither was reachable, where Qobuz and Tidal have had their Playlists shelf from the start. A Playlists button now opens a second row with the two, and a playlist behaves exactly like an album — open it to see its tracks, or send the whole thing to the queue in one gesture, on the box's own DAC as on a network renderer or an HQPlayer. The two collections number their playlists independently, so the same number exists in both; each one now carries the name of its own collection, and nothing anywhere has to guess which of the two a playlist came from. An account that has no playlist of its own says so, rather than looking broken. Creating and renaming playlists, and adding or removing tracks, are not part of this: what HIGHRESAUDIO has made is readable, what you make stays where you made it, for now.

  • [core] [ui] A HIGHRESAUDIO search can be narrowed by composer or by record label. Two fields under the search box, for the two questions the words alone cannot answer: the composer of a classical recording, which appears on no other field, and the label — ECM, Pentatone — that a collector browses by. They narrow the search rather than replacing it, and the ordinary search is untouched while both are empty. Two other filters the service offers are deliberately absent: choosing an audio format makes HIGHRESAUDIO discard the words typed altogether — three different searches, one of them nonsense, return the same fifty albums — and a mood takes about a minute to answer the first time. Both have been reported to them, and both come back the day they answer differently.

  • [core] [ui] Every HIGHRESAUDIO shelf is reachable, under the name HIGHRESAUDIO gives it. Their catalogue is arranged in fourteen shop categories; the browse offered three of them, and had renamed two of those three — High-Res Essentials was shown as Discover, Editors Choice as Editor's Picks. The three were written into the interface, so the eleven others could not be reached at all: the label highlights, ECM, Pentatone, Sony, Universal and Warner; the bestsellers; the editorial tips; what has just been added. The core now lists the categories and the browse builds one button per category from that answer, in the order HIGHRESAUDIO publishes them — so a shelf they add tomorrow appears without an update, and every name on screen is the one their own listeners recognise. Four of their titles come back in German whatever language is asked for, which is a quirk of their API their own application works around the same way: those four are relabelled for display (Hörtipps → Tips), and the original title stays what the box uses to ask for the shelf.

Fixed

  • [core] Searching HIGHRESAUDIO can no longer get in the way of the music. Their filtered search is slow the first time it is asked something — measured between nine seconds and a minute, then instant on a repeat once their own cache is warm — so it now waits thirty seconds rather than fifteen, and says "try again, it usually returns at once" instead of reporting a failure. But a request that waits is a request that holds a connection, and every exchange with HIGHRESAUDIO shares the same handful of them, including the one that resolves a track when you press play. At most two searches now run at a time, so pressing play while someone is searching can never be the thing that fails. A box that cannot reach HIGHRESAUDIO at all is told exactly that, rather than being invited to try again for as long as it stays offline.
  • [ui] A search that fails now says what happened. The interface had no way of showing an error at all: HIGHRESAUDIO answering "this took too long, try again" appeared as No results for … — a different statement entirely, and the wrong advice. Failures are shown as themselves now. A search of one or two characters with a filter set is refused with its reason too, instead of returning nothing: their catalogue cannot be searched on fewer than three, which made U2 look like a catalogue with no U2 in it.
  • [ui] A row of buttons that continues off-screen now says so. The browse filters sit on a strip that scrolls sideways, with its scrollbar hidden — deliberately, since a scrollbar drawn across labels that short is worse than none. That was survivable while the longest row held five buttons; with fifteen it left the last one sliced through the middle of a word and no sign that anything followed. The edge that still hides buttons now fades into the page instead of being cut, and where a mouse or trackpad is used — a screen you drag needs no button — a chevron sits in each margin to move the strip one shelf at a time. Neither ever claims more than there is: on a row that fits, there is no fade and no chevron at all, and each disappears the moment its end of the strip is reached. Under Firefox the same strips were also drawing a real scrollbar, which the rule hiding it never covered.
  • [core] A HIGHRESAUDIO account signed in elsewhere no longer empties your own shelves. The service allows one signed-in device per account, so any other phone or computer signing in ends the box's session — an ordinary event, and one Audiogravity is built to recover from by signing back in on its own. It recovered from two of the three ways HIGHRESAUDIO announces a finished session, and not from the third, which is the one their account-scoped answers use: your playlists, your favourites, your purchases and your downloads. The box did not notice, did not sign back in, and answered with an empty list — telling you the account owns no favourites, on an account that owns some. All three forms are now recognised. Verified against a genuinely ended session rather than a simulated one: signing in from elsewhere, checking the service refuses the box's old session, and watching the shelves come back untouched.
  • [core] A momentary refusal from HIGHRESAUDIO no longer costs a restart. The list of shelves is fetched once and kept, since it does not change while a box is on. But HIGHRESAUDIO answers some failures — a session it has just invalidated among them — with a perfectly successful reply containing nothing, and that empty answer was kept exactly like a real one: from then on every shelf was empty, on a box that had no way of knowing why and no way back short of restarting. Only a real answer is kept now, so the next visit asks again, and the browse's own Refresh is enough to repair a bar that came up bare.

[0.9.44] - 2026-08-21

Added

  • [core] [ui] [ops] MPD is part of Audiogravity, not one of its options. The installer set up Audiogravity and none of the engines it conducts, so a fresh box reached the interface with nothing able to play: no output to choose, no music, and nothing on screen saying that the missing piece was one button away under Audio Software. But MPD is not an engine among others — the local library, the queue, internet radio and all three streaming services reach the DAC through it, and even the UPnP bridge is a front-end driving it. It is now installed with the box, and the interface no longer offers to remove it; the request is refused by the core as well, for anyone calling the route directly. A box used purely as a Roon endpoint or an HQPlayer NAA can stop the service instead — a stopped daemon costs nothing, and the setup stays possible without a box that cannot play at all.

Fixed

  • [core] [ui] HIGHRESAUDIO is shown as its owner defines it. Above the album grid the source was announced by the plain word HRA, their logo appeared nowhere on that screen, and elsewhere the name was set as Highresaudio. Their mark now heads their section — their own positive and negative artwork, chosen by theme, with its flat background made transparent so it sits on any of the interface's backgrounds instead of painting a rectangle over them; the drawing itself is untouched. Where the name is written it is now written in full, in capitals — now-playing badge, queue header, Sources card — measured as fitting uncut in the narrowest of them. A brand's owner defines its spelling, and the corporate identity supplied with their API is part of what allows the integration to exist.
  • [site] [ops] A box left with a dead access token now repairs itself. Boxes installed during early access hold the shared token in their configuration and present it on every update. Since the releases repository became public that credential can only hurt: sending nothing succeeds, while sending something GitHub refuses is answered 401 — so those boxes were about to fail every one-click update, and could not fetch the fix that would free them, the fix travelling in the very download the token was blocking. The bootstrap installers, which a box re-downloads at each attempt, now test the token they are handed and continue without it when GitHub refuses it — only on a refusal; a network outage stays an error rather than being quietly passed over. The token is dropped rather than forwarded, and the installer clears the stored value when no token is given, so the dead credential leaves the configuration instead of being ignored for ever. One update is enough, with nobody opening a terminal on the box.
  • [ops] The unattended update refuses to execute anything that is not a script. It fetched each installer straight into a shell, so a truncated transfer or a proxy error page would have been run — during an update, where nobody is watching the screen. Both downloads are now checked for a shebang first, and the failure is written to the update log.
  • [site] An address that does not exist on the website now says so. Any unknown path answered 200 with the whole home page instead of a 404 — the hosting serves the site's index.html whenever a request matches no file, and the site carried no 404.html to serve in its place. Harmless on a normal page; not here, where the entire install path is scripts fetched by their address and piped into a shell: a stale or mistyped one delivered ninety kilobytes of HTML to bash rather than failing, and curl --fail could not help, the server having announced success. The site now carries a real 404 page, which also points at the current installer names for anyone arriving on the ones used before the August rename.
  • [site] The all-in-one installer refuses to execute anything that is not a script. It fetched each half of the install straight into a shell, so anything the address returned was run. Both halves are now downloaded, checked for a shebang, and only then executed; a download that fails stops the install instead of letting it finish on a cheerful "installed", which a pipeline used to allow by reporting only the last command's status.
  • [site] The all-in-one installer no longer ends on an address that does not work. It closed by telling you to open https://<this-host>/, written into the script rather than derived from anything. The ui installer, two lines above, had already printed the address it actually deployed — protocol, host and port — and put that same address in the QR code. The two agreed only where a box serves HTTPS on the standard port; a box left in HTTP mode behind a reverse proxy was sent somewhere nothing listens. The closing line now points at the address that was just printed.

Changed

  • [site] [ops] Installing Audiogravity no longer needs an access token. Early access shipped behind a private releases repository, so every installation had to carry a personal access token: a 90-character credential that had to reach the new owner by e-mail, be pasted into a command line, and then live on in the box's configuration for one-click self-update. It was the same token for everyone, it expired on a fixed date, and a single mishandled copy revoked it for every owner at once — which is exactly what happened. The releases repository is now public and the token is gone from the whole path: --token is accepted but optional in all three installers, and the install command is a single short line anyone can type by hand. Boxes installed during early access still hold the old token in their configuration and must have it cleared once — an invalid credential is refused even when the repository is public, where sending none succeeds.

Fixed

  • [site] The install command on the landing page survives a copy-paste. The indentation of its continuation lines was written with non-breaking spaces, which browsers copy literally and shells do not read as whitespace: the pasted command handed curl a second, unresolvable address and answered curl: (6) Could not resolve host: before doing anything useful. The lines carry real spaces now, kept intact by the stylesheet rather than by an entity.

[0.9.43] - 2026-08-19

Fixed

  • [ui] The note explaining a dash on a service card now actually appears. 0.9.42 replaced an untrue 0 with a dash wherever a counter is switched off, and added a line under the Services title saying what the dash means and how to get the figure back. The dashes arrived; the line never did. The interface splits the box's measurement report into one message per service before delivering it, and the fact that says this machine counts no memory at all travels on the report as a whole — so it was being dropped on the way. Two tabs had been waiting on that whole report since v0.9.4 without anything ever sending it: the Services tab could never explain its dashes, and the Profiles tab never noticed a service starting or stopping until it was reopened. Both are fed now. Most visible on a Raspberry Pi, whose kernel starts with the memory counter off, where every card shows a dash for memory and the reason is exactly what the missing line was there to give.
  • [ui] A lost connection to the box now says so in the browser console. When the live stream dropped, the diagnostic message logged undefined instead of naming the failure: the worker sent it under one label and the page read another. Two of its three error paths were already right; this was the third, and the one an owner actually meets.

[0.9.42] - 2026-08-19

Fixed

  • [core] [ui] A figure nothing measured is no longer shown as zero. Three of the numbers on a service card depend on counters that can be switched off, and Audiogravity published all three absences as 0 under a flat graph — which is exactly what an idle service looks like. So a box measuring nothing looked like a box doing nothing. The cards now show a dash, with no graph, and the Services tab says once what a dash means and how to get the figure back. Disk and network are counted per service, and only once IO Accounting and IP Accounting are enabled for it in the Systemd tab — two switches, effective immediately. Memory depends on the machine: a Raspberry Pi kernel starts with that counter off, which no setting in the app can change, so the note links to the manual and its one-line boot change. CPU and tasks are always measured and are untouched. The manual gains both explanations, in Administration and in Troubleshooting.

[0.9.41] - 2026-08-18

Fixed

  • [core] The Roon retry no longer works your box for nothing, nor hides the one thing worth saying. The retry that 0.9.40 introduced was too eager and too quiet at once. Too eager: a box whose Roon Core is simply switched off probed for it every single minute, for as long as the box stayed on — a socket test, a network broadcast and four system queries a minute, on a machine tuned to leave the music alone. It now eases off to a quarter-hourly check while nothing answers, and returns to the quick cadence the moment you open the Roon card, so the click you make inside Roon is still noticed within the minute. Too quiet: the journal reported an unchanged situation once, which was the intent, but keyed that on a value another part of the box overwrites — so a Roon Core that died in the middle of a session, the one event an owner needs recorded, was the one guaranteed to leave no trace. It now always says so, and a Core that moved to another address is named again rather than silenced by an older report. A connection attempt that hangs — a Core that answers the door and then never finishes the introduction — is also confined to a single dedicated thread; before, one such attempt per minute quietly consumed the pool of workers the rest of the box shares, until nothing else could run and the Pipeline tab stopped answering altogether.
  • [ui] The Pipeline tab can be configured from a phone. Its CONFIG button — the way into the description of your hi-fi chain — existed only in the desktop layout, so on a phone there was no way in at all. That is the device most people have in hand when they notice something is off, and it is where an empty signal path sends you. The Pipeline tab now carries the same title bar on a phone, with INFO and CONFIG, and a guest still cannot edit anything. Writing a whole chain remains far easier at a keyboard — the description is a text file — but a phone is no longer a dead end.
  • [ui] [core] A signal path with nothing in it now says why. On a phone, the Pipeline tab showed the track playing and then a blank space where the chain of devices belongs — no box, no converter, no speakers, and nothing to explain the gap. The view draws a device only when music actually flows through one of the connections you have described, and a new box carries an example chain declaring a converter over USB and optical. A box playing through a HAT board — a converter mounted on a Raspberry Pi's connector — matches neither, so every device reads as idle and the whole path disappears. Measured on two boxes, both blank for exactly that reason. The panel now names the situation instead: it puts the two side by side — playing through: HiFiBerry DAC+ Pro against your chain declares: USB Audio Output, Optical Output — so you can see which of them has to change, and points at CONFIG just above. When everything does match and no music is on its way, it says that too. The manual gains the same explanation, in the Outputs chapter and in Troubleshooting.

[0.9.40] - 2026-08-17

Added

  • [ui] [core] The Roon card says where you are, and what is left to do. The last step of connecting to Roon happens inside Roon — enabling the extension — and the box cannot see you do it, so an expanded Roon card said "No zones available" and stopped there. The person who got through it read the box's journal over SSH to find out what was expected of him. The card now names the situation: no Roon Bridge running on this box, no Core answering, waiting for your authorization — with the extension's name to look for, straight from the box so the two can never say different things — or connected, with the number of zones. When it is waiting, I have enabled it checks again on the spot instead of making you wait for the next look, and the zone list refreshes itself the moment a session comes up.

Changed

  • [core] Setting Roon up no longer means opening a terminal. It used to take an SSH session, a configuration file, your Roon Core's IP address typed in by hand, and a restart of the box's backend — for a product whose whole point is that you never do that. Now: install Roon Bridge from the Audio Software page, then enable the Audiogravity extension in Roon → Settings → Extensions. Nothing else. Your box finds your Core on the network by itself (measured: five seconds), and Roon's own Enable click is what grants access — it always was, so asking for the same permission a second time in a file only made the feature hard to reach. The manual's Roon section is rewritten around the two steps. Three settings are gone with it, and none of them did what its name promised: the port was never read (Roon fixes it, and the value shipped was a different port entirely, closed on the Core), the switch duplicated Roon's own authorization, and the extension's name only made it harder to find in Roon's list — one box announced itself as "Aty" while its owner looked for "Audiogravity". Boxes that ran Roon Server rather than Roon Bridge were checked too: they get the same treatment, where an earlier version of this change would have shut them out.

Fixed

  • [core] A Roon Core started after the box is now picked up on its own. The box tried to connect exactly once, at startup: turn your Core on afterwards and Roon stayed absent until someone restarted the backend. It retries now, quietly and at most once a minute, without ever making the interface wait for it. The log stopped crying wolf too — waiting for you to enable the extension is written as the ordinary step it is, and names the extension to look for, instead of appearing as an error next to the real ones.

[0.9.39] - 2026-08-17

Fixed

  • [ui] Closing a browser tab no longer writes a crash report in the box's journal. Every time a browser dropped a connection it had finished with, the server that hands out the interface logged a full Python error trace — 29 of them in three days on one box, all identical, all harmless. The plain case was already handled; over an encrypted connection the same thing arrives under a different name, which was not. Nothing was ever broken, but a journal full of tracebacks hides the day something genuinely is. Only that one cause is silenced: a real encryption fault still shows.
  • [ops] A service you stop now stays stopped. Stopping MPD — from its card, or by switching to a profile that leaves it out — lasted under two seconds on a recently installed box: it was running again before the page had finished refreshing. Measured on one: stopped at 07:40:32, running again at 07:40:34. Debian ships MPD as two units, the daemon and a second one that holds its port and starts the daemon the moment anything connects to it; your box connects to that port continuously to report what is playing, so it revived what it had just been asked to stop. Audiogravity writes MPD's own configuration with the address and port it must listen on, so that second unit does nothing the daemon does not already do for itself — it is locked out at installation, and MPD now runs when, and only when, Audiogravity says so. Boxes that never showed the fault escaped it only because something had already switched that unit off; they get the same treatment at their next update, and uninstalling Audiogravity puts the unit back — running again at once, or from the next restart if MPD is holding the port by then. One side effect worth knowing: that second unit also decided where MPD could be reached, and it accepted connections from the whole network. MPD now listens where Audiogravity's own configuration has always said it should — on the box itself only. Nothing you use through Audiogravity changes: the interface, the UPnP renderer and playback all reach MPD from the box (measured on a box where MPD has always been local-only: the renderer is on the network and works). What stops working is a third-party MPD application on a phone or another computer talking to the box directly — which used to be possible by accident of the packaging, on a port that asks for no password.
  • [ui] The configuration editor stays in the view you chose. Opening a service's configuration lands on Guided, as it should; choosing Structured or Expert worked for about a second, then the editor slid back to Guided on its own — and any edit begun in the meantime went with it, without the usual question about discarding changes. Nothing about the click was at fault: the box streams each service's running state continuously to keep the status dots on the tiles current, and the editor mistook one of those refreshes for "a different service has been opened", which is the one moment it is supposed to return to the default view. It now tells the two apart by the service itself. Opening another service still starts you on Guided.
  • [ops] The end of the installation stops sending iPhone owners to a dead end. It printed the address, a QR code and "accept the browser warning to look around" — then explained, further down than most people read, how to trust the box's certificate. On an iPhone that order is not merely inefficient: Safari offers no way past the warning, so the address does not open at all and the QR code leads nowhere. The certificate step now comes first, and says which case you are in — required on an iPhone, optional-but-useful on Android and on a computer, where the warning can still be waved through for a look. The QR code follows, once the address will actually open. The user manual said the same thing in two places and is corrected with it.

[0.9.38] - 2026-08-16

Fixed

  • [core] The box now checks for new versions every day, and refreshes what it knows before answering. The check ran exactly once, a couple of seconds after the box started: one left on for three weeks answered with what was true the day it booted, and the "update available" notification could only ever appear in that one moment. It now repeats daily. More importantly, it first asks the system to re-read its package sources — the versions it reports come from a local list that nothing else necessarily refreshes, and whether anything does depends on the distribution: DietPi switches those automatic refreshes off, Raspberry Pi OS leaves them on. Same box, same code, two different answers. Measured on a DietPi box: the main Debian index was five weeks old, so a security update to MPD would have stayed invisible that whole time. Pressing Check updates now refreshes too, which is what the button was always taken to mean. The cost is bounded: the refresh is skipped when the lists are already recent (measured at 4.6 seconds for twelve sources when it does run), and an update already announced is not announced again every day.
  • [core] An installation no longer fails because of a repository that has nothing to do with it. Installing anything began by refreshing the system's package lists, and that step reports failure if any configured source is unreachable — a developer tools repository, a Node.js one, whatever else the box carries. The install stopped there, on a package that was very often already available locally. It now carries on and lets the package manager itself say whether it can find what was asked for. The refresh also waits its turn instead of failing when another operation is already using the package system, which is what used to happen when a check and an install overlapped.
  • [core] A download that cannot fit is refused before it starts. Nothing measured the free space, so a disk filling up mid-download left a truncated file — and on an update, the old version had already been replaced by the time the tool complained. The check asks for the size of the download plus as much again for unpacking, and never refuses when it cannot measure.
  • [ui] A card says when a service is installed but not configured. Installing a package does not configure it — that is a separate, deliberate step — so a freshly installed service runs on the settings its package ships and can play to the wrong output while looking perfectly ready. Nothing said so. The card now does, for the services Audiogravity manages, and only when it genuinely knows: a page that could not read the audio configuration says nothing rather than accusing everything. "Configured" means the file carries Audiogravity's own marker, not merely that a file exists — every package ships one.
  • [core] An upgrade keeps the configuration already on the box, and now says so on every path. All four audio configurations are files the packaging system tracks, and it asks what to do when a new default arrives. Only the plain-repository path stated the answer; the other three left it to the default behaviour, which happens to keep the existing file when nothing can be asked — an accident of the environment rather than a decision, and one that can surface as a failed operation instead. All four now state it.
  • [ui] The confirmation says when an operation will interrupt playback. Updating a package restarts the service it drives — the package's own installation script does it — and uninstalling stops it outright. That was not said anywhere, so an update chosen mid-album simply cut the music. The dialogue now says which service is about to be restarted or stopped, for the packages Audiogravity actually drives.
  • [ui] A card whose last operation failed offers a way out of it. It offered nothing at all — no retry, no removal — on the one card that most needs an action. What it offers now depends on what failed: an installation that failed leaves nothing on the box, so the card offers to try again; an update that failed leaves the previous version in place, so it offers to update or to remove.
  • [ui] An update badge can no longer outlive the update. The browser kept its own copy of the versions it had last been told about, with no expiry, and re-applied it over a fresh "nothing new" answer from the box — so a version that had stopped being offered could keep a badge lit on that browser indefinitely. The box now remembers them itself and re-checks daily, so the copy served no purpose.
  • [core] HQPlayer NAA finally reports the version its publisher offers, instead of the one you already have. It is not part of any package repository, so the system's package manager answered with the version already installed — Audiogravity compared a version with itself and always concluded "up to date". Measured on a box: 5.1.5-67 installed while Signalyst published 6.1.4-71, two major versions later, with nothing on screen to say so. The code that reads the publisher's own page had been written all along; nothing ever called it. Its update also no longer uninstalls first: the newer package is installed over the old one, so the audio output no longer disappears for the length of a download, and a failed download no longer leaves the box with nothing.
  • [core] Roon can be updated from the interface at all. Roon publishes no version number anywhere — its download has a fixed name and carries no version file — so the interface, which insisted on a number to compare, refused the action and left no way to update it. Updating Roon means re-running its own installer, which always fetches the current build; that is what the button now offers, saying plainly that there is nothing to compare and which build you currently have. A package that should announce a version and does not is still refused: that is a symptom, and a blind reinstall would hide it.
  • [core] A rebuilt UPnP bridge for 64-bit ARM now reaches boxes without a new release. Upstream publishes nothing for that architecture, so Audiogravity builds and hosts the packages itself — but their file names carry the version, and those names were written into the release. Publishing a rebuild therefore changed nothing until the next release edited them: the quick operation held hostage by the slow one. The file list now comes from the checksum manifest already published beside the packages and already downloaded to verify them, which also gives the version to display. The manifest is checked against the packages the release expects, so an unexpected, missing or duplicated entry stops the install before anything is downloaded.
  • [core] [ops] Uninstalling Roon Bridge now actually removes it, and its version is shown. On a box where Roon was installed by DietPi rather than by Audiogravity, the software card showed no version at all, and uninstalling reported success while deleting nothing — the program stayed on disk, complete, and only its service entry went away. The cause was a single assumption written down as a fact: that Roon lives in /opt/RoonBridge with its data under /var/roon. That is true of Roon's own installer, the one Audiogravity uses; DietPi's installs the same software in /opt/roonbridge with its data under /mnt/dietpi_userdata. Neither is the "right" one, and the distinction is not the machine or its operating system but who installed it — so a box running DietPi where Audiogravity installs Roon itself lands back on the first layout. Both paths are now read from the running service, which knows, and both are re-read after every install and uninstall so they cannot describe a state that has moved on. Removal goes through a small privileged helper that works the paths out for itself rather than being handed them, and it never touches the shared folder that sits above them — on a box with both Roon products, that folder holds the other one's data. It also refuses to report a removal it did not make: deleting a folder that is not there succeeds silently, and announcing that as done is exactly how the original fault stayed invisible — the service entry, the only thing that could still have found the files, was deleted right afterwards. If a previous attempt already removed that entry, the helper looks for both layouts by name rather than giving up on the one it exists for, and says plainly that the data folder can no longer be worked out.
  • [core] What your box is told it can install is now checked against what the sources actually publish. The list of architectures a package supports was read from the names of the folders on the software publisher's server. Upstream upmpdcli has a folder for 32-bit ARM that contains only its add-ons, not the program itself — and the repository's own manifest declares two architectures, neither of them that one. So every box carried "32-bit ARM supported" for a package apt would have refused to install. The manifest is now read, which is what apt itself obeys, and the presence of the actual package is confirmed for your architecture before it is offered.
  • [core] A ten-second network outage no longer marks software as permanently incompatible. If a publisher's site could not be reached while the list of installable software was being worked out, the package was recorded as unsupported — indistinguishable from software that genuinely has no build for your machine, and with no way back short of pressing the refresh button by hand. A box installed without internet came up with HQPlayer NAA, the UPnP bridge and Roon all marked unsupported, and nothing said why. The two are now told apart and the reason is written down and shown on the card: unreachable is unreachable, and it is worth trying again.
  • [core] [ops] Installing the UPnP bridge on a fresh box works. Adding a publisher's repository means writing two files owned by the system — a signing key and a source entry — and the permission the box grants its own service covered neither. The install failed at its first privileged step, on any box that did not already happen to have both files. Both are now granted, in exactly the shape the rest of the system already uses.
  • [ops] The permissions the box grants its own service name the files they apply to. The two new permissions — writing a repository's signing key and its source entry — named their destination with a wildcard. The permission system matches those without treating / as a boundary, so a wildcarded destination also matched paths climbing out of the folder it appeared to describe. Both destinations are fixed values, so both are now written out in full. The permissions file is also re-read after being written, and removed if it does not parse: a single malformed line disables the whole file, taking with it the permissions the rest of Audiogravity already depends on.
  • [core] The same repository is no longer declared twice. Audiogravity checked whether its own source file existed before writing one. A box where the same repository was already configured under another name, or in the newer file format, got a second declaration of it — which apt reports as a duplicate on every update. Any existing declaration of that repository is now found, whatever it is called and whichever format it uses.

Changed

  • [core] [ui] Roon Server is no longer offered on a box that already runs Roon Bridge, and the reverse. Roon Server already contains everything Roon Bridge does, so installing both means two Roon endpoints on one machine competing for the same audio output — and Roon's own installer neither detects the other product nor warns about it. The card now says which of the two is in the way instead of leaving a greyed-out button unexplained. A box that already has both keeps showing both, and keeps its uninstall and update buttons: the rule governs what can be installed, it does not take away the means of undoing a situation it exists to prevent.
  • [ui] The page's own INFO panel describes what the page now does. It still said that a card is dimmed when your architecture is unsupported — one of four reasons now — and, on checking versions, only that you should press a button, with nothing about the box checking daily on its own. In-app help is read before the manual, so help that describes the previous behaviour sends people looking for work the software has already done. The user manual's Audio Software section is updated in the same breath.
  • [core] A refused installation now says why it was refused. Whatever the reason — no build for your machine, a source that could not be reached, a conflicting package — the answer was always "not supported on this architecture", which for a conflict is simply untrue.
  • [core] The generated software list records what it was worked out against. The file the box builds for itself — which publisher folder to download from, which repository suite to use — carried no trace of the system release, architecture or date it was resolved for. Nothing could tell a current file from one left over from before an operating-system upgrade, still pointing at the previous release's downloads. It now says.
  • [ui] The interface no longer highlights its own labels when you press and hold, or drag across it. Audiogravity is a control console rather than a document, but a browser treats every button label, tab and section title as text waiting to be selected. A long press on a phone raised a blue highlight and the Copy / Look Up bubble over a button you were only trying to press; a mouse drag across the screen left a trail of highlighted words behind it. Selection is now off across the interface — and deliberately left on everywhere copying is the point: anything you type into, the details of an event, the differences shown before a service change is applied, and the short technical values such as an order number or a media server's address. The logs and the Device ID already had their own copy buttons and keep them. The terminal is untouched: it has always drawn its own selection rather than the browser's, and keeping the browser out of its way is precisely what leaves its copying intact.

Removed

  • [core] The "Test Package (Dummy)" card is gone from the Audio Software page. It installed nothing and existed only to exercise the install workflow during development, but it was declared alongside the real software, so every box carried it and every owner saw it. Removing it from the list the release ships is enough for a new installation; an existing box also rebuilds that list from the copy already on it, so both are cleared on upgrade — otherwise the card would come back the first time the page was refreshed.

[0.9.37] - 2026-08-13

Fixed

  • [core] The volume comes back after a DSD album — before, it stayed at maximum. DSD is a 1-bit stream that any attenuation degrades, so Audiogravity raises every volume to 100 while one plays and puts them back afterwards. The putting-back had a hole: it only ran if something was still playing at that exact moment. When a DSD album simply reached its end, or was stopped before choosing something else, the restore was skipped — and the levels it was holding were discarded in the same breath, so no later moment could recover them. Everything that played next did so at 100, and the only way back was to set the volume by hand. Measured on a box: 58 before the album, 100 during, 58 lost. It now restores whether or not the music is still running, and never forgets a level it has not put back. One nuance decided the shape of the fix: a player that fails to answer looks exactly like a player that stopped, and writing a level on the first would attenuate a stream still in flight — so the end of DSD is now confirmed by asking, once, rather than assumed from silence.
  • [ui] The volume slider no longer snaps back to a level you already left. Drag it and, a moment after letting go, it could jump back to somewhere in the middle of your own gesture before settling on the value you chose. The previous release fixed one cause of this and it does hold — measured, nothing arrives out of order any more. What remained is that a drag sends one command per movement, so for about a second after you let go the box is still publishing the levels your finger passed through. The slider was covering that with a fixed one-and-a-half second pause, which measured out at a quarter-second of margin: enough on an idle box, not enough on a busy one, a slower one, or after a longer drag. It now holds the level you asked for until the box confirms that exact level, so the margin no longer decides. The same gap existed in the bottom player, where nothing had gone wrong yet — it is closed there too.

[0.9.36] - 2026-08-13

Added

  • [core] Qobuz, Tidal and HIGHRESAUDIO play through HQPlayer. Choosing HQPlayer as your output used to forbid all three: the answer was an immediate refusal saying the route did not exist, and turning HQPlayer off was the only way to hear them — a hard choice between your streaming subscription and your DSP engine. Nothing had to be invented to lift it. Your local library, internet radio and media servers already reach HQPlayer as a web address it fetches by itself, and the streaming services already publish exactly such an address on your own network, because network speakers have been pulling them that way for months. The same mechanism now points at them. Qobuz and HIGHRESAUDIO go one better: the address redirects straight to the service's own servers, so your box hands over an address and steps aside rather than carrying hours of music — measured at four percent of one processor core during playback, which is the cost of watching, not of relaying. Tidal is the exception and has to pass through the box, because it delivers its audio in a form that must be converted first; measured on a genuine first listen, HQPlayer plays that conversion while it is still being written, knows the full track length, and never runs its buffer dry. Nothing is re-encoded on any of the three paths: what reaches your DAC is bit-for-bit what the service sent.

Fixed

  • [core] With HQPlayer as your output, the title, the artist and the cover follow the track — they used to stay on the first one for the whole album. Play an album and the progress bar moved on correctly at every track change while the screen kept showing track one, from the first note to the last. This was never fixable by asking HQPlayer: it reads no tags at all from a stream it fetches over the network, and answers "HTTP stream" for a perfectly tagged Hi-Res file. So Audiogravity now keeps the list it sent and reads which entry HQPlayer is on — the answer comes from its own side. It follows on everything it pushes: your library, media servers and the three streaming services alike. Two situations still show no title, and both are honest: a playback you started from HQPlayer's own remote, which Audiogravity never sent and cannot name, and a playlist you rearrange from HQPlayer itself — noticed, and left unlabelled rather than captioned with the wrong album. Internet radio is a separate matter and still shows the station rather than the song, because there it is HQPlayer that receives the stream and it keeps only the station's name.
  • [core] Playing an album while another one plays now replaces it, instead of queueing behind it. With HQPlayer as your output, starting a second album left the first one playing and added the new one after it — while the screen showed the new album's title and cover. You were looking at one record and listening to another, with no way to tell which. The command that empties HQPlayer's queue turns out not to remove the track that is loaded and playing; it is now stopped first, which empties it completely. This affected every source that goes through HQPlayer, your own library included.
  • [core] A track found by searching a media server plays through HQPlayer. With HQPlayer as your output, playing a search result from a UPnP server failed outright: the request was handed to the local library index, which was asked to find a file named after a web address and, reasonably, found nothing. Playing the same track from the server's folders always worked — only the search results were affected, and the failure looked like a broken server rather than a wrong turn inside Audiogravity.
  • [core] Tidal in a quality that cannot play now says so, instead of going quiet. Tidal's lossy qualities deliver AAC, which cannot be converted losslessly — so nothing came out, on any output, with no explanation. Audiogravity now checks what Tidal actually serves before starting, not what your account asked for, and names the cause. That distinction matters: an album unavailable in lossless is served as AAC even when your quality setting is right, and that is precisely the case a setting cannot warn you about. The check costs one request per album, not one per track.
  • [core] A refusal names the track the way your screen shows it. When HQPlayer cannot decode a format, the message quoted the file name; it now quotes the title, and an album still refuses as a whole rather than stopping halfway through.
  • [ui] [ops] The certificate you are told to install can now actually be fetched — on an iPhone it could not be, at all. Boxes installed with HTTPS were pointed at https://<box>/ca.crt to collect the authority to trust. That address cannot work, and the reason is circular: fetching it over HTTPS means already trusting the very certificate you are coming to collect. Safari on iOS refuses outright — no warning to accept, nothing to tap past — so the procedure was impossible on the device it was written for, and with it the app install that depends on it. The box now hands the authority out over plain HTTP on port 8081, and that port serves exactly one public file: every other address on it is redirected straight to the HTTPS interface — no interface, no API, no proxy in the clear. Being unencrypted costs nothing here, and is the point: a certificate authority is public by design — it is the thing you are meant to hand out — and it carries no key. If something else already occupies the port, the box notes it and carries on playing.
  • [ops] [site] The steps printed by the installer and written in the manual were wrong, and had never been performed. They were written from memory rather than from a device, so they named menus that do not exist and stopped short of the one action that matters. The iPhone steps are now the ones verified end to end on a phone: open the address in Safari and allow the download, then Settings → Profile Downloaded → Install, then Settings → General → About → Certificate Trust Settings and switch on Audiogravity Local CA — installing the profile alone changes nothing, that last switch is what makes the box trusted. The Android instructions no longer give a menu path at all: it differs between Android versions and between manufacturers, so any fixed path would be wrong for most phones. They describe the goal instead — install the file as a trusted authority from the system settings — with the two things that actually stop people: the phone must have a screen lock or Android refuses to store an authority, and since Android 11 only the Settings app may start the install, so tapping the downloaded file does nothing. The manual states plainly that, unlike iPhone, the Android path has not been verified on a device.

[0.9.35] - 2026-08-12

Added

  • [core] You can now move around inside a track played through HQPlayer, and its progress bar works. The bar was there but inert: the elapsed time counted up, the bar never moved, and dragging it did nothing. Audiogravity had HQPlayer down as unable to seek, and never asked it for a track's length. Asking its remote-control connection directly showed both assumptions were wrong — it accepts a jump to any position, and it works out a track's length by itself, including for a stream nobody described to it. Both are now used: the bar has a scale, it advances, and you can drag it. A live stream still shows none, which is the honest answer for something with no end. Two things follow that are worth knowing: the length HQPlayer measures is preferred over the one the source supplied, because it is right more often; and a playback started from HQPlayer's own remote — which used to show nothing at all — now at least shows how long the track is and where you are in it.

Fixed

  • [core] A track from your media server played through HQPlayer shows its length, and the progress bar moves. The elapsed time counted up correctly while the bar sat frozen at the far left and the remaining time read −0:00 — a bar with no total has nothing to measure progress against. Your server had published the length all along: for the track this was found on, it announces seven minutes twenty. Audiogravity read that value, then lost it twice over on the way to the screen — once because the record it fills when you browse a server had no room for it, and once because the record it fills when you search never copied it across. Both paths now carry it through. A track whose server publishes no length, or a live stream that has none, still shows none — which is the honest answer.
  • [core] The volume slider no longer jumps back to where it started before landing on the value you chose. Dragging it produced a brief, unmistakable flick back to the old position. The player describes itself to your screens from three places at once — a regular check, an immediate refresh after each command, and network-speaker events — and nothing kept them in order. A check that began reading just before your change finished reading just after, so it published the old volume after the new one had already gone out, and every screen dutifully followed it back. Measured on a real drag: 47 published, then 48, then 46. Each description now carries the instant its reading began, and one that started earlier than something already published is dropped. This was never only about volume — any control could be walked back the same way — but volume is where a drag makes it happen almost every time.
  • [ui] [security] The interface no longer hands out its own private key to anyone on your network. The box's certificate and its private key were stored inside the folder the interface serves its pages from, and the server made no distinction: asking it for /ssl/key.pem returned the key itself, in full, to any device on the network — no password, no login. That key is what proves the box is the box; with it, someone on your network could impersonate it. The store now lives outside anything the interface serves, an upgrade moves it there and deletes the exposed copy, and the server refuses that whole path as a second lock — including the spellings that try to slip past it. Only the public certificate, the one your devices are meant to install, is published.
  • [ops] The certificate your box makes for itself can finally be trusted — which is what lets it become a real app. If you install with HTTPS, the box signs its own certificate, and browsers warn about it because no authority vouches for it. Accepting that warning was enough to look around, so the problem stayed invisible; it was not enough to install Audiogravity as an app, because a phone only grants a site an icon, an offline start and notifications once it genuinely trusts the certificate. And trusting this one was impossible: it identified itself only in a field that Chrome, Safari and Firefox stopped reading in 2017, so it was refused outright, and installing it by hand as trusted did not repair that. Android's Install app prompt, in particular, could never appear. The box now creates its own small certificate authority, and signs the interface's certificate with it. You install that authority once per phone or computer — the installer prints where to get it and the exact steps for iPhone, Android, macOS and Windows, and the manual has a chapter for it — and from then on the box is trusted, warnings stop, and the app installs like any other. Once per device, and it holds for years.
  • [ops] Upgrading a box that is already installed replaces its certificate, and only when that certificate is broken. Every box installed before this release carries the unusable kind, so upgrading swaps it for one signed by the box's new authority — nothing to undo first, and if you had installed that old certificate on a phone it simply becomes obsolete. A certificate you supplied yourself is a different matter: a valid one is detected and left alone, whoever issued it, with no authority created and nothing published; the box only steps in if it has actually expired or carries none of the identification browsers require.
  • [ops] A box that changed address no longer serves a certificate for the old one. The certificate was written at the first install and never looked at again, so moving the box to another address — or a router handing it a new one — left it presenting a name that no longer matched, forever. It is now re-examined before every start and once a day, and reissued when the address changed or the expiry approaches. The authority is never reissued, so your phones and computers keep trusting the box through all of it, with nothing to redo.
  • [ops] The installer's QR code takes a third less room. It was drawn with the widest margin the standard allows, which on a terminal meant a block of 58 columns by 29 lines for a symbol that is only 21 modules across. The margin is now the minimum: same symbol, same precision — the modules are drawn exactly as before, which is what makes it scannable — in 46 by 23.

[0.9.34] - 2026-08-11

Added

  • [ui] A whole class of blank-panel bugs can no longer reach you. Nothing changes on screen today; what changes is what can ship tomorrow. The interface's stylesheets have been checked automatically for years, and its JavaScript had never been checked at all — so a rename that landed on one side only, leaving code reading a name that no longer exists, was found by whoever happened to open the right panel. That is not hypothetical: it happened twice in one afternoon in the audio pipeline diagram, and each time the panel stopped drawing entirely until the page was reloaded, because once that kind of error is thrown every later attempt throws again. The check that catches it takes about a second, refuses the commit that carries the mistake, and runs over every file on every test run so it cannot be walked past. It looks for exactly one thing — a name used and never defined — and deliberately nothing else: an interface is a matter of taste, a name that does not exist is a crash.
  • [ops] The installer ends with a QR code, and tells you how to make it an app. The interface lives at a private address on your own network — it cannot be shortened, guessed or looked up, so until now it had to be typed by hand onto the phone most people actually use it from. The installer now prints a QR code of that exact address at the end, followed by the three steps that turn it into a fullscreen app with its own icon on Android, iPhone and desktop, taken word for word from the manual. The certificate warning is explained where you meet it — right after the code, phone in hand — instead of three screens earlier. The code is only drawn when a person is watching: an upgrade, or the one-click self-update which runs detached with no terminal, prints nothing and installs nothing. On a narrow terminal it switches to a compact rendering, and on a very narrow one it steps aside and leaves the address alone rather than print a symbol that cannot be scanned.

Fixed

  • [core] [ui] A control that does not work no longer pretends it did. Press pause, skip a track, set the volume, toggle repeat or shuffle: the interface flipped immediately, and the player's answer was never read — the command counted as done the moment it was sent. Almost always the two agree; the exceptions were an interface showing a pause that never happened, or a volume the output cannot set because the DAC has no volume control of its own. Every transport command now waits for the player's verdict, and on a refusal the button or slider simply returns to the true state — without a message, deliberately: a control that visibly does not act is the honest answer, and a notification would be noise for something this rare. One case had stayed misleading for good: the volume slider kept showing the level you dragged to until the panel was closed. It now glides back to the real level a moment after you let go — invisible when the change took, honest when it did not. And after every command, refused or not, the player broadcasts its actual state, so whatever any screen was guessing is corrected within the second.
  • [core] Switching outputs can no longer end in silence that calls itself success. The switch enabled the output you chose and disabled the others in whatever order the list happened to come, and never read the answer. If the new output refused to open — in use by another program, cable pulled — the others were sometimes already off: every output disabled, and the screen reporting the switch had worked. The new output is now enabled before anything is disabled, in one atomic exchange, so a refusal leaves the previous output playing; and the refusal is reported with the player's own reason instead of a success that isn't one. The output you picked is also addressed by its identity rather than by its name, which two outputs in one configuration are allowed to share — picking the second of two same-named devices used to switch to the first.
  • [core] [ui] Adding music to the queue says when it did not work. Play an album, queue a track, start a radio station, remove a row from the queue: the answer was always "done", because the player's reply was never read. The cases where that matters are ordinary ones — a station whose address the player rejects, an album that left the library index since the page was drawn, a queue row someone already removed. Each looked like a click that worked and left you facing a screen that disagreed with the music. They now report what actually happened, with the player's own reason, and the queue view says so instead of writing it to a log nobody reads. Two exceptions are deliberate, because failing the whole request would lie in the other direction: when several tracks are queued and the list stops partway, the ones that made it in are kept and counted; and the titles attached to queued streams are decoration, so failing to write them never cancels an enqueue that worked.
  • [core] A library that never got indexed no longer reports that it did. After you point Audiogravity at a new music folder, it asks the player to re-index — and that request is sent in the exact moment the player is restarting, so it can be dropped. The reply was not read: the log said "rescan triggered", the five follow-up attempts were skipped, and the new library simply never appeared. The reply is read now, and a dropped request is retried like any other failure.
  • [core] A player that hangs can no longer pin a button with it. Every exchange with the player now carries a strict deadline of two and a half seconds covering the whole conversation — including one step, the greeting the player sends on connection, that previously had no limit at all: a player wedged mid-restart could hold a tapped button forever. One nuance is handled honestly: a command that times out was already delivered and may still act, so the interface does not assume it failed — it re-syncs on the state the player broadcasts right after.
  • [core] A Tidal track can be seeked the first time you play it. Until now the first play of any Tidal track ignored the progress bar: the track reaches the player as a stream being written while it plays, and the player only learns where a stream's parts are once it has read it through — so it declined every jump until the next play. What it needed already existed: a complete, seekable copy of the track finishes downloading a few seconds in, and was simply never used for the listen in progress. It is now — when you drag the bar, the track is quietly reopened from that copy and the jump lands, which looks exactly like an ordinary seek since a seek always pauses for an instant anyway. Two consequences worth knowing. Dragging the bar in the very first seconds, before the copy is ready, is still declined — and now says so instead of quietly doing nothing, which also means the position you see stays the position you are at. And a seek on a live radio stream is likewise reported as declined rather than silently ignored: a live stream has no end to jump to. Dragging the bar again while a reopen is already under way is declined too, rather than queued: each queued jump would have reopened the track in turn, so a few impatient drags meant hearing the track restart several times over.
  • [core] Tracks queued from Qobuz, Tidal, HIGHRESAUDIO or a UPnP server show their length. They showed --:-- where a track from your own library shows its duration. The reason is that MPD, which plays them, only learns how long a stream is by decoding it — so a track sitting in the queue, waiting its turn, had no length to report. The length was known all along: the service that supplied the track had said so when it was queued, and the value was being dropped a line later. It is now kept, and kept on disk, so it survives restarting the box. Internet radio still shows --:--, which is the honest answer for a live stream. Tracks already in your queue from before this update keep showing --:-- until they are queued again — their length was never recorded, so there is nothing to recover. The player's own remaining-time display draws on the same value, so it can no longer disagree with the queue about the track they are both showing.
  • [ui] HQPlayer's volume reads as a volume again. The DSP panel printed the level as -13.800000190734863 dB. Nothing was wrong with the level itself: HQPlayer keeps it in a smaller number format than the one it is read into, and the trailing digits are the noise of that conversion — a −13.8 that cannot be written exactly on the way across. It now reads -13.8 dB, and the slider beneath it moves in tenths to match, where it used to snap to the nearest half and leave the handle pointing somewhere other than the printed figure.
  • [ui] The library's tabs are named on a phone too, and the bar no longer runs off the screen. The five tabs — Browse, Search, Queue, Library, Radio — carried their names on a desktop and were reduced to bare glyphs on a phone, which is the screen the library is browsed from most: five unlabelled symbols, and nothing to say which was which. The names now sit under the icons everywhere. Making them fit meant taking back the room they needed: the source you are browsing has left the bar for the head of the content it describes — where it reads better anyway, since it says what you are looking at rather than where you can go — and the bar has stopped spending a sixth of a phone's width on its own margins. If it still runs short, a long source name or a narrow screen, the bar can be dragged sideways, and the selected tab is brought back into view whenever it changes.
  • [ui] Tapping the highlighted tab gets you out of where you are. Four screens show a tab selected that is not quite them: the outputs list shows Library, and the artist, Roon and UPnP browsers all show Browse. Tapping that tab — the obvious way back — did nothing at all, because the interface treated it as a tab you were already on. Easy to miss while the names were hidden; a clearly named, clearly selected, completely dead control once they were shown.
  • [ui] The buttons beside the tabs say what they do. Refresh, Outputs, and the way back from the outputs list were bare symbols carrying a tooltip, which a phone never shows. They now give their names like the tabs, set apart by a hairline because they do something rather than take you somewhere, and they are real buttons — reachable with a keyboard, which they were not.
  • [ui] Internet radio has an icon of its own. It was concentric arcs around a dot, drawn from the same vocabulary as the Wi-Fi and cast icons that sit near it, and at the smallest size the three were hard to tell apart. It is a radio set now: an object rather than a wave. The same drawing had also been standing in for the selected state of a radio button in two settings screens — a pun on the word rather than a meaning — and those use a proper filled circle.
  • [ops] The address the installer prints, and the certificate it issues, are the one your phone can actually reach. Both were taken from the first entry of the list of every address the machine holds — a list with no useful order. On a box running Docker, a bridge, a VPN or simply a second network interface, that first entry can be an address nothing outside the box can route to: the installer named it, and the certificate was issued for it. It now asks the routing table which address a packet leaving the machine would carry, which is the question actually being asked, and falls back to the old method on a machine with no default route. Certificates already generated are untouched.
  • [ui] On a phone or a tablet, every field is the size of its own label again. iPhone Safari zooms the page in whenever you tap a field whose text is declared smaller than a certain size, and it never zooms back out — so Audiogravity raised every field on a touch screen to that size. It worked, and it was the wrong price: a form was drawn with its fields visibly larger than the labels above them, on every screen that has one. What that guard actually inspects, it turns out, is the size a field declares — not the size it ends up drawn at. Fields still declare the size Safari wants, and are now drawn at the size the form was designed with. Measured on an iPhone: the page stays put, and the pair reads as one again. On an iPhone older than iOS 16.4 the field simply keeps its former size — bigger than its label, but still never zooming, because the protection never depends on the new part.
  • [ui] Dropdown lists were being protected from something that cannot happen to them. They were held at that same enlarged size from the start, on the assumption that every form control zooms. Nobody had checked. A dropdown opens a picker wheel, not a keyboard, and measured on an iPhone it does not zoom the page at any size — so the rule was buying nothing and costing the lists their proportions. It showed worst on the HQPlayer DSP card, where the Filter, Shaper and Mode lists sit under deliberately small labels and were being drawn a third larger than them. Lists now keep the size their own screen gives them, on touch as on desktop.
  • [site] The editions page listed the Pro library without HIGHRESAUDIO. Two sentences on that page still named only Qobuz and Tidal — the feature list for the Pro library, and the note about needing your own subscription — although HRA has been a first-class source since 0.9.10 and is named throughout the landing page. Anyone comparing the editions before buying, HRA included, would have read that their service was not covered. Both sentences now name all three.
  • [site] The third-party notices claimed more of HIGHRESAUDIO than HRA has actually done. The notice stated that HRA "reviewed the interface built on it". What HRA saw were early screenshots; the application itself has not been in their hands, and their own condition — to test and confirm before a public release — has not yet been met. A notice that credits a partner with a review they have not given is a claim about someone else, published on their behalf, and it is the kind of sentence that costs a working relationship if the partner reads it first. It now says what happened: HRA supplied their API and saw early screenshots of it. The rest of the paragraph is unchanged, and still makes no claim of partnership or endorsement.
  • [ui] [core] In your local library, All, Recent and A–Z now show three different things. They showed the same list. Two reasons: the albums already arrive sorted alphabetically, so A–Z re-did work that was done — and Recent never looked at a date at all. It took the first fifty albums alphabetically and called them recent, which is why a library starting with & and 10,000 Hz Legend opened on those under every heading. Recent now means what it says: most recently added to your library first, whatever the record's release year — a 1959 pressing copied in this week sits at the top. Albums added on the same day keep their alphabetical order, so a library that was indexed in one go stays readable instead of appearing in the order the disc happened to be read. And Recent is no longer capped at fifty: the whole library scrolls in that order.

[0.9.33] - 2026-08-09

Fixed

  • [ui] The sign-in screen can be read, and the button says when it is working. Its grey texts — the line under the logo, the version at the foot, the placeholder in each field, the link back to the password form — were set in a grey that falls below the contrast the text of an interface owes its reader: legible on a studio monitor, gone on a laptop in daylight. They are darker now, and the version line is no longer set at the size reserved for short uppercase labels. The message shown when a password is refused was stated in red on a pale red ground, which is the least legible way to say the one thing on that screen a person has to read; the alert keeps its red border and tinted ground, and the sentence itself is now plain dark text, with the icon beside it carrying the meaning for anyone who cannot see the colour. A failed passkey attempt is shown in that same alert rather than as a line of small red text unlike anything else on the page. And the spinner on the sign-in button, in dark mode, was drawn in white on a white button — the button appeared to do nothing at all while it worked. It now takes the button's own text colour. Three of the colours and sizes involved were asking for values that were defined nowhere in the interface, which is why a stylesheet that looked correct did not behave correctly.

  • [ui] An error inside a list can be read. The panel shown when a list of albums, stations or servers fails to load stated itself in red on a pale red ground, exactly as the sign-in message did and with the same result. It keeps its red border and tinted ground and states itself in plain dark text. It also takes that red from the colour each theme defines for an error, rather than from one theme's value that the others were quietly inheriting.

  • [ui] The faint grey text throughout the interface can be read. Secondary and tertiary text — labels, captions, metadata, placeholders, version lines: a hundred and seventeen places — was set in greys that look considered on a good monitor in a dim room and disappear on a laptop outdoors. They were between 2.08 and 3.5 times lighter than the background where legibility asks for four and a half. Every one of them, in all three themes and in both light and dark, now clears that bar, with the three levels of the hierarchy still distinct from one another. Nothing changes hue; what changes is how far each grey sits from the page behind it.

  • [ui] Coloured text — a status, a licence, an error, a warning — is readable at last. Those colours were doing three jobs with one value: outlining a badge, tinting its fill, and writing the words on top. A red dark enough to read is too dark for the outline it was chosen for, so the words lost. They now have their own reading shade, same hue, while the outlines and fills keep the colour you know: a green status was 2.15 times lighter than its background and is now 5. The whole interface is covered — roughly a hundred and fifty places, badges, statuses, validation messages and licence panels among them.

  • [ui] Two more things that had never followed your theme. A placeholder square behind album art in the queue was painted a fixed light grey, so it stayed light in every dark theme; and the legend of the audio pipeline no longer disagrees with the lines it describes. Both were reading colour names that existed nowhere in the interface, which is a mistake nothing reports — the colour written beside the name simply renders, forever.

  • [ui] Audiogravity no longer flashes white before it appears. The appearance you chose — theme and dark mode — was applied by the interface's own code, which cannot run until half a megabyte of it has been fetched and parsed. Until then every screen was drawn in the light palette, so anyone using the dark theme met a bright white flash on every cold start, on a device people mostly listen to in the dark. The browser's own bar flashed with it, having been given one fixed colour that belonged to a single theme whatever yours was, and the sign-in screen appeared in the default theme before switching to yours a moment later. All three are read and applied before anything is drawn now, so the first thing on screen is already the appearance you picked.

  • [ui] The interface is set in its own typeface, and the box serves it. Two separate things kept that from being true. The typeface was fetched from Google's servers on every load, which is a request to a third party, a visitor's address handed to it, and a wait before the first screen can be drawn — on a device that may sit on a network with no route out at all, it is simply a typeface that never arrives. And the default theme replaced it with whatever font the operating system provides, so on most installations the file was downloaded and then discarded: the interface looked like a Mac on a Mac, a PC on a PC, and never like itself. Inter now ships with Audiogravity and is served by the installation, on every screen including sign-in, with nothing requested from the internet and no font host left in the security policy. The default theme uses it like the others. It is stored alongside the rest of the interface, so it is there offline from the first visit.

  • [ui] Values you read digit by digit are in a typeface made for it. Frequencies, temperatures, identifiers, addresses, log tables and the terminal were set in Courier New — a typewriter face from the 1950s, thin on a modern screen and the one thing on an otherwise current interface that looked its age. They now use JetBrains Mono, served by the installation like the rest. Its figures are the same width, so a value that refreshes in place no longer makes the layout shift under the eye. The reverse was corrected with it: a handful of things that are read rather than transcribed had drifted into that typeface and have come back to the interface one — three status badges, a confirmation sentence, a title and an error label.

  • [ui] A theme can now carry its own typeface. Themes are the part of the interface open to contribution, and a theme that declared a typeface was obeyed everywhere except the sign-in screen, which pinned its own. The rule is now a single one with no exception: a theme states the typeface it wants, and the whole interface follows.

  • [ui] Seven screens no longer need the internet to appear. Configuration, Profiles, Performance, Systemd, Services, Audio software and the dashboard were built on a small library the interface fetched from a public code distributor every time it loaded. A missing typeface would have been a matter of appearance; missing code is not — those pages rendered nothing at all on a machine with no route out, or behind a network that blocks that distributor. The library ships with Audiogravity now, inside a file the interface already downloads, so it costs no request at all. It was in fact already installed and unused, and one screen was loading it that way, so the interface had been carrying two copies of the same library at two different versions.

  • [core] The music keeps priority over the rest of the box. Nothing on the machine told the system that playback matters more than the control interface: when the core did something heavy — installing a package, scanning, or simply misbehaving — it competed on equal footing with the player for processor, memory and disk cache. On an idle box this changed nothing and still changes nothing; the core spends about a tenth of a percent of one processor at rest, measured over three days. But under contention the player now wins two thirds of the processor, and a burst of memory from the core is slowed down rather than allowed to push the player's own data out to disk. A runaway is capped and restarted in seconds instead of dragging the whole machine into swap while you are listening.

  • [ui] An interface that has lost its key no longer hammers the box it cannot reach. A browser whose stored key was gone — cleared storage, a broken installation — kept sending every request anyway: cover art fetches even carried the literal word "null" where the key belonged. Measured on one box: 598 rejected requests in a day, each costing the player's machine a request cycle and a journal line, none able to succeed. Such a client now sends a single request, reads the refusal, and goes quiet: it shows "Not connected" with the reason instead of an interface that looks alive and loads nothing. Installations configured to run without a key are recognised by that same first request and keep working as before. The box's budget is for music; a client with a screen and a processor of its own carries its own failure now.

  • [ui] Small text you actually read is a size larger. Timestamps, log times, IP addresses, error messages, tooltips, track durations and the descriptions under a source were set at the same size as the badges beside them — a size chosen for things you identify at a glance rather than read. Twenty-nine of them move up one step. Badges, format tags and short labels stay where they were: the interface is a dense working surface and that density is deliberate.

  • [ui] The manual's title is bigger than its section headings again. It asked for the size above them and was given the same one, because two steps of the interface's type scale had been created with the same value — so the two rendered identically on a desktop and only came apart on a narrow window. The scale had eight names for seven sizes; it has seven of each now, declared in one place instead of two a hundred lines apart, which is how the duplicate went unnoticed. Nothing else changes size: every screen that used the affected step keeps the size it had.

  • [lic] The licence pages breathe evenly. The space between things had grown to twenty-one different values, a pixel or two apart in places — differences nobody chose and nobody can see individually, but which add up to a page that reads as slightly restless. There are seven now, and where a value had to move it moved outward rather than inward, so nothing is tighter than it was.

  • [lic] The page where you fetch your licence is set at a size meant for reading. It shared its text sizes with the administration console, which is a dense working surface — rows compared against rows, consulted at length. The portal is the opposite: opened once, often from a phone, by someone who has just reinstalled a machine and wants a file. Every size on it moves up one step. Behind that, the twenty-two sizes those pages had grown — eight of them within two pixels of one another, differences no eye resolves — are now six, and every one of them is a deliberate step.

  • [lic] Tapping a field on the licence portal no longer zooms the page in. iPhone Safari zooms in on any field whose text is smaller than a certain size, and does not zoom back out — the page stays magnified and the button you were reaching for is off-screen. The stylesheet already guarded against this, but six fields carried their size written on the element itself, which overrides any stylesheet: the guard was there and had no effect. Among them the device-id field on the portal, which is the first thing a customer touches.

  • [lic] The campaign preview showed the unsubscribe footer at the wrong size. The preview is drawn in a frame that does not load the interface's stylesheet, so the footer's size was dropped there and shown at the browser's default, while the message actually sent to recipients used the intended smaller text. A preview is consulted precisely to decide whether to send; the two now match, and are checked against each other.

  • [lic] The page where you re-download your licence no longer depends on anyone else. It loaded the code it is built from — and its typeface — from a public code distributor on the internet. That worked until it did not: an outage there, a company network that blocks it, and the page rendered nothing at all. Blank, with no explanation, at the moment someone who has reinstalled their machine comes to fetch their licence file. Both now come from the licence server itself, which is the only machine that has to be reachable for that page to work. The same applies to the administration console, where scripts fetched from a third party ran with full access to a page holding customer records.

  • [lic] The licence pages can be read in daylight, and in the light theme. Their light theme had never been given colours of its own, so it wore the ones drawn for a black page: greens, ambers and reds that measured as little as 1.4 times lighter than white where legibility asks for four and a half. A status, an expiry warning, a refusal — the things on those pages that exist to be read — were the least readable text on them. Every colour now has a value for each theme and a separate shade for when it is text rather than a border or a fill. Several were asking for colours defined nowhere at all, which is a fault nothing reports: a badge marking a customised template was drawing black on black and simply could not be seen, and the links to release notes, to support and to an announcement's address had lost any sign of being links.

  • [lic] The licence pages are set in the same typefaces as the rest of Audiogravity. They used whatever the system provided, so the customer portal looked like a different product on every machine that opened it. They now use Inter, and JetBrains Mono for the values read character by character — licence keys, fingerprints, order numbers — both served by the licence server rather than fetched from a font host.

  • [ui] The service state and its button are the same size again. On the network stability test, the green ACTIVE marker sat beside the STOP SERVICE button and was visibly shorter and narrower — a fixed height where the button derives its own from its padding, a smaller type size, and a monospaced family where the rest of the interface's uppercase labels use the interface font. The two now share a box, so they read as one control rather than two that do not line up.

  • [ui] The red "Remove" backdrop no longer frames a row nobody is removing. On an iPhone, swiping a list left drew a red frame around the station under the finger, and it stayed there — including while a station was being chosen. The backdrop that a swipe uncovers was painted under every row at all times, and a row that had been touched once kept the displacement it was given, which on iOS is enough for a sliver of that backdrop to show around its edges for the rest of the session. The backdrop now exists only while a row is genuinely being dragged left, and a row returns to being an ordinary row as soon as the finger lifts. Four lists share this gesture — radio stations, UPnP servers, renderers and the playback queue — so all four are fixed, and three neighbouring faults with them: dragging a row and then merely tapping another one used to strand the first, a rightward drag or the drift of a vertical scroll uncovered the backdrop under a row that had not moved, and a removed row handed its red backdrop to whichever row took its place.

  • [ui] The licence panel says the same thing once. A trial announced the days left three times over — the badge, a sentence and the progress bar — one of them assembled from a template, so it read "27 day(s) remaining". The price was stated twice on the way to paying. The badge and the bar keep the count, the offer keeps the price, and the panel is two lines shorter. A box whose trial has ended is still shown the price, which is where the previous release's fix had left a gap: it removed the repeated price without noticing that the wording shown after a trial ends carries no figure of its own, and neither does the "About Licensing" help.

  • [ui] The purchase sentence in the licence panel is a whole sentence again. The price comes from the licence server, so it is empty until the first fetch lands — and stays empty where none is configured. It was dropped into the middle of the sentence regardless, which read "one-time payment of , no subscription": a stray comma, on the panel where the customer is being asked to buy. The clause is now left out when there is no price. The tests covering that sentence were checking a copy of the logic rather than the panel itself, so they could not have seen this; they now render the panel.

  • [ui] The System tab shows the core's logs again. The log panel asked the system journal for a service named audiogravity-backend, which was renamed some time ago. A journal asked for a unit that does not exist does not complain — it answers with an empty log — so the panel showed nothing at all, with no sign that anything was wrong. It now asks for the service that is actually running.

  • [ui] The interface calls the core "the core", as the rest of Audiogravity does. The System tab's action was labelled "Restart Backend", and the confirmation, the notification that followed it, the log panel's title and the connection indicator all used the same word — one component for which the documentation, the site and the manual have said "core" throughout. The indicator's tooltip was also the only text in the interface still written in French, and the development-mode prompt for an API key pointed at a backend/.env that does not exist. Nothing changes in behaviour; the addresses the interface calls are untouched.

[0.9.32] - 2026-08-06

Added

  • [site] The user manual has somewhere to be read. Every link to it pointed at GitHub, so anyone following "Where is the documentation?" landed on a code host — and the search traffic for our own manual went there too. The twelve chapters are now published as ordinary pages on audiogravity.app, with a chapter sidebar, working cross-references and the same look and theme as the rest of the site. They are rendered from the same Markdown the interface reads, and regenerated automatically whenever a chapter changes, so the copy on the site and the copy in the app can no longer disagree. Reading it is also now reachable from the top of the page, not only from the sentences that happened to link to it. Every manual page carries the app mark beside the wordmark at the top, and that is the way back to the site — alongside the link already at the foot of each page. Its contents page states which release it describes, taken from the version the release process already writes, so it follows a release without anyone remembering to update it. That is said once, at the entrance: a version on every chapter would invite "where is the manual for my version?", which has no answer while only one is published.
  • [site] Who is selling this. A short note beside the prices says who wrote Audiogravity and on what system, in the first person. Asking for a lifetime licence while saying nothing about who is behind it is a fair thing for a visitor to hesitate over.

Fixed

  • [core + ui + lic] A licence bought for a fixed period no longer lies to its owner, twice. These licences had never been issued, and the whole path was wrong end to end. While one was running the interface announced it as "Lifetime license active", with the end date shown nowhere — so a customer who bought a year believed he had bought the software outright. Then, on the day it ended, the same interface told him his licence file was "invalid or bound to a different device" — read as corrupted or stolen, and worth a support ticket every time. His licence panel now states the end date while the licence runs, and on the day it ends says so plainly: the licence ended on that date, and Audiogravity continues in Starter Edition. His order number stays on screen, and the panel offers renewal rather than the sales pitch shown to someone who never bought anything.
  • [lic] A mistyped end date is refused when it is typed, not months later on a customer's box. The date could be entered as free text on an order, and nothing checked it. A slip such as 31/12/2026 was stored as written: the administration went on showing the order as valid, while the customer's box — which refuses anything it cannot read — reported his licence as tampered with. Nothing connected the two, and the customer was the one who found out. The field is now refused on entry, an order already holding an unreadable date is shown as such in the administration instead of appearing healthy, and the licence server refuses it rather than ignoring it.
  • [core + lic] A licence now ends at the same moment for everyone. Each side worked out "today" in its own timezone, so a box in New Zealand and the server in Europe disagreed for about thirteen hours a day on whether a licence had run out. Both now settle it in UTC.
  • [ui] A licence bought for a period is no longer labelled a trial. The licence server marks every dated licence trial internally, including one sold for a year, and three separate screens relayed that word to the customer — the licence panel, the key check and the activation certificate, the last of these right after payment.
  • [site] The menu is usable on a phone again — in portrait it was not there at all. Below a certain width the section links were simply hidden, so on any phone held upright there was no way to reach a section but to scroll the whole page: eight entries, none of them reachable. Held sideways they did appear, but squeezed into a bar too narrow for them, with two of the labels breaking onto a second line and the first one touching the logo. The links now take a row of their own beneath the name and scroll sideways when there are more than fit — the same answer the licence server's own header uses. The threshold was set from where the single row genuinely stops working, not from where a phone stops, which is why landscape is fixed too.
  • [site] The top bar's links no longer sit on top of the buttons beside them. Between roughly 680 and 1000 pixels — a tablet, a small laptop — the section links overflowed their own row and were drawn over the theme switch and the releases button. Nothing looked broken to a page-width check, because the page was never wider: the elements simply overlapped. The releases button now steps aside on those widths, as its link is in the footer anyway, and the links sit closer together.
  • [site] Wording of the comparison table and the purchase steps. Four rows carried parenthetical asides that were the tallest thing in a phone-width column; they say the same thing without them — the streaming row keeps a short "(official APIs)", which is the part worth stating and the only one that fits. The first purchase step no longer names the payment provider in its title, and the activation step and the platform note both say "device" where one said "machine".
  • [site] The page calls the third-party engines "engines" throughout. Roon, MPD, HQPlayer and AirPlay were described as "players" in six places — the opening paragraph, the shared description search engines and social previews use, the footer line, a feature heading and a comparison row — while the section that introduces them is titled "Every audiophile engine" and the body text beneath that heading already said "engine" twice. The word "player" now means one thing on this page: the listening screen.
  • [site] What Audiogravity's relationship with HIGHRESAUDIO is, and is not. The third-party notices grouped HRA with Qobuz and TIDAL under a blanket "not affiliated with, endorsed by, or sponsored by", which stopped being accurate once HRA supplied their API for this integration and reviewed the interface built on it. HRA is now stated separately and precisely: a working arrangement, not a commercial one, with no claim of partnership or endorsement made anywhere.
  • [ui] The user manual fits a phone screen. Opening an illustrated chapter shifted the reader sideways and left it there: the screenshots are authored at a fixed width of 360 pixels, and nothing in the reader capped them, so on any screen narrower than that plus the reader's own margins the page had to be dragged back. An iPhone 13 offers just enough and hid the problem; every narrower model does not. Two other things pushed the same way and are fixed with it — a table wider than the screen took the whole reader with it instead of scrolling on its own, and a long command written inline in the text (an installation flag with a URL) could not be broken, so it set a width the reader could not go under. Tables now scroll by themselves and keep their structure for a screen reader. Checked on the twelve chapters across four screen sizes, from the smallest iPhone up.
  • [site] The landing page fits a phone. It had to be dragged sideways on every iPhone: the page wanted 470 pixels and the narrowest current model offers 375. Two things caused it — the top bar kept a GitHub button beside the logo long after the menu links had been hidden, and the "Audio priority" table kept a fixed label column that left the sentence beside it under a hundred pixels to wrap in. The bar now drops the button with the links, and that table stacks its labels above their lines on a narrow screen. Measured on six screen sizes in both themes, portrait and landscape.
  • [site] The "Coming soon" ribbon no longer sits on the announcement. Pinned to the top-right corner and above everything else, it landed squarely on the announcement text once that text wrapped onto three lines, which is what it does on a phone.
  • [site] The page fits around the notch and the home indicator. Added to a home screen it runs full-screen, where the sticky bar would have sat beneath the clock and, held sideways, the first characters of every line would have been cut by the rounded corner. The full-height opening section also now follows the screen rather than counting the strip the address bar sits over.
  • [site] Scrollbars and controls follow the theme. The browser draws its own widgets in the light scheme unless told otherwise, whatever the page colours say.
  • [ui + lic + site] The four installable apps no longer identify themselves under a retired domain. The interface and the licence administration announced themselves to the browser as net.di-marco.*, a domain being wound down, and the landing and the customer portal announced no identity at all — which leaves the browser deriving one from the address, so moving a page would have split one app into two. All four now use app.audiogravity. Note for anyone who already pinned one to a home screen during the beta: that identity is what the browser recognises an installed app by, so an existing icon keeps the old one and stops following the manifest — remove it and add it again. Five installs are concerned; doing this later, with a live fleet, would not have been an option.
  • [site] The landing's icons are drawn, not typed. Fifty-eight of them — every tick in the comparison table and in the plan lists — were the ordinary text character ✓, so each was drawn by whatever font the visitor's own machine offered: a different shape, weight and width on macOS, iOS, Windows and Android, right down the column where the product is compared with its alternatives. They are now drawn shapes on the same grid and in the same weight as the rest of the page, and the thirty-odd hand-built shapes already on the page were replaced by their equivalents from the set the app itself uses, so the landing and the interface no longer draw the same idea two different ways. The GitHub mark has no equivalent in that set and keeps its own logo. Each repeated shape is now written once and referenced where it is used, rather than copied into every cell — the page is 5.8% smaller, and a correction to a shape is a correction in one place.
  • [site] The page fits the smallest iPhone. The title is a single word that cannot be broken, and at its set size it alone was wider than a 320-pixel screen leaves — one word holding the whole page open, which is why that screen still had to be dragged sideways after the rest was fixed. It is set smaller there, and only there.
  • [site] The comparison table shows when it continues past the screen. iOS draws no scrollbar, so a table cut at the right edge reads as broken rather than scrollable. A soft fade now marks that edge, and disappears once the last column is reached — at that point nothing continues, and a cue still showing would only be painting over the figures.
  • [site] The figures under the opening section no longer show dividers with nothing beside them. The four figures — trial length, price, machine count, places left — are separated by vertical rules, drawn one per figure. As soon as they wrapped onto several lines, whichever one ended a line kept its rule and pointed at empty space. That happened on every phone width, and again on a desktop window around 1100 pixels where the section splits in two. They are now laid out in a grid, so a rule is drawn only where there is genuinely something on both sides of it.

[0.9.31] - 2026-08-04

Added

  • [lic] The licence administration can be installed on a phone. Added to a home screen it opens as its own app: it starts on its own mark over a black field rather than the browser's blank white page, keeps its own copy of the pages so a launch does not wait on the network, and answers a lost connection with a readable page instead of the browser's error. Orders, licences and installations are never kept — those are read live every time, because a cached copy would show a state the server no longer holds. Signing in now lasts twelve hours rather than until the app is closed: iOS ends a backgrounded app freely, which meant re-entering the token at every launch. The token is erased automatically once the twelve hours are up.
  • [lic] The installations table sorts by any column. Click a heading to order by device, status, version, address or dates. Versions and addresses compare part by part rather than as text, so v0.9.9 comes before v0.9.30 and 192.168.1.9 before 192.168.1.254 — the order you actually want when looking for the box left behind on an old build. Devices missing a value stay at the bottom either way.
  • [lic] The trial start date the server holds for a device is now visible, and can be forgotten. That date decides how long a box's free trial runs, and until now nothing showed it: a support question about a trial could not be answered, and a date recorded from a device whose clock was wrong could never be corrected. It appears in the installations table, and an administrator can clear it — an audited action. Clearing it repairs a wrong date; it does not hand out a new trial, which is what a licence with an expiry date is for.
  • [lic + ops] The licence server reports its own version. It carried a hardcoded 1.0.0 that had never moved and contradicted the releases it was cut from, so establishing which build was live meant probing the API's behaviour. Its version now comes from the same single source as the app and the interface, its health endpoint returns it, and the admin header shows it next to the title — the answer is now a glance, or a single request. The version badges of the licence-server and interface READMEs are propagated along with it, instead of being edited by hand every release.
  • [lic] The installations table can be searched, filtered and paged. A single field matches any part of a device, address, operating system, architecture or version, so finding one box no longer means reading the whole list; a second control narrows to one licence status. Only fifty rows are drawn at a time, and the count above the table states how many match out of how many were loaded — and says so plainly when the server's ceiling is reached, because a table quietly showing a subset reads exactly like a complete one. Sorting and paging apply to everything that matches, not to the rows on screen. The request behind it is now bounded server-side instead of accepting any size asked of it.
  • [lic] The admin and the customer portal now carry the product's own lockup and typeface. Both show the app icon and wordmark at the same figures as the landing page, in the same typeface, so the three no longer look like three different products. The administration top bar also gained direct links to each section, which scroll to it without hunting down a long page.

Fixed

  • [core + lic] Licences with an expiry date are accepted again — they never were. The licence server writes the expiry under one name and the app read another, so a time-limited licence could not be validated: the app found no date, assumed the licence had lapsed, and — since the file was there but refused — reported it as tampered with. Anyone given an evaluation licence would have been told their own licence was forged. The app now reads the right field and lapses a licence on the same day the server does. No customer was affected: none had been issued. A test on each side now locks the field names, so a rename breaks a build instead of an install.
  • [lic] Rate limits now count each caller, not the whole fleet at once. Every limit on the licence server was written per caller and behaved as one shared allowance, because behind the reverse proxy every request arrives carrying the proxy's own address. "Thirty a minute" meant thirty for all installations together rather than thirty each — invisible today with a small fleet, and a throttle waiting to happen as it grows. The real caller is now read from the header the proxy sets, and only when the request genuinely came through it.
  • [lic] The key-validation endpoint is rate-limited. It was the only route on the licence server without a limit, which left the one place that answers "does this key exist?" open to being asked without end. The key space makes guessing hopeless, but that is an argument for the key space, not for leaving the door open.
  • [lic] A recorded trial start can no longer be pushed further back. The server remembers when a device began its trial, so deleting the local files no longer buys a fresh one — that part is unchanged. What changed is that only the first report counts: the earlier rule took the smaller of the two every time, which meant anyone able to reach that endpoint with a device's fingerprint could report a start in 1970 and end that device's trial permanently. A wrong first date is corrected from the admin, as before, and the action is recorded.
  • [lic] The licence administration no longer holds a connection open for nothing. The page carried a development convenience — a link kept open so it reloads itself whenever a file is edited — and nothing stopped it from running on the deployed server, where no file is ever edited. Every open tab kept a connection to the server for a minute, was cut, and reopened it two seconds later, since the day the interface was written. Nothing broke and nothing showed, which is why it went unnoticed until the server's own error log was read. The page now asks only where development actually happens, and the server no longer offers the route at all unless it is told to.
  • [lic] The icons are drawn, not typed. Ticks, crosses, arrows, the refresh and copy buttons, the theme switch and two emoji were ordinary text characters. Twelve of the fifteen do not exist in the typeface the pages load, so each was drawn by whatever font the visitor's own machine offered — a different shape, weight and width on macOS, iOS, Windows and Android — and the copy button used a character rare enough that a system may show the missing-character box in its place. Every icon now comes from the same drawn set the app itself uses, on one grid, in one weight; a check keeps a stray character or a hand-drawn shape from creeping back in.
  • [lic] The licence server's pages fit a phone screen. The administration bar was held on a single line that needed 552 pixels — wider than every iPhone in portrait — so the theme and sign-out buttons sat off the right edge and the page had to be dragged sideways. Worse, a page wider than the screen carries anything anchored to it out as well: the order detail panel, whose own layout was correct, opened partly off-screen. The bar now wraps, the section links take a row of their own instead of being squeezed to nothing, and the whole interface was re-measured on five screen sizes from the smallest iPhone to a tablet.
  • [lic] Dates, numbers, dropdowns and search boxes follow the theme. The stylesheet named the three kinds of field it covered, so the nine other kinds in use kept the browser's own styling — white rectangles in the middle of a black page, as anyone who set an expiry date could see. Fields are now selected by what they are not, so a kind added later is styled by default rather than forgotten, and the browser is told which scheme to draw its own date pickers and dropdown lists in.
  • [lic] Tapping a field on a phone no longer zooms the page in. Every field was set below the size at which iOS enlarges the page on focus — and it never zooms back out, so one tap left the interface stuck at the wrong scale. Fields are raised on touch screens only; the density the interface was designed at is unchanged on a desktop.
  • [lic] The interface fits around the notch and the home indicator. Both pages ask to run full-screen when pinned to a home screen, but neither reserved the room iOS needs at the edges, so the sticky bar would have run under the status bar. The page height also now follows the screen rather than counting the strip the address bar sits over, which is what left content unreachable at the bottom.
  • [ops] The customer licence portal is reachable again — from the app, it never was. The "Download .lic" and "Upgrade v1 → v2" buttons open a page hosted on the licence server, but the proxy in front of it only ever exposed the API calls that page makes, never the page itself: both buttons landed on a redirect to an unrelated site. Anyone reinstalling a box had no way to fetch their licence back. The page and the three files it needs are now served; everything else, including the whole administration interface, stays behind its client certificate.
  • [ops] Links that carry values no longer arrive empty. The proxy passed on the address but dropped everything after the ?, so every unsubscribe link sent in a campaign email and every v1 → v2 eligibility check reached the server stripped of the order number and token that identify the request — and were answered "missing". Both routes were measured with and without values before and after the fix.
  • [lic] Campaign emails can no longer go out with dead unsubscribe links. The address those links are built on had to be typed by hand, the example beside the field still named the licence server's former domain, and nothing checked what was entered — a wrong value only showed itself once the mail had left, in other people's mailboxes. A fresh install now starts with the right address, and an unusable one is refused wherever it is saved, not only when a campaign is sent. An address already configured is never overwritten.

[0.9.30] - 2026-08-03

Fixed

  • [core] The free trial is kept in a single place, ending the licence-state noise on upgraded boxes. The trial start used to be mirrored to a second, system-level file that no installer had ever created — so it never worked on a deployed box, and where a leftover copy did exist the two disagreed and the core logged a licence error every few seconds. The trial now lives in one file, and the licence server remains the authority that restores a box's true start date. The obsolete file is removed on upgrade.
  • [core] The test suite can no longer write outside its sandbox. Two paths escaped it: the obsolete trial file, and the data directory holding the key that encrypts stored service passwords — on a machine where that directory was writable, running the suite could have overwritten a live installation's key. Both are sandboxed, and any future escape now fails the build instead of silently touching the system.

[0.9.29] - 2026-08-03

Added

  • [core + lic] The free-trial period is anchored to a device's genuine first-use date. The licence server records, per device, the earliest trial start it has seen and returns it signed on each check-in; the box adopts it. The trial length therefore reflects real first use and stays consistent across a reinstall, rather than restarting from local state alone. Tamper-proof: the anchor is signed and bound to the device, so a box can neither forge it nor move it forward.

Fixed

  • [core + lic] Trial and expired boxes now receive announcements and update notifications. These fleet broadcasts previously reached only a box with a currently-valid licence — a box on its free trial (which never contacts the verify endpoint) and an expired one (whose verification stops short) were silently cut off and could sit on an obsolete version with no notice. They now ride the public config channel every box already polls, so every box is notified. The update offer is bounded to the box's own major (derived server-side from the reported version), so a box is never pushed across a major — that stays a deliberate paid upgrade.
  • [ui] The announcement banner shows proper icons instead of emoji. Version/promo/alert/info announcements, the dismiss button and the "Learn more" link now render Lucide SVG icons, coloured per type, consistent with the rest of the interface.
  • [core] A trial box whose two anti-tamper files disagree on the start date self-heals instead of locking up. A legitimate desync (OS reinstall, config restore, a permission change) used to wedge the box in a permanent "tampered" state and log an error on every check; it now reconciles to the earliest valid start — tamper-safe, since the earliest can never lengthen a trial — and repairs both files. A start that genuinely predates the OS install is still rejected.

Changed

  • [core] The public licence config (PayPal URL, prices, contact) is served from the 24 h cache instead of re-querying the licence server every time the licence panel opens — this config is near-static, so the per-open round-trip was wasteful.

[0.9.28] - 2026-08-03

Added

  • [core + lic + ui] The free-trial length can be changed for the whole fleet without shipping an update. The 30-day trial is now a floor the licence server can lengthen: set a number in the admin, the server signs it, and every box adopts it at its next check-in — a promo or a longer beta becomes a setting, not a release. It cannot be abused locally: the value must carry the server's signature and can only extend beyond the built-in floor, never shorten below it, so an unreachable server or an edited cache falls back to the built-in length. The licence panel now counts down against the real total (N / total days).

[0.9.27] - 2026-08-02

Security

  • [lic] The licence server's privileged routes fail closed when their secret is missing. The endpoints only Audiogravity's own backend should reach — recording a box's last-seen, releasing a device binding, checking a key, verifying a licence — are guarded by a shared secret. If that secret was left unconfigured the guard did nothing and the routes stood open to anyone who could reach the server, which is internet-facing. They now refuse every call when the secret is unset, and the secret is compared in constant time. The installer now requires the key rather than offering to skip it, and warns on upgrade if an existing server was left without one.
  • [lic] The box heartbeat endpoint is now rate-limited per client. The shared secret that guards it filters random traffic but is present on every box, so it cannot be treated as private forever. The heartbeat is now capped at 30 calls a minute per source — plenty for a real box, which checks in about once a day — so the endpoint cannot be used to flood the device table.

Changed

  • [core + ops] The licence server moved to its own domain, lic.audiogravity.app. New installs point there, and an upgrade rewrites an existing box's endpoint (host only — same port, same behaviour). The previous address stays live in parallel, so a box that has not upgraded keeps working until it does; nothing is required of the user.

Fixed

  • [core] A box no longer reports its licence as invalid when the licence server is merely unreachable. The online check only ever receives its verdict — valid, revoked, expired, not-found — as a normal reply. A rejected shared secret, a rate limit or a server error would arrive differently, and the box mistook that for a definitive "invalid" instead of "could not reach the server", so a transient hiccup could grey out a perfectly good licence. Those cases now read as unreachable and the last known state is kept.
  • [lic] A licence-email campaign no longer loses recipients on a passing provider hiccup, nor double-sends after a crash. A send that failed once — a rate limit, a timeout — was marked failed for good and never retried, while the progress showed nothing missing. Each recipient is now retried on a later day, up to three attempts, before being given up on. And because an attempt is now stamped before the send rather than after the whole batch, a restart midway no longer leaves rows looking un-sent and re-sending them the same day. An upgrade also no longer briefly doubles the day's send allowance.

[0.9.26] - 2026-08-01

Fixed

  • [ui] The window that asks for your password no longer pushes its Confirm button off the screen. On an iPhone, tapping the field to type zoomed the whole page in — Safari does that to any field whose text is smaller than 16px — and the dialog's right-hand side, Confirm included, ended up outside the visible area. The update simply could not be validated. Audiogravity has carried a rule since 0.9.20 that lifts every field to 16px on a phone precisely to prevent this, but that one field had its size written directly on the element, where no rule can reach it.

  • [ui] Text fields no longer zoom the page on a tablet, or on a phone held sideways. The protection above was tied to narrow screens rather than to touch ones, so it stopped applying past 768 pixels — which is every iPad, and every iPhone in landscape. It now follows the pointer: a finger gets it, a mouse does not.

  • [ui] The pencil that edits a radio station can now be aimed at. Two things made it nearly unclickable. It was the smallest of the three icons on the row and, in compact mode, its target was 24 pixels across; it is now 36 by 44, and all three glyphs are the same size. And the swipe-to-remove gesture armed after 8 pixels of sideways travel anywhere on the row, including on top of the buttons — so a tap that drifted slightly started removing the station instead of opening the editor.

  • [ui] Pressing a button on a list row no longer drags the row away. The fix above applies to every list that supports swipe-to-remove: radio stations, the playback queue, UPnP servers and renderers. A gesture that starts on a button now belongs to that button.

  • [core] Radio search no longer hammers the catalogue when it is unwell. Audiogravity believed it had five mirror servers to fall back on. It has one: three of those names no longer exist, and the other two are aliases of the same machine — which the catalogue itself confirms, advertising a single server. So a failed search meant four requests fired at the very machine that had just refused, inside a third of a second. It now makes one, and asks again only when the first got no answer at all — a timeout or a broken connection, never an explicit refusal.

  • [core] Audiogravity now goes quiet when the catalogue asks it to. If the catalogue answers that it is being asked too often, the box stops sending anything for as long as it requested, instead of carrying on.

  • [core] The catalogue's servers are now found the way its documentation prescribes. Audiogravity asked an address on the catalogue's own server which servers were available — so when that machine went down, the mechanism meant to route around it went down with it. It now resolves them by DNS, which is served by separate infrastructure and answers even when the catalogue does not.

  • [core] A radio search you have already run keeps working while the catalogue is down. Rather than an error, Audiogravity shows you the result it last obtained for that exact search, for up to a day — and those stations still play. They were only listed at first: starting one asked the catalogue again, which was still down, and the answer was "station not found" for a station right there on screen.

  • [core] A station that cannot be reached is no longer reported as missing. Playing a station while the catalogue was unreachable said it did not exist, which sent people looking for a station that was never gone. Audiogravity now says the catalogue is unavailable, and gives the reason. "Not found" now means what it says: the catalogue answered, and it does not know that station.

  • [core + ui] "Search failed" now says what actually failed. Every radio search failure showed the same three words, whether the fault was on your box or at the provider — so an outage two thousand kilometres away sent people looking through their own settings. The message now distinguishes a catalogue that is down from one asking to be queried less often, and says when to try again.

  • [ui] The key your box publishes is now the one that wins. The interface preferred a copy kept in the browser over the key the installer publishes, so a wrong value stored once outranked it for good: every upgrade republished the right key and lost, and the interface stayed locked out with nothing on screen to explain why. The published key is now authoritative, and a leftover copy is discarded.

  • [ui] Typing in radio search no longer sends a query for almost every letter. The pause before a search left the box was shorter than the gap between two letters typed with a thumb: one session sent sixteen queries for a handful of words.

  • [ui] Leaving radio search no longer drops its results into the tab you switched to. A search still on its way arrived after you had moved on and replaced what that tab was showing, so hits for what you had typed appeared under My Live Radio or Favorites. They could then be swiped away, which asked the box to remove a station that list never held. Leaving the Search tab now abandons the search with it, and the longer pause introduced above — which widened the window this could happen in — is safe to keep.

  • [ui] Saving a radio station while the catalogue is down now says so. Starring a station, or adding one to My Live Radio, showed "Could not update" whatever the reason — the same wording for a provider outage as for a fault on your box. Both now report what the catalogue said, and when to try again, exactly as search already did.

  • [ui] A station that will not start no longer blames a part of the box that is working. The message named the local player as the thing to go and check, when the cause is just as often the catalogue being unreachable, or the chosen output refusing the stream. It now names both possible causes — the catalogue, or the chosen output refusing the stream — instead of pointing at one that is working. (Telling the two apart, core-side, is a follow-up.)

  • [core] A network speaker you added by its address is now found even when multicast does not work. Audiogravity looks for speakers by shouting on the network; where that shout cannot travel — some routers, some subnets — you can give the address directly instead. That direct probe was skipped whenever the shout itself failed, which is exactly the situation it exists for. The two are now independent, and a search that cannot send anything at all stops immediately rather than listening to silence for its full duration.

  • [core] Audiogravity now stops sending play reports when the radio catalogue asks it to. A refusal on the report channel was noted and ignored, so every new station played sent another report to a server that had explicitly asked for quiet. It now holds off for as long as it was asked — and only the reports stop: searches, which you are waiting on, still go through.

  • [core] A box with no multicast route no longer writes a warning a minute. Looking up cover art from the local player retries every minute during playback, and the network search it uses had been made to log at warning level along with the ones you trigger yourself.

Security

  • [ops] A failed self-update can no longer be turned into a way to gain root on the box. The privileged helper that performs an update ran with administrator rights and took the address it downloaded-and-ran, and the file path it wrote, from whoever called it. Anything already running under Audiogravity's own limited account could therefore hand it a malicious address and have it executed as administrator — a local privilege escalation, not reachable from the network, but real. The helper now fixes those values itself and accepts only a checked list of the rest; a supplied address or path is ignored. Proven on real hardware: an injected address is refused and never runs.

Fixed

  • [ops] Undoing a failed self-update now puts the whole box back, not just the program. An update reinstalls far more than the core binary — the service definition, the update mechanism itself, the package registry, the privileged helpers — but undoing one only restored the binary, then reported success. A box could be left running parts of a version that had just failed, including an update mechanism that could no longer update it, while its log said everything was fine. The undo now restores everything the update overwrote, reloads the service definition, and says so honestly — if any part could not be put back, it reports that instead of a clean success. Verified on a real box: nine files an update overwrites are all restored, and the box can update again afterwards.

Added

  • [core] Audiogravity now tells the radio catalogue which stations you play, as its documentation asks. That count is what ranks stations by popularity — the ranking Audiogravity sorts your own search results by, and which it had until now been consuming without ever contributing to. Only stations from the catalogue are ever reported, never the ones you add by hand, and nothing is sent when playback fails to start. This sends the station's identifier to a third party: set RADIO_REPORT_PLAYS=false in the configuration and no play ever leaves the box.

  • [core] Audiogravity now says who it is when it talks to your network speakers. Requests to a UPnP renderer — reading its description, subscribing to its events, sending it transport commands — went out with no identification, so a speaker's log showed an anonymous Python client. Every outbound request now carries the same Audiogravity signature the rest of the box already used.

  • [ops] The new radio setting reaches boxes that are already installed. Upgrading adds it to the existing configuration rather than leaving the machine on a default it never agreed to.

  • [ops] Uninstalling no longer leaves a service running on the machine. Performance tuning installs a small service that re-applies the CPU governor at every boot; the uninstaller never removed it. A machine you had removed Audiogravity from therefore kept having its processor scheduling changed at each start, by a service with nothing left on the machine to explain where it came from. It and its helper script are now removed with everything else.

  • [ops] --purge now really removes everything. It left behind the backups of your audio services' configuration files in /var/backups/audiogravity — copies that can contain your MPD password or Shairport authentication keys — along with the CPU governor setting and the box's state directory. All three now go with the rest.

Removed

  • [ui] The API key is gone from the Settings panel. It was offered as a field, with a reveal button, to every logged-in user. It protected nothing — the key is already inside the page your browser loads, because the installer puts it there — and it was the only thing on a running box that could store a wrong key and cut that browser off from your music. It has no replacement: nothing needs to be typed there any more.

  • [core] The SSDP network search has one implementation instead of three. Finding a network speaker, finding a media server and finding the local renderer for cover art each carried their own copy of the same search — so a fix or a hardening applied to one reached neither of the others. They now share one, each keeping only what genuinely differs: what it looks for, and whether it also probes hosts directly.

Documentation

  • [site] The manual explains where radio stations come from, and what is sent back. A new section says that search depends on an outside catalogue Audiogravity does not run, that your own stations keep playing when it is down, what a play report contains — the station's identifier, nothing else — and how to switch it off.
  • [site] The Radio Browser notice describes what Audiogravity actually does. It claimed the box auto-discovers the nearest mirror through an HTTP endpoint; discovery is now by DNS, there is no nearest mirror to pick from, and the play reporting was undocumented.
  • [site] The manual no longer documents the API key field removed above. Its screenshot of the Settings panel is retaken as well, so the picture matches what the panel now shows.
  • [site] The third-party notices now list every outside service the box calls. Only Radio Browser was named, while cover art also queries MusicBrainz (release data, CC0), the Cover Art Archive (a cooperation between the Internet Archive and MusicBrainz, which asserts no licence over the images themselves) and the iTunes Search API as a last resort. Each entry states what it is used for, what its provider requires of a client, and what Audiogravity sends.
  • [site] The manual now says what uninstalling actually leaves behind. It gave two commands and nothing else, so following it to the letter left you believing the machine was clean while your whole configuration was still there. It now states what each form removes, documents --purge, and names what is deliberately kept: the audio software installed from the Audio Software page — MPD, Shairport-Sync, upmpdcli, HQPlayer NAA, Roon Bridge — which are ordinary system packages that keep working on their own.

[0.9.25] - 2026-07-28

Fixed

  • [core + ui] The window that shows an installation no longer stays stuck on "Installing…" after the work is done. Its log lines only ever reached the browser as they happened, and nothing was kept for a browser that missed them — a reconnection, a phone that put the page to sleep, a second of bad network. The window then froze on the last line it had received and stayed there, while the installation carried on and finished without it. Measured on a real box: the package was installed in six seconds, and the panel was still announcing work in progress six minutes later. Audiogravity now keeps the log of the operation in progress and the interface asks for what it missed — when it reconnects, and each time the operation moves on. Nothing is lost any more, and nothing is shown twice.
  • [core] An installation no longer stays silent for as long as it lasts. Everything the package manager printed was collected and only written to the log once the command had finished, so between "Installing…" and the end there was nothing at all — a healthy installation looked exactly like a dead one, which is what made the previous problem impossible to tell apart from a real failure. Lines now appear as they are produced.
  • [core] A successful removal no longer reports an error. Removing AirPlay ended on a red ERROR — rmdir: failed to remove '/var/lib/shairport-sync' — while the removal had worked perfectly. That line comes from the shairport-sync package's own removal script, which produces it deliberately and ignores it (the directory is obsolete and usually absent). Audiogravity was reading the severity of each line from its wording, so anything containing "failed" was painted red, whoever had written it. What the package manager prints is now shown as plain information; the verdict comes from the result of the operation itself.
  • [core] Simulation mode works again. Asking to simulate an installation instead of performing it failed every time, on both architectures, with an internal error message that said nothing about it.
  • [core] An installation that exceeds its time limit now says so. It reported an internal error instead of the timeout, so a package manager stuck waiting on something looked like a defect in Audiogravity. The limit itself is also enforced again: it had stopped applying to commands that spawn others — which is every installation. The abandoned command is also cleaned up properly instead of leaving a dead process entry behind.
  • [core] The log of a long installation no longer grows without limit in memory. The cap of 500 lines was silently dropped at the start of each operation, and output arriving without a line break was accumulated with no bound at all.
  • [core] Reading an installation log while it is being written no longer fails. Consulting the log of an operation in progress could fail outright, at the exact moment it is of any use.

[0.9.24] - 2026-07-28

Fixed

  • [site] The install command now really installs what the release notes describe. The one-line setup command fetches two smaller installers from audiogravity.app, and those two copies had not been refreshed since 30 June — five releases. So the 0.9.23 fix that lets a bare machine install itself was announced but never reached the only command users are given: a machine without Python still hit the old dead end. The access token was also still mandatory there, months after it had been made optional. Both copies are now in step with their source, and the release procedure checks it before every release.
  • [core] The configuration check no longer reports a file as missing while the editor has it open. Each audio service keeps its settings in a file, and where that file sits depends on how the service was installed: /etc when it came from the distribution's packages, /usr/local/etc when it was built from source. The location was written down in /etc/audiogravity/audio-config.json, which can only name one of them — so it was wrong on the other kind of machine. The Config tab coped by looking in both places; the configuration check did not, and flagged the file as absent on the very machine that was editing it. Every ARM box was in that situation for AirPlay. Both now ask the same question of the same source.
  • [core] A profile that starts one delicate service is no longer reported as inconsistent. Marking a profile as requiring care warned as soon as it also started an ordinary service — "Roon + HQPlayer" is delicate because of HQPlayer, and Roon Bridge beside it is ordinary. The only way to silence it was to mark those services delicate too, which changes how the box behaves when one of them refuses to stop. A warning you can only clear by degrading behaviour is one people learn to ignore. It now appears only when a profile asks for care while starting nothing that warrants it.

Changed

  • [core] Audiogravity now knows by itself where each service keeps its configuration. The path no longer has to be written in audio-config.json: the appconfigfile field is ignored, and new installations ship without it. Boxes set up earlier still carry the line — it does nothing, breaks nothing, and can be deleted. The way the file is read is decided the same way: a file is parsed according to the service it belongs to, rather than guessed from its name.

[0.9.23] - 2026-07-28

Fixed

  • [ops] Installing on a bare machine no longer stops at a missing prerequisite. The one-line install command checked that Python was present and gave up if it was not — while the installer inside the package it was about to download would have installed Python itself. That installer was never reached: the check runs first, so the repair was unreachable from the only install path anyone is ever told about. The command now installs what is missing before carrying on. It always could — it already runs with administrator rights.
  • [ops] The installer no longer reports installing things it did not install. Setting up a box pulls in a few system tools — the latency benchmark, the network test, the component that makes Tidal playable. If the machine could not refresh its package lists (no network, an unreachable mirror), the installation step was skipped entirely and the installer still printed "System dependencies installed". The box finished setup believing it had those tools and discovered otherwise much later, somewhere unrelated. The opposite case was no better: if the installation itself failed, the whole setup stopped dead on the package manager's raw error, before the service had even been created. Each step is now reported on its own, and what counts as success is what is actually on the machine afterwards, not what the package manager claimed. Anything still missing is named, together with the feature that will be unavailable — and the setup finishes either way.
  • [ops] Installing the interface on a machine without openssl no longer stops without saying why. The interface is served over HTTPS by default; that needs a certificate, and the installer creates one with openssl — a tool Debian marks as optional and does not guarantee is present. When it was missing the installer stopped dead immediately after announcing it was generating the certificate, and the error was discarded, so nothing at all appeared on screen to say what had gone wrong. Measured against the previous version: the install simply ended, silently. openssl is now installed when needed — only in HTTPS mode, since an installation sitting behind a reverse proxy never uses it — and if that cannot be done, the reason is printed. In the same step, a machine that does not have a network address yet no longer receives a certificate whose name field is empty.
  • [ops] Push notifications no longer depend on which packages a machine happened to already carry. Setting up a new box generates the keys that let it send notifications, and that step needs one cryptography library. It was fetched with pip, which is not installed on a stock ARM machine — so on ARM the step failed without saying so, and notifications worked only where that library happened to be present for an unrelated reason. It now comes from the machine's own package manager, which carries it on both architectures. Where pip did exist, this also stops Audiogravity from writing into the system's Python installation and hiding the version the operating system put there — something Debian deliberately protects against.

[0.9.22] - 2026-07-27

Changed

  • [ui] Opening the app no longer downloads one album cover per track waiting in the queue. The list of what is queued shows a small cover beside every track, and that list lives in the Library tab — which is built into the page from the very first moment, even when you are looking at another tab. A hidden image is still fetched, so a queue of sixty-seven tracks asked the box for sixty-seven covers before you had touched anything, all of them the same album seen sixty-seven times over. Measured on a phone with a cold start, that was half of everything the app requested while starting — 271 requests against 137 once fixed — and it grew with the queue: a few hundred tracks waiting meant a few hundred requests. Covers are now fetched only when they are about to come into view. Nothing on screen changes.
  • [ui] The Pipeline tab on a phone no longer keeps the box busy while you look at it. Left open, it asked the server for the entire pipeline every five seconds — and that answer is expensive to produce: measured at around 570 ms of work each time, which adds up to close to seven minutes of processor time every hour, on the very machine that is playing the music. It was also pointless: the core already recomputes the pipeline when it actually changes and announces it. The tab now reads the pipeline once when it opens and is told about every change afterwards, so an open tab costs nothing. The only thing still asked for on a timer is which output each source is steered to, which has no announcement of its own and is now checked every fifteen seconds instead of every five. Nothing on screen changes, and the pulsing dot next to each active source now fades in colour rather than shrinking — one less permanent drawing layer per stream on the devices least able to spare them.
  • [core] The message shown when the box refuses to start is now in English, like the rest of Audiogravity. When a security key is missing or too weak, the core refuses to start and prints what is wrong and how to repair it. That message — and only that one — was still in French.

Fixed

  • [core] Twelve more formats HQPlayer cannot read are refused up front instead of failing halfway. Audiogravity only refused the formats it had been told about one at a time; anything else was pushed to HQPlayer, accepted, and failed a moment later — pointing at the sound card rather than the file. Checked against Signalyst's own list of accepted formats, the refusal list now also covers AC3/E-AC3, DTS, Musepack, TAK, TTA, Shorten, Speex, AMR, MKA/WebM and AIFC (compressed AIFF). Nothing HQPlayer actually reads was added: the ten accepted source formats and the M3U/M3U8/PLS playlists are pinned by tests, because refusing a track that would have played is worse than a late failure.

[0.9.21] - 2026-07-26

Fixed

  • [core] A track a media server names .mp4 is now refused up front, like the same track named .m4a. HQPlayer decodes neither — they are two names for one container — but only the second spelling was on the refusal list. Serve your library through Plex, which publishes every M4A as file.mp4, and Audiogravity pushed the track anyway: HQPlayer accepted it, then failed, and you were told about a sound-card problem that did not exist. The same file served by MinimServer was correctly refused. On a library that is largely M4A, this was the common case rather than an edge one. The player also showed an empty format for those tracks, and now names the codec.
  • [ui] The handle that brings the player back no longer needs several attempts. The small tab at the bottom of the screen restores the Now Playing bar, and it used to ignore roughly one touch in two. Two reasons, both fixed: it only reacted to a movement under 10 px or an upward swipe past 30 px — doing nothing in between, which is exactly what a finger produces on a target 18 px tall — and that target was less than half the size Apple and Google recommend. The tab looks identical, but the area that answers the finger now extends above it, and any touch that is not a deliberate downward drag brings the player back.
  • [core] Asking for a real-time stream that does not exist now says so, instead of going silent. The live updates Audiogravity pushes travel on named streams — one for the machine, one for the player, and a few narrower ones. Asking for a name that does not exist used to open the connection and then send nothing at all, for ever: no event, no error, nothing to say the name was wrong. Audiogravity's own documentation even offered one of those wrong names as an example. An unknown name is now refused immediately, and the refusal lists the names that work.

[0.9.20] - 2026-07-25

Fixed

  • [core] A track sent to HQPlayer that never makes a sound now says so — and no longer holds up your click. Audiogravity checks that playback really started, but that check used to run while you waited: up to twelve seconds before the answer came back. Worse, it gave up too early on a heavy chain — measured on a DSD128 track upsampled to DSD, playback starts twenty-eight seconds after being sent, and HQPlayer stops answering at all while it warms up, so a perfectly good play was announced as failed fifteen seconds before the music started. The check now runs in the background: playback starts at once, and if no sound comes out the reason appears under the output, where a busy sound card is already explained. What counts as proof changed too — a position that advances, not HQPlayer's word: with the sound card held by another player it reports "playing" with its position frozen for as long as you leave it, which used to be taken as success. The message clears on its own as soon as the music really plays, however you restarted it.
  • [core + ui] Setting your machine up no longer needs a licence. Editing a service's configuration file, and the guided setup that generates the whole audio stack, were both reserved for Trial and Pro — so a Starter user could run the box but not configure it. Both are now open, and the Config tab is no longer greyed out. The guided setup stays administrator-only: it rewrites every service's configuration and mounts network shares, which is a question of role, not of edition.
  • [ui] The source list now updates the moment you connect or disconnect a service. Connecting HIGHRESAUDIO, Qobuz, Tidal or HQPlayer left the list unchanged — the screen you were looking at read it once, when it opened, and never again. You had to leave the page and come back. UPnP servers were worse: removing one never disappeared from the search picker at all. Every one of those actions now refreshes the list straight away.
  • [core] Importing a configuration is checked again in Starter Edition. The import itself was never licence-gated, but the validation that runs just before it was — so on Starter the check answered 403, the app logged it to the console and imported the file anyway. The gate removed a safety net from unlicensed users instead of protecting anything. Configuration validation is now open; the topology validation stays licence-gated, since it serves the Pipeline view.
  • [core] A source Audiogravity cannot identify is refused, instead of being taken for your local library. The classification ended on a catch-all, so anything unrecognised was declared "local" — and therefore eligible to be pushed to HQPlayer, which resolves tracks through your library and would have found nothing there. Anything Audiogravity cannot name positively is now refused with its reason, whatever the selected output. Nothing that played before can start being refused: the local path already required the very same recognition.
  • [core + ui] HQPlayer and network speakers no longer show up in the library source picker. While one of them was playing, it appeared among the sources you can browse, next to Qobuz and your local library — an entry with no catalogue behind it. They are outputs and processors, not sources: they stay visible in the player, which is where they belong, and are no longer offered as somewhere to browse.
  • [ui] The output line names your output even when nothing is playing, and tells you when it is stopped. It read the output's name from a field that is empty on a box whose audio layout declares no converter, which showed as "No output selected" while a perfectly good DAC sat there. The name, its state and its errors now all come from the same output list, so a speaker that is selected but idle is named — and marked Stopped rather than looking as if it were playing.
  • [core + ui] A selected speaker that cannot be reached now says so. The player state told you an output's name and what it was doing, but not whether it could be contacted at all — an asleep speaker was recognisable only by its playback state being blank, which also means "reached it, it said nothing". The two are now distinct, and the output line reads Unreachable instead of leaving you to guess.
  • [ui] "Up next" follows the output carrying the sound, not an internal handle. The fullscreen player decided where to read the upcoming track by testing an internal routing name. It now tests where the audio actually comes out, which is the same answer today and will keep working for any network output added later.

Changed

  • [core] A speaker you selected but cannot be reached now refuses the play, instead of quietly using the local output. Sending music while the chosen speaker was asleep, off the network or still reconnecting made it come out of the local DAC without a word — the wrong device, with no explanation. Audiogravity now says so and asks you to wake it up or pick another output. This is the rule it already applied when two outputs were selected at once.
  • [core] "HQPlayer" is no longer one of the origins a track can carry. An origin says what the music is — your library, Qobuz, a radio station. HQPlayer is a processor: what it plays keeps its own origin. Nothing produced that label any more; it was still advertised to clients.

Refactored

  • [core] One place now decides where music goes — all of it. Half the routing decision (HQPlayer or the local output) already lived in one module, but the other half — the local player or a network speaker — was copy-pasted at seven play paths, each asking the speaker manager itself. A path added without that paragraph silently stayed local, which is exactly the failure that decision point exists to prevent. Both halves are now answered in one call. No change for you: the same music goes to the same place, including the rule that "add to queue" never goes to a speaker (they have no persistent queue) while it does go to HQPlayer, which has one.

  • [core] The Now Playing split is finished. The modules extracted earlier were still reached through a compatibility shim left in the old file; every caller now points at the right module and the shim is gone, along with two places that reached into another module's private data instead of asking it.

  • [core] A track HQPlayer cannot decode is now refused straight away, whatever it came from. HQPlayer does not handle ALAC, AAC, OGG, Opus, WMA or APE — and until now only internet radio checked. Send it an ALAC album from your library, or an AAC track from your media server, and HQPlayer accepted the request, then failed to start: several seconds later Audiogravity reported it as a sound-card problem, which it never was. The check now sits on the one path every source goes through, so the answer is instant and names the real cause — "…is M4A, which HQPlayer cannot decode. Turn HQPlayer off as output to play it on the local output." An album is checked before anything is sent, so a single unplayable track is caught up front instead of stopping the music halfway through — with the trade-off that a mixed album (say a FLAC record with one ALAC bonus track) is refused whole rather than played without that track; the message names the offending track. Formats Audiogravity cannot identify are still attempted rather than blocked on a guess: only what HQPlayer is documented not to handle is refused.

Removed

  • [core] The two leftover routes that pushed music straight to HQPlayer are gone. They dated from before Audiogravity decided server-side where each track should play: back then the app had to pick HQPlayer itself. Since that decision moved to the box, nothing called them any more — the app even had a test forbidding their use, because going through them would skip the arbitration and let content reach HQPlayer that it cannot decode. Nothing changes for you: every way you start music already went through the normal path. Removing them means there is now exactly one door into HQPlayer, so a rule applied there cannot be bypassed.

[0.9.19] - 2026-07-20

Added

  • [core + ui] When an output can't play, Audiogravity now says why instead of staying silent. The sound card is exclusive by design (that is what makes bit-perfect playback possible), so only one player can hold it at a time. If you press play while another one still holds it, a notification tells you straight away — "Output in use by another player — stop it to play here" — and the fullscreen player repeats it under the output, with the engine's exact message as a tooltip. Previously you got silence and no explanation, and the reason was only readable if you thought to open the fullscreen player. The notification appears once when the problem starts, not on every refresh. Costs nothing: the reason was already in the data Audiogravity reads every second, it was simply discarded.
  • [ui] A refused playback now tells you why, instead of looking like a click that did nothing. When Audiogravity cannot start a track, the backend explains the reason precisely — but the three places you start playback from (browse, search, UPnP browser) dropped that explanation into the browser console, which nobody reads. All three now surface it as a notification, in the server's own words. This covers every refusal, not one case: an unreachable UPnP server, an expired stream, a source that cannot reach the current output. Enqueueing from the UPnP browser gets the same feedback as everywhere else too — it was the last place where adding a track could fail in complete silence. Successful playback stays silent by design: the music starting is the feedback.
  • [core] The reported output failure now always describes the current attempt. MPD keeps its last playback error indefinitely — the protocol says starting playback clears it, MPD 0.24 does not — so a problem solved minutes earlier stayed on screen for good, and would now have triggered a spurious notification too. Every playback command Audiogravity sends is prefixed with a reset, so the message you see is about the music you are trying to play right now.

Removed

  • [core + ui] "Spotify Connect" no longer appears as an available source. Audiogravity does not support Spotify, yet it still advertised it in the source list (a leftover service entry) — offering something that could never play. The service and its now-dead badge/label are removed. A Spotify stream reaching Audiogravity through AirPlay is unaffected: it is shown, correctly, as AirPlay.

  • [core + ui] The "Use HQPlayer as output" setting now follows your system, not your browser. It lived in each browser's local storage, so your phone and your laptop could disagree on where music should go — one sending playback to HQPlayer while the other sent it to the local output. The setting is now stored on the Audiogravity box, shared by every client and remembered across restarts. Better still, the routing decision itself moved to the backend: the app no longer chooses, so the two can no longer diverge at all. Audiogravity also now refuses to enable it when it could not actually play — no HQPlayer configured, or its network audio daemon (NAA) not running — telling you why instead of sending every play into silence. Playing a streaming source while HQPlayer is the output likewise says so plainly (only the local library can be routed through HQPlayer today) rather than doing nothing. The toggle also stays reachable while it is on even if HQPlayer stops answering (host asleep, app closed): the setting keeps routing playback from the box, so hiding its only off switch would have left every play failing with no way to stop it. You may need to re-enable the toggle once after updating.

Fixed

  • [core + ui] Your HQPlayer output setting can no longer be switched off behind your back. The app checked whether HQPlayer's audio daemon was running and, if not, turned the setting off for you. Harmless when the setting lived in your browser — but it now lives on the box, so any device merely showing the HQPlayer panel while that daemon restarted (which happens every time you change HQPlayer's audio output) silently disabled the setting on all your devices at once. Enforcement moved to the box: it refuses to enable the setting without a running daemon, and refuses to send a track to a dead one, telling you which daemon is missing instead of changing your choice. The switch also ignores a second tap while the first is still being saved — rapid flips could leave it showing the opposite of what was stored.
  • [core] Disconnecting HQPlayer now gives the sound card back. Disconnecting cleared the connection without stopping HQPlayer, so its audio daemon kept the card open — and with the address gone, Audiogravity had no way left to stop it: local playback stayed broken until a restart. It now stops HQPlayer before disconnecting. The app was already doing this on its own; anything else (a script, an older version) was left with a stuck card.
  • [core] Starting playback when HQPlayer's daemon is down now fails immediately, with the reason. The setting is remembered across restarts, so a box rebooting with that daemon stopped kept trying: every track waited a couple of seconds then failed, on every device, with no clue why. Audiogravity now checks first and answers straight away — "its network audio daemon (NAA) is not running on this box".
  • [core + ui] A cast you started is no longer labelled as coming from somewhere else. When you sent a radio station or a media-server track to a network speaker, the player could badge it External — the label meaning "another app is driving this speaker". Audiogravity was trying to recognise the music from its stream address, which is impossible for a radio URL and only works for a media-server track until an internal list fills up or the box restarts. The origin now travels with the track when Audiogravity casts it, so a station shows as Radio and a server track under its server's name. It still says External when someone really does take the speaker over — the address is checked, so the label stays honest.
  • [core] Music loaded on your speaker by another app is no longer mistaken for your library. A third-party controller pointing the speaker at a network share (an address that isn't a web link) was labelled Library, as if the track came from Audiogravity. It now shows as External.
  • [ui] The "External" badge shows its name and icon instead of raw text. The label was missing from the app's badge list, so it displayed the bare word external with a generic note icon.
  • [core] A slow-starting track through HQPlayer is no longer reported as a failure. Audiogravity checks that playback really began — but it gave up after 2.5 seconds, while HQPlayer reports "stopped" during its warm-up. On a heavy processing chain (upsampling to DSD, convolution) a perfectly good track was announced as failed just before the music started. It now waits long enough for that warm-up, and checks immediately instead of always pausing a quarter-second first, so a track that starts at once is confirmed faster. It also stops clearing its shared status reading, which used to make every other part of the app query HQPlayer separately exactly when it was busiest.
  • [ui] The progress bar no longer jumps backwards after you move it. Moving the cursor sent the bar to the right spot, then a status update already in transit — still carrying the old position — pulled it back for a moment before it jumped forward again. It looked like the move had failed, and invited a second attempt. The bar now holds your target until the player confirms it, and gives up holding on its own if the device really refused the move.
  • [core] The player no longer credits an idle speaker for sound coming out of your DAC. When a network speaker was selected but stopped, Audiogravity treated it as the destination — so a track playing locally (AirPlay, for instance) was labelled "→ ". Worse, the app only reports problems for the destination it believes is active, so a genuine local sound-card error stayed invisible in exactly that situation. The destination is now whichever output is actually carrying the audio; when nothing plays, the selected speaker keeps its place so an idle setup stays visible.
  • [core] A stopped network speaker can be started again from the player. After a restart of Audiogravity — or once another source had played — the player had no handle on the selected speaker, so pressing play did nothing and returned an error. The handle is now derived from the output list, which covers every case.
  • [ui] "Up next" follows the track you are looking at. With a cast running and another source playing at the same time, switching to that other source still showed the speaker's queue as its upcoming track. The upcoming track is now read from the item on screen.
  • [core] Radio now respects your output choice — it used to always play locally. The "send playback to HQPlayer" decision was wired into the library only, so starting a radio station while HQPlayer was your output pushed it to the local player instead — and since HQPlayer still held the sound card, you got silence and a "device busy" error. Every way of starting music (library, media server, radio) now goes through the same routing decision, so a new one cannot silently miss it again.
  • [core] Adding a track to the queue no longer stops what you are listening to. With HQPlayer as output, "add to queue" on a media-server track was treated as "play now": Audiogravity cleared the queue and jumped to that track, while telling you it had been added. Adding now appends, and leaves the current track playing. Adding also no longer overwrites the Now Playing card with the queued track's title and cover.
  • [core] Roon works again while HQPlayer is your output. Any Roon playback was refused with "this source cannot be played through HQPlayer — turn HQPlayer off to play it on the local output", which was wrong on both counts: a Roon zone is its own output and never uses the sound card HQPlayer replaces. Roon plays normally again, whatever your local output is.
  • [core] Selecting two outputs at once now tells you, instead of both pretending to work. Nothing prevented having a network speaker selected and HQPlayer as output: playback went to HQPlayer while the player showed the speaker as the destination. Audiogravity now says the two choices conflict and asks you to turn one off — until the output arbitration lands, it refuses rather than guessing which device you meant.
  • [core] Sending music to HQPlayer no longer reports success when nothing plays. Audiogravity used to report the outcome of its conversation with HQPlayer, not the outcome of the playback: it answered "done" when HQPlayer had accepted the track but could not actually start (typically because the local sound card was still held by another player), and conversely answered with an error on a play that was working fine. It now checks that HQPlayer really started and reports the reason when it did not — "its audio output is unavailable (most often the local sound card is still held by another player)".
  • [ui] Turning off "Use as output" for HQPlayer now gives the DAC back — local playback works again. The toggle only changed where Audiogravity sent new plays; HQPlayer stayed loaded and its network audio daemon (NAA) kept the sound card open, so playing locally afterwards failed silently (Failed to open ALSA device "hw:0,0": Device or resource busy). Switching the toggle off — or disconnecting HQPlayer — now stops HQPlayer so the card is released. Casting to a network renderer was unaffected (it never uses the local card).
  • [core] Pausing a cast no longer makes the mini-player vanish. A network-renderer cast (and any virtual source such as HQPlayer) is now kept visible in the mini-player while Paused — it was treated as "not playing" and filtered out, so pausing it collapsed the mini-player (the fullscreen was unaffected).
  • [core + ui] Seeking on a network renderer is more robust, and the progress bar follows. A seek now tries both ABS_TIME and REL_TIME (some renderers advertise both but honour only one, e.g. upmpdcli); the new position is refreshed immediately so the progress bar jumps to it instead of appearing stuck for up to the 5-second poll interval (with an optimistic bar update for instant feedback). If the device rejects every seek method (UPnP 501 for an HTTP stream it can't seek) Audiogravity disables the seek control for that track instead of erroring repeatedly — it re-enables automatically on the next track.

Changed

  • [core + ui] HQPlayer now appears as what it is — a processor in the audio chain, not a source. When Audiogravity sends music to HQPlayer, the player badges the item with its real origin (Library) and the signal path shows the full processing chain (e.g. Library → HQPlayer Embedded → Music Streamer → DAC) — HQPlayer is a DSP step, not where the music comes from. When HQPlayer is driven outside Audiogravity (its own remote), the item shows External with the live format — the HQPlayer API does not expose the track title for external playback, so that sub-case is deliberately title-less for now. Transport commands and the DSD amplifier protection keep routing to HQPlayer through its (unchanged) internal routing identity; the DSD volume-slider exception and DSD-lock badge in the player now key on the routing identity too, which also extends them correctly to network-renderer casts (both are self-managed devices).
  • [core + ui] A cast to a network renderer now shows the content for what it is — the renderer becomes an output. When you send music to a network renderer (a Marantz/Linn player, another Audiogravity box), the player now badges the item with its real origin — Qobuz, Library, the UPnP server's name, or External when another app drives the renderer — and shows the renderer as the output ("→ renderer name", output bar), instead of presenting the renderer itself as the source. Under the hood every now-playing item now carries two separate identities: what it is (display: origin, played_on) and what plays it (routing: control_id) — commands always go to the device that plays the content (MPD locally, the renderer for a cast), so play/pause/next keep working no matter how the item is badged. The player state gains a runtime outputs[] list (local DAC + selected renderer, with live transport state, built from in-memory caches only — zero added polling) with active_output_id and queue_next; a selected-but-stopped renderer stays visible through it (same on-screen result as before, no more synthetic "Stopped" item). The renderer's transport state now lives in one place (outputs[]): the player UI no longer consumes the renderer_status channel (kept for the Settings card), reactivity preserved via an event bridge (renderer GENA events → player state push, ~1s). The DSD amplifier protection is re-keyed on the routing identity, so a cast badged "Qobuz" can never have the renderer's physical volume forced to 100. Both control endpoints accept the new control_id handle (source_id still works).

Refactored

  • [core] The Now Playing service is split into focused modules (behaviour-preserving). The ~1550-line now_playing.py — which mixed playback-state aggregation, transport control, cover-art resolution and album tracklists — is broken into stream_meta.py, album_tracks.py, cover_art.py and transport_control.py; now_playing.py (now ~630 lines) only orchestrates and delegates. No API, event or behaviour change. The split adds 29 unit tests for the previously untested cover-art, album-tracklist and transport-dispatch paths, and folds the two duplicated MPD-escape helpers onto the canonical core.mpd_client.escape.

[0.9.18] - 2026-07-19

Added

  • [core+ui] Casting to a network renderer now shows a full player. When you cast audio to a native UPnP renderer (a Marantz/Linn network player, another Audiogravity box, any AVTransport device), it now appears as a first-class now-playing source in both the mini-player and the fullscreen player, with working transport controls, cover art, format/origin badges and signal path — the same experience as any local source. When the renderer is the selected output but stopped, it stays visible as Stopped (with the renderer as the output label) instead of the player collapsing to an empty state. Previously, casting to such a renderer showed no mini-player and its controls were routed through a fragile, inconsistent side-path.

Fixed

  • [core] Play/pause now works reliably on a native UPnP renderer. Pause/Play were sent through a helper that silently did nothing whenever its cached "allowed actions" list (fed by device events some renderers deliver unreliably) went stale — so play/pause looked dead while volume and next kept working. Audiogravity now sends the AVTransport Pause/Play command directly to the device.
  • [core] Renderer state (play/pause, track, position) now updates in real time instead of lagging up to a minute. The device-event callback was registered so that it rejected the actual UPnP NOTIFY requests (HTTP 405), and its no-API-key exemption never matched the real callback URL (HTTP 403) — so Audiogravity never received live events from a renderer and refreshed only via slow polling. Both are fixed; renderer state now follows within about a second. As part of this, after a pause/play command Audiogravity no longer briefly re-broadcasts the pre-command state.

Security

  • [core] The unauthenticated UPnP renderer event callback is hardened. The callback (which LAN renderers must reach with no API key) now only accepts events whose sender IP matches the renderer's own address, so a random host on the network can't inject spoofed playback state (it fails open for renderers that advertise a hostname rather than an IP, where the device-subscription id still guards). The API-key exemption for this callback is now an anchored path pattern that can't silently widen to a future nested route.
  • [core] Shared auth/infra hardened after a core code review. users.json (bcrypt password hashes + passkey credentials) is now written atomically at 0600 — a crash mid-write can no longer truncate it into an empty file that would wipe every account (admin lockout), and it is no longer world-readable. Three unused debug endpoints were removed: POST /publish/{channel} (which let any API-key holder inject arbitrary events into any dashboard SSE channel — spoofed alerts/state), GET /events/history/{channel} and GET /debug/sse/connections. Password verification (/auth/login and the admin re-authentication on privileged actions) now runs bcrypt off the event loop, so a login attempt no longer freezes other requests / SSE for ~250 ms. A non-ASCII API key now fails closed with 403 instead of raising a 500, and stdlib log records get the same secret-redaction backstop that structlog already had.
  • [core] The systemd services module is hardened after a seventh code review — including a root-RCE fix. Editing a service's tuning properties wrote them straight into a root systemd override file: fields like CPUAffinity, TasksMax, LimitNOFILE/LimitNPROC had no format validation, so a value containing a newline could inject an arbitrary directive (e.g. ExecStartPre=) that ran as root when the service restarted. Every property is now strictly validated (no newline/control characters), and the override generator refuses any embedded newline as a second layer. Lifecycle actions (start/stop/restart/enable/disable) and property edits are now restricted to Audiogravity-managed units — a standard user can no longer stop ssh, disable the AG backend, or kill networking through this API. The server always re-validates the override (skip_validation from the request is ignored), and the override file is forced to root:root ownership so it can't be rewritten without sudo.
  • [ui] Passkey login adapts to the anti-enumeration change on /auth/webauthn/login/begin. The server now returns 200 with empty allowCredentials for an unknown or passkey-less username (instead of a 404), so the login flow no longer relies on that 404: when a typed username has no passkey the UI shows a clear "No passkey registered for this account" message instead of prompting for — and possibly offering — another account's resident passkey.
  • [core] Cover-art proxy, auth and config writes are hardened after a sixth code review. The cover-art proxy (GET /audio_pipeline/cover?token=url:…) took a client-supplied URL and fetched it with no restriction — a server-side request forgery (SSRF) path to internal/link-local addresses (e.g. cloud metadata, AG's own loopback API). Cover fetches now go through an SSRF guard: only http/https, and hosts resolving to loopback or link-local are refused, while private LAN stays reachable (UPnP/MinimServer album art lives there); redirects are disabled so a public URL can't bounce into a blocked address. POST /auth/webauthn/login/begin no longer returns a distinct 404 for an unknown vs. passkey-less user (it always returns options), closing a username-enumeration oracle. Config writes now force root:root ownership so a same-filesystem move can't leave an /etc file writable by the service user. Config backups are pruned to the last 10 per service (they previously accumulated one per change, including each output switch).
  • [core] Package/engine downloads are hardened after a fifth code review. Downloads of artifacts that AG then runs or installs as root (engine install scripts, .deb files, GPG keys) are now HTTPS-only — a plaintext HTTP URL is rejected, closing a man-in-the-middle path to arbitrary root code (APT repositories are unaffected: they legitimately use HTTP because apt verifies a pinned GPG signature). Downloads are also capped at 500 MB (the existing limit was never enforced) and given read timeouts so a stalled or oversized response can't fill the disk or hang an install. A trial with a start timestamp in the future (clock skew) no longer reports more than the full trial length.

Fixed

  • [core] sysinfo / profiles / performance hardened after a fourth code review. Applying a CPU governor no longer blocks the event loop: the sudo/sysfs writes run off the loop (a synchronous write per core could stall audio on a multi-core box), and an unknown CPU id now returns a clean error instead of a 500. A failed self-update launch (missing/misconfigured root wrapper) is reported as failed immediately instead of leaving the box showing "updating" for ~15 minutes. The critical-temperature and profile-activated push notifications keep a strong task reference so they can't be garbage-collected before they send (and the profile notification's completion callback no longer risks an error on cancellation). Bare except: clauses in the sysinfo service were narrowed so they no longer swallow task cancellation.
  • [core] Player / push / steering hardened after a third code review. DSD playback no longer risks getting stuck at volume 100: a transient failure while forcing volumes to 100 at DSD start now keeps the saved original so it is restored when DSD ends. Push notifications only prune a subscription on a real HTTP 410 (Gone) status — a transient error whose text merely contains "410" no longer silently drops a valid device — and expired endpoints are removed in a single batched save instead of one full-file rewrite each. The subscriptions file is written 0600 from creation (shared atomic private-writer) with no world-readable window. The steering restart-fallback that rewrites mpd.conf now targets only the alsa output block and refuses to act on an ambiguous multi-output config instead of collapsing every output to one device. Steering's output-switched SSE events keep a strong task reference so they can't be dropped before firing.
  • [core] Playback modules hardened after a second code review (UPnP renderer / HQPlayer / library). Qobuz browsing now recovers from an expired user session by actually re-logging in before retrying, instead of replaying the same token and misreporting the failure as rotated app credentials (which sent the user to re-extract the bundle for nothing); the HQPlayer status read tolerates a malformed numeric field instead of 500-ing the whole endpoint; a single-track HQPlayer play keeps its album in the now-playing card; and the UPnP renderer's short-lived background tasks (the play-state confirmation poll and the SID-mismatch resubscribe) now keep a strong reference so the GC can't collect them mid-flight.
  • [core] Streaming modules hardened after a code review (Tidal / Qobuz / HIGHRESAUDIO / Radio). A batch of reliability fixes with no user-visible feature change: Tidal now drops an expired access token whenever a refresh fails (revoked session, network error) instead of replaying a dead token — this stops a per-request re-login storm and the misleading "client credentials rotated" error it caused; Qobuz sign-in no longer leaves a half-open connection (reported "connected" while streaming has no working secret) when the app-secret probe fails, and a bundle/network failure during sign-in now shows the styled error page (HTTP 502) instead of a raw Internal Server Error in the browser popup; the Qobuz app-bundle cache is now genuinely single-flight (concurrent cold-cache callers no longer each download the ~200 KB bundle); the Tidal remux proxy always cleans up its temporary files when a client disconnects mid-track (previously an orphan .mpd/.part could leak on cancellation); Radio favourites/library mutations no longer serialise behind a slow catalogue lookup, and the Hi-Res only search over-fetches so a page isn't starved down to a handful of stations. Hardened along the way: MPD argument escaping now strips newlines at the single choke point every MPD write passes through (defence-in-depth against control-character injection via a stream URL). Tidal also reuses the shared pooled HTTP session (keep-alive) instead of opening a fresh connection per track.

[0.9.17] - 2026-07-18

Added

  • [core + ui] Add a NAS share (CIFS/SMB) straight from the music-library picker. The Initialize and Guided library pickers gain an Add network share (NAS) panel: enter the host, share and (unless it is a guest share) credentials, and Audiogravity mounts the share at /mnt/<slug> and actually mounts + tests it before confirming — read-only by default. On failure the exact mount error is shown and every artifact is rolled back; on success the share is selected as the library. The panel also lists the shares it created (MOUNTED / ON-DEMAND / LIBRARY badges) with a remove action that refuses to break the active library unless forced. New routes GET/POST /audio-stack/mounts and DELETE /audio-stack/mounts/{slug}; systemd .mount/.automount units (mount-on-access, robust to a NAS that is off at boot), CIFS credentials in a 0600 file only, connectivity test done as a non-blocking start + state poll (no orphan mounts). NFS stays terminal-only by design (mounting it would pull RPC daemons onto the box). No fstab is touched. Mountpoint privilege goes through a root-owned pinned helper (ag-mountpoint.sh) — no /mnt/* sudoers glob.
  • [core + ui] The music library re-indexes automatically after a change, with a live indexing indicator. Changing MPD's library — Initialize, the Guided Apply changes, or mounting a NAS share and selecting it — now triggers an MPD database rescan. The minimal config keeps auto_update off, so without this the Library view kept the old (or empty) content until a manual update. A transient "Indexing library…" row (new GET /audio-stack/library-scan-status, polled ~1.5 s) shows the scan is running and clears when MPD settles; it also re-attaches to an in-flight scan if you leave and return to the Config tab, and targets the port pinned in mpd.conf. The rescan is best-effort: the common case (MPD already up) fires inline so the scan is queued before the response returns, while a not-yet-ready MPD retries off the request path so provisioning never blocks.
  • [site + ui] The FAQ becomes a manual chapter — and the landing derives from it. A new manual chapter (11-faq.md) is now the single source of truth for the ten FAQ entries; the landing's FAQ accordion is generated from it by audiogravity.ops/scripts/sync-landing-faq.py (marker-delimited block in index.html, chapter links absolutised to the published manual — never edited by hand again). The in-app reader picks the new chapter up automatically through the live-derived TOC; the offline fallback list follows.
  • [site] The manual grows to 11 chapters, illustrated. Two new chapters — Quick start (zero-to-music on one page, now the in-app reader's landing view) and a Glossary (~20 audiophile/UPnP terms) — plus new how-to sections: mounting a NAS share (CIFS/NFS fstab recipe, matching what the library picker actually detects: network shares and local mounts under /mnt), installing the app (PWA — on iOS the installed app is required for push), getting HTTPS for passkeys/push (LAN domain + Let's Encrypt DNS-01 + reverse proxy to the UI in plain-HTTP mode on 8080), backup & restore (including the editor backups under /var/backups/audiogravity), a lost-admin-password recovery, and UPnP-discovery / AirPlay-visibility network troubleshooting (SSDP/mDNS, IGMP snooping, VLANs). Eleven real screenshots (iPhone rendering, Minimal light theme) illustrate first-run, listening, library, outputs, administration and updating. A coverage pass also fixed wrong or missing facts: sign-in documents the seeded default admin/admin123 account (and says to change it), the Admin tab is named and mapped (users, announcements, updates, licence), the tab map marks Library as Pro, and installation no longer claims an account-creation step.
  • [ui] The in-app manual reader renders the expanded manual. The chapter list grows to 11 (Quick start lands first), sidebar numbers derive from the chapter slugs (so the new chapter 0 doesn't shift them), and chapters are enhanced at cache time inside an inert template: images are absolutised against the manual base and marked loading="lazy" before they ever reach the DOM — no wasted wrong-origin fetches, and chapter revisits reuse the enhanced HTML with zero rework.
  • [ui] The manual sidebar derives from the published table of contents. The reader now fetches and parses the manual's README.md "Contents" list at open time, so a chapter added or renamed in the site repo appears in the app without a UI change; the hardcoded list remains only as an offline fallback. Sidebar numbering comes from each chapter's own number.
  • [ui] Manual screenshots ship as WebP at 2× — ~0.3 MB instead of 1.65 MB. All eleven illustrations are re-encoded (visually identical in their 360-px slots), and the capture pipeline that generates them is now committed (tools/shoot-manual.cjs: Storybook stories + the login page, iPhone viewport, Minimal light theme, demo covers, in-browser WebP encoding) so they can be regenerated reproducibly after any restyle.

Changed

  • [site] Landing readability pass. Feature-card copy tightened to scannable length (the audiophile vocabulary — bit-perfect, gapless, FIFO, DSD — stays), the hero's disabled trial button is removed in favour of a single primary CTA (Early access — 30-day trial included), and below-the-fold images (hero carousel, PWA screenshot) are lazy-loaded.
  • [site] Landing accuracy and freshness pass. Corrected claims: the push-notification triggers now match the code (service down, critical CPU temperature, update available, profile activated — there is no DAC-disconnect alert and the threshold is fixed); output rerouting is described as gapless on MPD with AirPlay warning before its receiver restarts (three spots); "fourteen modules" becomes twenty-one; the zero-cloud stat names the cover-art/radio-catalogue lookups; the Auto-discovery card is retagged Pro (it lives in the Library tab); and Install mentions the one-click engine installs before the guided setup. New content: a full-width one-click self-update feature card closes the features grid, two FAQ entries (Where is the documentation? — the illustrated manual, in-app and online — and How do I update?), and the library/transport cards gain the 0.9.16 features (favorites star, artist drill-down, per-source queue badges + filter).
  • [site] Complete the user manual with everything shipped since 0.9.13. Listening gains a full Queue section (streaming albums with durable titles, header named by the real source with the station logo, per-source badges + display-only filter on mixed queues, remove by button or reindex-safe left-swipe, filter-aware Clear) and a Portrait lock section; Library & streaming documents the search artist drill-down and the streaming-album favorites star; Outputs & engines describes the gapless MPD output switch, the AirPlay restart warning and honest failure reporting. An API sweep (178 live routes crossed with the manual) adds the Settings panel section (themes, compact mode, animations, API key, Face ID/Touch ID), push notifications (the four wired alerts), announcements, and the Audio Software UPDATE ALL badge; left-swipe removal is noted on the UPnP server and renderer lists, and the Admin-tab update marker in Updating.
  • [site] Document the audio-engine installation step in First run. The installer sets up Audiogravity itself, not the audio engines it conducts — the manual now says so (Installation) and First run gains an explicit "Install the audio engines" step (Software tab: MPD first, then AirPlay / UPnP / Roon / HQPlayer NAA as needed) before the guided configuration.
  • [site] Fix manual links that rendered as raw Markdown. GitHub Pages serves .md files unrendered, so the two links escaping docs/manual/ broke in the in-app reader: Requirements now points at the manual's own "What you need" section, and the release-notes link uses the GitHub-rendered URL (the same canonical target as the landing footer and the update banner).
  • [ui] Add Storybook stories for the manual-illustration screens. New populated stories for the playback queue (mixed-sources badges + filter), the fullscreen player, the output selector and the album browse grid — real components with demo data, no backend needed. The shared lib-* style injection is extracted as injectLibStyles() so stories can mount library organisms outside the page shell.

Fixed

  • [core] NAS mounts negotiate the SMB dialect instead of pinning SMB3. The CIFS mount dropped its hardcoded vers=3.0, which a NAS/Samba that enforces a different minimum protocol rejected with mount error(95): Operation not supported; the kernel now negotiates the highest dialect both ends support (3.1.1 down to 2.1, SMB1 still excluded by default). CIFS credentials also derive their directory from the configured config_dir (not a hardcoded /etc/audiogravity/creds), so a dev server that does not run as the service user can mount too.
  • [ui] The dev-server CSP allows the manual's illustrations. The img-src directive of the development Content-Security-Policy (index.html meta) now includes https://audiogravity.app, so the in-app manual's screenshots render on the Vite dev server instead of being refused (a box repointing AG_CONFIG.manualBase must allow its origin the same way).
  • [ui] Components import the child elements their templates use. ag-config-panel (ag-switch) and the fullscreen player (ag-volume-popover) relied on the app-wide registration in main.js and rendered without their toggles/volume control when mounted standalone (Storybook, isolated tests). The same gap was then fixed across the board — 30 more components — and a guard test now enforces the invariant: every <ag-*> tag used in a component's template must be reachable through that component's own import graph.
  • [ui] Swiping in from the right edge to open Settings no longer also deletes a queue track. On the queue (and the other swipe lists), a right-edge swipe to open the settings panel could simultaneously trigger a row's swipe-to-remove and delete an item. A touch gesture that starts in the screen-edge band is now reserved for the panel-open swipes (Settings from the right, sidebar from the left), so it opens the panel only; a mouse drag is unaffected. The reserved-band width is shared between the tab-swipe and the row-swipe guards so it can't drift.
  • [ui] iOS Safari no longer zooms in when you tap a text field (login and throughout the app). Per-component field styling (14px) was overriding the rule that keeps mobile fields at 16px to prevent the zoom; the anti-zoom rule now takes precedence, so iOS stops auto-zooming on focus.
  • [ui] The self-update banner no longer renders the product name as "Audiogravi ty". A flex layout on the "Updating…" title split the superscript with a gap; the title is kept inline again.

[0.9.16] - 2026-07-13

Added

  • [core + ui] Add or remove favorites for Qobuz, Tidal and HIGHRESAUDIO albums. A star toggle on streaming album cards (browse grid) and search rows adds/removes the album to your account's favorites on that service. The star shows the accurate state (filled when already a favorite), updates instantly (optimistic, reverts if the service call fails), and stays in sync across the browse and search views. New backend routes POST/DELETE /library/favorite and GET /library/favorite-ids dispatch to each service's favorites API (Qobuz favorite/create|delete, Tidal favorites, HRA MyAlbum).
  • [core + ui] Swipe a queue row left to remove it. Up-next tracks can now be removed with a left-swipe (in addition to the Remove button), on the queue and — through the same shared gesture — the radio, UPnP-server and UPnP-renderer lists. Removal is keyed on the stable MPD song id (deleteid / QueueItem.queue_id), so it is reindex-safe: the intended track is always the one removed, even if the queue reordered. Internally the four previously hand-copied swipe implementations are unified into one shared controller + Lit directive.
  • [core + ui] The playback queue now shows Qobuz, Tidal and HIGHRESAUDIO albums, with per-source badges and a source filter. Selecting a Qobuz/Tidal/HIGHRESAUDIO album used to open an empty Queue — those services play over the shared MPD engine, and the queue endpoint returned nothing for them. The queue now lists their tracks with real titles, and those titles are written back as MPD tags at enqueue so they survive a core restart. When the queue mixes providers, each up-next row shows a small source badge (Qobuz, Tidal, Radio, Library…) and a filter bar lets you narrow the view to one source; the filter is display-only, so track positions, removal and playback are unchanged, and the currently-playing track is always shown. Remove and Clear now work on these queues too, and Clear respects the active filter. Now-Playing's next-track peek fetches a bounded window instead of the whole queue.
  • [ui] A portrait orientation lock. A new Portrait Lock switch in Settings (on by default, phones + tablets) keeps Audiogravity in portrait. On Android the installed PWA is locked via the web app manifest; on iOS — which has no orientation-lock API — a full-screen "Rotate your device" overlay blocks landscape, with a Continue in landscape escape that turns the lock off (so a device physically fixed in landscape is never stranded). Desktop/mouse installs are unaffected (the toggle and the OS lock are scoped to touch devices), and the overlay marks the rest of the app inert while shown so keyboard / assistive tech can't reach the hidden UI behind it.

[0.9.15] - 2026-07-13

Added

  • [ui] An in-app user manual. A new Manual entry at the end of the tab bar (far right in the horizontal layout, bottom in the vertical sidebar) opens a full-screen reader: a chapter sidebar plus a reading pane that fetches the Markdown manual published at audiogravity.app and renders it in place. The Markdown renderer (marked) is lazy-loaded on first open, so it adds nothing to the app's startup. Links inside the manual never navigate the app away — chapter links switch chapters in the reader, in-page anchors scroll, and external/other-doc links open in a new tab (and resolve correctly for middle-click / "copy link address"). The manual base URL is repointable per box via AG_CONFIG.manualBase.
  • [site] A "Made in EU" badge and a manual link in the landing footer. The footer now shows a Made in EU badge after "Made with care". The Resources column drops the redundant "Landing" link, leads with Releases, and adds a User manual link to the GitHub-rendered manual.

Fixed

  • [core + ui] The playback queue is labelled by what is actually playing, not the transport engine. Radio, Qobuz, Tidal and HIGHRESAUDIO all play over the MPD engine, so a radio stream queued from the library showed the header "Queue of Local Library" with a missing cover. Queue items now carry the stream origin (GET /library/queue gains an origin field, mirroring now-playing), so the header reads by the real source (e.g. "Radio") and a recognised radio stream uses the station logo as its cover.
  • [core] A stuck "update already in progress" no longer blocks the self-update. is_in_progress() judged staleness from started_at, which the detached shell updater never writes (it only refreshes updated_at), so a killed update could leave the state file reading "installing" forever and every later update returned "An update is already in progress". Staleness now anchors on updated_at (the shell's heartbeat) with a started_at fallback, so a dead update self-recovers after the stale window. Recovery is documented in the manual (Troubleshooting).
  • [ui] The admin About licensing info modal opens again. Its handler still called a helper removed by an earlier XSS-hardening refactor and threw; it now builds the modal content as a Lit template.
  • [ui] Loading spinners keep animating with reduced-motion (or animations off). The global motion-reduction rules froze every animation, including functional loaders — so a real operation (e.g. a self-update) looked stuck. The icon spinner (.ag-spin) and the CSS loader (.spinner) are now exempted.
  • [ops] The core service is packaged with NotifyAccess=main. With NotifyAccess=all, a stray sd_notify(ERRNO=…) from any child that inherited $NOTIFY_SOCKET could leak into the unit's status and surface as a misleading "Error: code: N" in systemctl status.

[0.9.14] - 2026-07-12

Fixed

  • [ui] The library no longer offers to switch to "Local Library" while you are streaming Qobuz/Tidal/HIGHRESAUDIO. Those providers (and local files) all play over the MPD engine, so the now-playing banner compared the transport engine against the browsed source and spuriously prompted a switch to "Local Library" when you played Qobuz from the Qobuz tab. The banner now compares at the source level (using the stream origin, like the library view does), so it stays quiet when the playing source matches what you are browsing — and when it does fire (a genuinely different source), "Switch" targets the right browse source (e.g. Tidal, not the MPD engine).
  • [ui] The mini-player source badge shows the generic origin ("Radio", "Qobuz"…) instead of the station/provider name. A long radio station name (e.g. "CLASSIC VINYL HD") overflowed the badge; it now shows the generic origin label like the source badge on the fullscreen cover. The full station name stays visible in the track title.
  • [ui] The brand renders consistently as Audiogravity across the UI. Assorted screens spelled it "Audiogravity"; the wordmark is now applied wherever the text renders as HTML (update banner, several help modals, the install prompt) and plain "Audiogravity" where a tag would show literally (tooltips, dialog titles, attributes).

[0.9.13] - 2026-07-12

Added

  • [core + ops] A fresh box now ships with a few default internet-radio stations. A radio.json.example (FIP, Classic Vinyl HD, Le Son Parisien) is deployed to /etc/audiogravity/radio.json on a first install when the file is absent — the same copy-if-absent pattern as audio-config.json and audio-topology.json. An upgrade never overwrites it, so your saved stations are preserved. Documented in the manual (Library & streaming, Updating).

  • [ui] The Admin tab flags an available software update. Mirroring the announcements/news bell, the Admin navigation tab now shows a small download indicator when the license server reports a newer Audiogravity release — a required update uses the warning colour, an optional one the accent colour. It is fed by the Admin page's update banner over a new update-badge window event (the same /license/online-status signal), so — exactly like the news bell — it appears once the Admin page has been opened at least once.

  • [core + ui] Library search — tap an artist to browse their albums. An artist in library search is now a drill-down: tapping one opens an Albums by … view (with a back control) instead of failing to queue — an artist is a navigational entity, not a playable item, so the add-to-queue action is gone from artist rows. Backend: GET /library/albums?artist_id=… now resolves an artist's albums for Tidal (artists/{id}/albums) and HIGHRESAUDIO (name-based via quickSearch filtered to the exact artist — HRA has no artist-id catalogue endpoint) alongside the existing Qobuz / MPD support. UPnP/DLNA and Roon artist rows stay inert (no artist-albums endpoint / session-scoped browse keys — use their dedicated browsers). The artist view uses its own browse instance, so returning to browse keeps your place.

Changed

  • [core] The default topology labels the AirPlay receiver "AirPlay Receiver". In the shipped audio-topology.json.example, the shairport-sync node is now labelled AirPlay Receiver (was AirPlay), so a fresh box's signal path reads AirPlay → AirPlay Receiver → … instead of the confusing duplicate AirPlay → AirPlay (the first hop is the source origin, the second the on-box receiver service). Existing boxes keep their own topology untouched — rename the shairport-sync label there to get the same.
  • [ui] Dropped a stray developer host from the shipped Content-Security-Policy. The img-src directive in index.html / login.html whitelisted a leftover dev reverse-proxy origin that nothing in the app ever loads from; it has been removed. The production CSP is served by Nginx/Lighttpd regardless, and local dev-server access is unaffected.
  • [license server] The customer license-key email is redesigned to match the Audiogravity landing. The plain HTML message is replaced by a brand-coherent, email-safe template: an amber accent bar, the AG wordmark with the app logo embedded inline (Content-ID cid:aglogo, so it renders without the recipient having to load remote images — no blocked-image placeholder), a highlighted key block, numbered activation steps, and a sign-off that points explicitly to [email protected] (the Reply-To, since the message is sent from license@). Layout is inline-styles/table-based for broad client support (border-radius degrades to a square on Outlook desktop). The logo ships with the server (pics/apple-touch-180.png) and the email still sends if it is ever absent (logo simply omitted). The admin template preview resolves the inline-logo cid: to the statically-served asset so the logo renders in the editor iframe too.

Fixed

  • [core] Qobuz library search no longer returns an error for every query. Qobuz returns an explicit null for an absent artist / album / performer / image field, and dict.get(k, {}) returns that null rather than the {} default — so a chained .get raised AttributeError and failed the whole search (a single imageless artist broke every query, HTTP 500). A shared _dig() helper now null-safely walks these nested fields at every Qobuz search / album / track / queue site, not just one. Search-failure logs now carry the traceback.
  • [ui] Admin views no longer crash their render when a data endpoint returns a non-array. ag-rt-monitor (real-time processes) and ag-config-panel / ag-passkey-manager (passkey list) fed a property straight from apiGet into .map(); if the response was ever a non-array — an unexpected payload shape, or an empty/undefined body that didn't throw — the render threw .map is not a function. Each load now coerces the response to an array (matching the inline Array.isArray guards used elsewhere), so the view degrades to empty instead of white-screening. These were surfaced as unhandled errors in the Storybook test run.
  • [ui] The mobile navigation sidebar no longer gets stuck hidden after visiting the pipeline screen. An interrupted edge-swipe (a gesture that turned vertical, or ended without committing — made easy to trigger by the pipeline's horizontally-scrollable output switcher) could leave a leftover inline transform on the sidebar. Because inline styles override the class-based CSS, the menu stayed off-screen while its state said "open" — it then refused to open by button, swipe or the burger icon, and only fully closing and reopening the app fixed it. The leftover transform is now cleared on every drag-end path.
  • [ops] A self-update no longer risks skipping the co-located UI update (version skew). The detached self-update.sh runs install-core.sh, which reinstalls — and therefore overwrites — self-update.sh itself mid-run. Because bash reads a script lazily by byte offset, that same-run overwrite could misalign the read and silently skip the steps that follow the reinstall — including the co-located UI update — leaving the core on the new version while the UI stayed on the old one (login screen showing the previous version). The script's body is now wrapped in a { … } block, forcing bash to read it fully into memory before executing, so a mid-run replacement can no longer derail it. The UI-update skip path also logs which check made it skip (missing UI directory vs missing ag-ui-server unit) instead of returning silently, so a future skew is diagnosable from self-update.log. Finally, the co-located UI service is now detected by querying its unit directly (systemctl cat, retried a few times) instead of grepping the full systemctl list-unit-files output — that enumeration could come back empty/partial while systemd was still settling the reload the core reinstall had just triggered, false-negatively skipping the UI update; a real end-to-end update surfaced this second, rarer skew path.
  • [license server] The "Release notes" link on the update banner no longer leads to a dead page. When the licence server auto-published a discovered release to the fleet, it carried over the GitHub release page URL of the private releases repo as the notes link — which 404s for end users. The auto-publish paths now drop that URL so the offer falls back to the public RELEASE_NOTES.md, matching what a manual publish already did. (Boxes already offered the previous release must be re-published once for the corrected link to take effect.)
  • [ui] The update banner spells the brand as Audiogravity. The "A newer … release is available" line rendered the name as plain "Audiogravity"; it now uses the same styled wordmark as the rest of the app.

[0.9.12] - 2026-07-10

Added

  • [core] audio-topology.json is now validated — structurally and semantically. A new POST /config_validation/validate-topology endpoint checks the user-declared hi-fi chain: a malformed file or an unknown device type is a blocking error; a broken reference (an output pointing at a target_device_id/target_input_id that doesn't exist) or a streamer output whose connector maps to no real audio output is a non-blocking warning (the topology only feeds the signal-path view, so it never blocks the app). The same file is validated at startup (best-effort — problems are logged, never fatal). The config_validation Pydantic models for the topology (AudioTopology / HiFiTopology / AudioDevice / AudioInput / AudioOutput) — previously defined but unused — were corrected and wired to this endpoint. A generic, fully-commented audio-topology.json.example now ships with the box and validates cleanly.
  • [ui] The topology editor validates before saving, and can download / upload the file. In Audio Pipeline → CONFIG, saving now runs the topology through validate-topology first: structural errors block the save and are shown, broken links / unmappable connectors are surfaced as warnings you confirm before writing. New Download and Upload buttons on the editor let you save audio-topology.json to your computer and load a file back in for review (the same save-time validation then applies) — parity with the audio-config editor. The generic JSON config modal gained an opt-in allowFileTransfer for this.
  • [site] User manual — the audio topology is documented. A new Audio topology (signal-chain map) section in the manual explains what audio-topology.json is, its role (it drives the Audio Pipeline graph and is never rewritten by Audiogravity), how physical outputs are resolved live from real hardware while the file declares the downstream chain, its structure (with an example), how to edit / download / upload it, and how to keep it up to date. Cross-linked from the Outputs & engines chapter.
  • [site] User manual — the audio configuration file is documented. A companion Audio configuration (services & profiles) section explains audio-config.json — the registry of managed services and one-tap profiles that feeds the Services and Profiles tabs — its structure (services / profiles / topology_link), how it's exported/imported from Settings, and how it's validated on import (structure plus systemd-unit and config-file existence).

Fixed

  • [core + ops] Reading the DAC hardware volume no longer spams the logs, and refreshes on change instead of polling. The core read the DAC's ALSA volume via sudo amixer on every signal-chain refresh, so sudo audit-logged the command in a loop; the service also lacked direct mixer access (hence the sudo). The service user is now added to the audio group so it reads/writes the mixer directly (no sudo, no log spam), and the 5-second poll is replaced by an event-driven alsactl monitor that only re-reads when a control actually changes — so the signal-chain view reflects an external volume change (Roon/AirPlay…) instantly while doing no work in between. Volume handling is unchanged for every DAC. Takes effect after reinstalling and restarting the core.
  • [core] HQPlayer no longer floods the logs when it's switched off. The backend used to poll HQPlayer's control API on every now-playing refresh and log "HQPlayer unavailable" roughly every 10 s whenever HQPlayer wasn't running. That poll only fed the now-playing view, which is relevant only when HQPlayer streams to this box through the local NAA (networkaudiod). It is now gated on an event-driven NAA-active flag (seeded once, then refreshed only on a networkaudiod systemd job — no periodic polling), so when the NAA is off the poll and its log spam disappear, along with the wasted cross-network connection attempts. Controlling HQPlayer from AG (status, filters, modes, volume, connection) is unchanged. Also demotes the "unavailable" log to fire only on the reachable→unreachable transition.
  • [ui] API.md now documents the config-validation endpoints (/config_validation/validate and /config_validation/validate-topology), and the audio-topology.json view/save routes are described accurately (the save route writes the whole topology file with a backup and hot-reload — it was previously mislabelled "save node positions").

[0.9.11] - 2026-07-10

Added

  • [license server] Campaign emails are throttled and drained across days within the provider's daily send cap. A campaign's recipients are snapshotted into a new mailing_queue table; a background loop delivers them in batches, capping campaign sends at MAILING_CAMPAIGN_QUOTA_PCT% (default 80) of MAILING_DAILY_QUOTA (default 100) attempts per UTC day — successes and failures both count, so a failing campaign can't hammer the provider — leaving headroom for transactional license mail, which is never throttled. A large list therefore drains over several days on its own; already-sent recipients are never re-sent, and opt-outs / revocations after the snapshot are honoured (skipped). POST /ls/admin/mailing now returns {total, sent, pending, failed, skipped, daily_cap} and the admin mailing panel shows per-campaign progress (sent / pending). New env vars MAILING_DAILY_QUOTA / MAILING_CAMPAIGN_QUOTA_PCT / MAILING_DRAIN_INTERVAL / MAILING_DRAIN_ENABLED; new module mailing_drain.py. No new runtime dependency.
  • [license server] Operator manual. A new docs/manual/ in the license-server repo — an 11-chapter Markdown guide (introduction, installation, getting started, selling & delivering licenses, managing orders, the customer portal, email & campaigns, fleet updates & announcements, maintenance, troubleshooting). The operator README was also refreshed to match the current feature set and API, and a THIRD_PARTY_NOTICES.md was added listing the server's open-source components and their licenses.

Changed

  • [core + ui] MPD switches its physical output live, without a restart. mpd.conf now declares one bit-perfect audio_output block per detected output (USB / optical / analog / HDMI), and switching hot-enables the target over MPD's control socket (enableoutput/disableoutput) instead of rewriting the config and restarting the daemon. Switching MPD's output is therefore gapless and a cast in progress (UPnP renderer through MPD) survives it. The switch opens the target device transactionally, so a genuine failure is reported rather than raced. AirPlay has no equivalent, so changing its output still restarts the receiver — the Outputs panel now warns first that this interrupts any AirPlay playback in progress (AirPlay is sender-driven and cannot be resumed). Migration: an already-deployed box keeps its single-output mpd.conf until it is re-provisioned (INITIALIZE / "reset to minimal"), which regenerates the multi-output config; until then MPD output switching automatically falls back to the restart path (functional, just not gapless).
  • [core] Physical output switching now resolves the ALSA device from real hardware instead of a static topology file. Steering (/steering/switch-output, /steering/outputs) and the pipeline visualization derive each output's live hw:X,Y by classifying detected ALSA hardware (USB / S/PDIF / analog / HDMI) and joining it to the declared hifi_topology by its connector — the static system_device_id was removed from audio-topology.json. This fixes the output mapping drifting after a reboot or USB re-enumeration (it was previously correct only by luck of the card-enumeration order), makes an HDMI output selectable, and keeps the pipeline's active-output indicator correct across reboots. API response shapes are unchanged. hifi_topology stays a user-declared, manually-maintained description of the hi-fi chain (Audiogravity never rewrites it).
  • [license server] Outbound email routed through Resend, with distinct license and communication sender identities. The transport (single _smtp_send SMTP choke point) is unchanged in mechanism, but broadcast/communication mail now uses its own campaign From (SMTP_CAMPAIGN_FROM, e.g. news@) separate from transactional license mail (SMTP_FROM, e.g. license@), and every message can carry an optional Reply-To (SMTP_REPLY_TO, e.g. support@). The envelope sender (MAIL FROM) is now the bare address parsed from the From, so display-name forms (Name <addr>) no longer leak into the envelope (RFC 5321). Two new admin settings smtp_campaign_from / smtp_reply_to on PUT /ls/admin/settings/smtp (persisted, applied at runtime, editable in the SMTP admin card) plus the matching env vars; default SMTP host is now smtp.resend.com. No new runtime dependency (stdlib smtplib); lit added as a dev-dependency so the settings component can be unit-tested (Vitest).

Fixed

  • [ui] The mobile signal-chain (pipeline) view now matches the rest of the app. It was styled with its own hardcoded palette, radii and font sizes, so it looked off from every other screen and did not follow the dark theme. It now uses the shared design tokens throughout, drops the coloured gradient tint and the green "glow" on status dots, inherits the standard page background, and the "Signal chain" heading is the same neutral colour as "Now playing" (no longer tinted per source).
  • [ui] Signal-chain output switcher no longer overflows when several outputs are detected. Now that the pipeline lists every real output (USB / optical / analog / HDMI), the output pills in the mobile signal chain stay on a single line, are more compact, and the row scrolls sideways instead of wrapping into oversized cards or getting cut off.
  • [core + ui] Physical output switching now reports honestly instead of always claiming success. /steering/switch-output no longer returns success: true unconditionally — it reflects whether the switch was actually applied (config rewritten + service restarted) and only emits the output_switched event in that case. The Outputs panel now surfaces a failed switch as an error toast (carrying the backend reason) instead of swallowing it, ignores clicks on the already-active output or while a switch is in flight, and always re-fetches the real state afterwards so the UI reflects what actually happened. (Detecting a restart that succeeds but leaves the device busy still needs the native-MPD switch, tracked for a follow-up.)
  • [ui] Library search — the source-filter badges no longer wrap onto multiple rows. On a narrow screen the source chip strip under the search box now stays on a single line and scrolls sideways (scrollbar hidden) instead of wrapping.

[0.9.10] - 2026-07-06

Added

  • [license server] Release-update signal — the LS tells installations when a newer AG release is available (first phase of self-update): the server discovers the latest release from the GitHub releases repo and offers it to the fleet through a new update field on POST /ls/portal/verify ({available, latest, mandatory, notes_url}), gated so it is newer than the device's ag_version and within the licence version_scope. A two-state model separates discovered (polled from GitHub) from published (offered to devices); a diffusion mode (semi-auto — admin publishes manually, default — or auto — published on discovery) is set from the admin UI. New Release Updates admin card: toggle mode, run an on-demand discovery, publish a discovered or specific version (optionally mandatory; an explicit version is validated as major.minor[.patch] and rejected otherwise, so a typo can't become a silent no-op offer), withdraw a bad release, and edit the discovery source (repo + token, stored in the DB overriding the env defaults; token revealed on demand). Discovery uses stdlib urllib (no new runtime dependency) and polls at a low frequency. The release-notes URL offered to devices defaults to the public site RELEASE_NOTES.md (never the private releases-repo page) and is overridable per publish. New admin endpoints under /ls/admin/updates*; new env vars GITHUB_RELEASES_REPO / GITHUB_TOKEN / UPDATE_DISCOVERY_INTERVAL / UPDATE_DISCOVERY_ENABLED / RELEASE_NOTES_URL. The LS is the availability signal only — it downloads and installs nothing (that lands in the AG core in a later phase).
  • [core] the backend surfaces the license server's update signal — GET /license/online-status now carries an update object ({available, latest, mandatory, notes_url}) captured from the 24 h verify check-in, so the UI can show "update available". The core is a pure pass-through: it performs no version comparison and downloads nothing (the actual self-update lands in a later phase).
  • [ui] "Update available" banner + one-click update in the Admin page — when the license server reports a newer release (via online-status.update), the Admin tab shows a banner (ag-update-banner) with the offered version, a required badge for mandatory updates, and a release-notes link. An admin can install it in one click: a confirmation (playback briefly stops) + password prompt triggers POST /sysinfo/actions/update, then the banner shows live progress (downloading → installing → verifying) polled from /sysinfo/update-status, tolerating the core restart, and reloads on success (or reports the automatic rollback on failure).
  • [core + ops] Core self-update engine — POST /sysinfo/actions/update (admin + password) reinstalls the core to a newer release without a terminal. Because the service is KillMode=mixed, the core cannot swap its own running binary in-process, so the updater runs detached in a transient systemd unit (systemd-run, outside the service cgroup) that: snapshots the current binary, re-runs the normal install-core.sh reinstall (stops the service, swaps the binary, restarts — .env preserved), health-checks the new binary on /health (requiring it to report the target version, not merely answer 200, to catch a silent no-swap), and rolls back to the previous binary on any failure. Progress is exposed via GET /sysinfo/update-status (idle → downloading → installing → verifying → done | rolled_back | failed), persisted to disk (world-readable so the unprivileged core can read it) so it survives the restart; a crashed updater no longer wedges the box in "updating" (the in-progress lock goes stale after 15 min). Ships self-update.sh in the package, launched via a root wrapper pinned in sudoers (not a bare systemd-run grant), with the updater scripts and their directory owned by root so the service user cannot tamper with them. An omitted download token falls back to the box's configured RELEASE_DOWNLOAD_TOKEN (private releases repo, Early Access). No OS reboot (playback interruption from the service restart is accepted). The one-click UI trigger lands in a later phase.
  • [ops] Fix: core upgrades now actually replace the binary — the internal installer copied main.dist without removing the previous one first, so on an upgrade cp -r nested it into core/main.dist/main.dist and the old binary kept running. The stale main.dist is now removed before the copy (the service is already stopped at that point). Required for self-update to take effect.
  • [ops] Self-update also updates the co-located UI (mono-host) — after the core binary is swapped and health-checked, the detached updater also reinstalls the UI (install-ui.sh) when it lives on the same machine, so a single "Update now" brings the whole box to the new version. Best-effort: a UI failure does not roll back the core (already verified good) — the app then shows a version-mismatch warning. No-op when the UI is on another host (multi-host).
  • [ops] The UI installer is non-interactive on upgrade — it now resolves the API key, core host/port, and UI mode/port from the existing install (ag-config.js + the ag-ui-server systemd unit) and only asks when a value is unknown and a terminal is present; with no terminal (the detached self-updater) it uses the existing values or fails cleanly instead of hanging on a prompt. This is what lets the self-update actually reinstall the co-located UI. Structured as a flag → existing → prompt → default chain (the flag step is reserved for a future scripted-install feature). A fresh interactive install is unchanged.
  • [ui] Version-mismatch warning — a banner (ag-version-skew-banner) appears in the Admin tab when the interface and the core are on different versions (compared at major.minor from GET /status), e.g. after a partial or multi-host update, prompting the user to update the other component. Silent when the versions match or the core version is unknown.
  • HIGHRESAUDIO (HRA) streaming integration (core + ui): official HRA-Streaming API. New highresaudio module — email/password login, lazy session renewal (re-login when the session expires, no polling), password encrypted at rest (Fernet). Browse (Favorites / Discover), search, and playback of native-master 24-bit FLAC via MPD and UPnP renderers.
    • Core: modules/highresaudio/ (GET/POST/DELETE /highresaudio/connection, GET /highresaudio/stream/{track_id} public proxy); LibraryService browse/search/queue for src_highresaudio; GET /library/highresaudio-discover; now-playing origin + source badge (HRA); virtual source injection.
    • UI: ag-highresaudio-output login-form card in Sources, source labelled Highresaudio in the library/sources list (badge stays HRA), HRA browse pills Favorites / Discover / Editor's Picks / Bestsellers (+ any HRA shop category via /library/highresaudio-category).
    • New reusable core/secret_store.py (Fernet encrypt/decrypt with a local 0600 key) for secrets at rest.
  • Streaming format is always the album's native master (bit-perfect); no quality selector — the HRA API serves master resolution only.
  • Cast the local music library to a UPnP renderer (core): local files are now HTTP-served (signed, Range-capable) so a remote UPnP renderer can pull and play them — POST /library/queue with a local source routes to the active remote renderer via play_queue, exactly like Qobuz / Tidal / HRA. The local DAC and the on-host upmpdcli renderer stay MPD-direct (bit-perfect, no HTTP round-trip). New public, HMAC-signed endpoints GET /library/stream/{path} (audio, HTTP Range/206) and GET /audio_pipeline/library-cover/{path} (album art), reachable without an API key by the renderer. New shared core/library_files.py (library roots resolved from MPD's music_directory, path-traversal guard, Range file serving) + HMAC URL signing in core/secret_store.py.
  • Now-playing source badge on a remote renderer is derived from the stream URI (LIBRARY / QOBUZ / TIDAL / HRA) instead of always showing UPNP.
  • Audio-stack per-service provisioning & guided configuration (core + ui + ops): generate and manage minimal, bit-perfect configs for the audio services (mpd, upmpdcli, shairport), each wired to its own output.
    • Core: POST /audio-stack/provision now always overwrites the config (auto-backing up the existing file first) — the previous only-if-absent policy never wrote anything, since distro packages ship default configs. Both the initial provision and per-service regenerate require the admin password. Each generated config carries a # Managed by Audiogravi<sup>ty</sup> marker; GET /audio-stack/status reports configured from that marker (not mere file existence, which was always true). Outputs are pinned per service (audio-topology.json outputs map) so mpd and airplay can target different DACs (e.g. MPD → USB hi-res, AirPlay → optical). New targeted-patch endpoints POST /audio-stack/output and POST /audio-stack/library change one service's output or mpd's library in place (rewriting only that directive, preserving the rest) — admin, no password.
    • Core: provisioning pins the selected USB DAC's ALSA card index at the OS level — it writes /etc/modprobe.d/ag-audio.conf (options snd_usb_audio index=<current> vid=… pid=…) so the DAC always keeps the same index across reboots / USB re-enumeration. The hw:X,Y baked into the service configs therefore never drifts and no boot-time re-resolution is needed. Written at the DAC's current index (matches the running layout — no reboot), no-op for non-USB outputs (PCI cards are stable), and skipped when the user already pins the device in their own modprobe.d. Reuses the existing privileged-write path (no new sudoers).
    • UI: the config editor gains a Guided mode (default for mpd/upmpdcli/airplay) to change the audio output or music library in a couple of clicks (targeted patch, preserving manual edits) or reset a service to a working default. A first-time-setup flow — a CTA that opens the provisioning panel in a modal — appears only on a new box (no service carries the AG marker) and disappears once set up; each provisionable tile shows a CONFIGURED badge. New reusable ag-prov-output-picker / ag-prov-library-picker molecules + ag-guided-config organism.
    • ops: build-core-package.sh removes the obsolete ag-audio-reresolve.service oneshot from earlier installs on upgrade (the modprobe index pin makes it unnecessary).

Changed

  • [core + ui] Audiogravity's own on-host UPnP renderer is no longer offered as a selectable output. The co-located upmpdcli exists only to receive external UPnP casts and bridges to the same local MPD → same DAC as the Local DAC output, so selecting it was a redundant duplicate. Discovery still surfaces it — shown as a non-selectable "This device · receives external casts" info row for transparency — but AG now refuses a connect to it (400) and, on upgrade, drops it from any persisted active output (falls back to the Local DAC). This removes the ambiguous transport routing a co-located renderer created (it could otherwise hijack a per-source command or resume an idle box). New is_local flag on GET /upnp-renderer/discover; shared core.utils.net.is_local_url.
  • [core] HQPlayer file streaming now honours HTTP Range (206 Partial Content) — the /hqplayer/stream/ endpoint shares the new core/library_files.py file server, which implements real range serving (it previously advertised Accept-Ranges without honouring the header).
  • [core] the streaming renderer-queue paths (Qobuz / Tidal / HIGHRESAUDIO) share a single _stream_queue_renderer skeleton — the local-library cast joins the same base; behaviour-preserving refactor (identical play_queue output, covered by tests).
  • [ui] the local music source is labelled "Local Library" (was "MPD / Music Player Daemon"), with a library icon — "MPD" is the playback engine, not the source. Display only; the internal src_mpd id is unchanged. The MPD service pages (audio config / services / systemd) are untouched.

Fixed

  • [install] a fresh core install no longer aborts before the service is created — the installer's optional VAPID key generation exited non-zero (and imported a non-existent cryptography.hazmat.cores module) on a box without the cryptography lib or a pre-existing vapid_keys.json; under set -e that killed the install before ag-core-server was created (binary placed, service missing). VAPID generation is now guarded and non-fatal (push notifications simply stay disabled).
  • [core] password-gated admin actions (audio-stack provision, license deletion, system reboot) now work when JWT auth is disabled — they re-authenticated by looking up the token's user, which does not exist under the JWT-disabled bypass admin, so they always returned 401. A shared core.jwt_handler.verify_admin_password helper now skips the redundant password check when jwt_enabled is false (the request is already trusted by require_admin); behaviour is unchanged with JWT enabled.
  • [core] a Qobuz / HIGHRESAUDIO track cast to a local MPD output no longer fails (403) once its CDN URL expires — the MPD queue now holds a stable proxy URL that 302-redirects to a freshly-resolved CDN URL on each fetch (MPD follows the redirect), instead of a raw, time-limited CDN URL that would 403 for a track still sitting in the queue past its ~hour TTL (paused, deep in a long queue, or resumed later). AG relays no audio bytes (the 302 sends MPD straight to the CDN) and no per-track CDN resolution happens at enqueue. Tidal already used stable proxy URLs (unchanged).
  • Streaming proxy (Qobuz + HIGHRESAUDIO): pre-signed CDN URLs are now fetched byte-for-byte, so a reserved character in the signed token is never re-encoded (which could cause a 403 and a track failing to play). The Qobuz and HRA proxies now share a single core.http.proxy_cdn_stream helper.
  • [core] player transport controls now route through the UPnP renderer that owns the queue — when a streaming source (Qobuz / Tidal / HIGHRESAUDIO) is cast to a network renderer, next / prev / pause / stop / seek / volume go through the renderer's AVTransport (advance_queue / pause / …) instead of driving the underlying MPD directly. Local MPD-library playback is unchanged: with an empty renderer queue the renderer call is skipped and control falls back to MPD.

[0.9.9] - 2026-06-30

Changed

  • [ops] Deployment nomenclature renamed frontend→ui and backend→core — build/packaging scripts (build-core.sh, build-core-package.sh, build-ui-package.sh, dev-core.sh), release assets (audiogravity-core-<ver>-<arch>.tar.gz, audiogravity-ui-<ver>.tar.gz) and public install scripts (install-core.sh, install-ui.sh) now use the new names. Old release assets from past versions keep their original names.
  • [core] systemd unit ag-backend-server → ag-core-server — data directory /opt/audiogravity/backend → /opt/audiogravity/core (AUDIOGRAVITY_HOME). The service-metrics map key exposed over SSE renamed backend → core. The sudo systemctl restart ag-core-server path is covered by the existing generic systemctl sudoers rule (no new alias needed).
  • [ui] systemd unit ag-frontend-server → ag-ui-server — deploy directory /var/www/audiogravity-frontend → /var/www/audiogravity-ui; version constant FRONTEND_VERSION → UI_VERSION; dev proxy env var BACKEND_PORT → CORE_PORT (dev.sh + vite.config.js).

Added

  • [ops] scripts/migrate-deploy-layout.sh — standalone, idempotent migration for hosts on the old layout. Run once as root before reinstalling: it backs up the existing layout (tar), stops/removes the old units, renames /opt/audiogravity/backend→/opt/audiogravity/core and /var/www/audiogravity-frontend→/var/www/audiogravity-ui, and rewrites AUDIOGRAVITY_HOME — preserving .env, secrets, users and tokens. Works on single-host and split-host topologies (migrates only the component present locally). /etc/audiogravity is left untouched.

Fixed

  • [core] UPnP renderer — auto-reconnect race condition — after connect() completes, the manager checks whether the renderer was removed during the async call and tears it down cleanly if so; prevents a connected/disconnecting oscillation when a renderer is removed while a reconnect is in progress.
  • [core] player — active UDN not persisted when MPD output selection clears it — select_mpd_output now calls _save_config() in the else branch so the cleared active UDN survives a restart.
  • [core] track_number missing for Qobuz, Tidal and MinimServer streams — now_playing.py reads track_number from ext_stream_meta even when MPD already has a title; library/service.py registers track_number for all Qobuz, Tidal and UPnP album queue paths; library/upnp_service.py extracts upnp:originalTrackNumber from DIDL-Lite with a safe int() guard against empty strings.
  • [ui] fullscreen player — source badge moved onto cover art — the origin badge (❖ QOBUZ / ❖ TIDAL / ❖ UPNP) now appears top-left on the cover art; the track badge (A1 · TRACK 01) appears bottom-left. Duplicate source badge below the cover removed; • MPD fallback label removed.
  • [ui] output panel — wording — section label "UPnP Renderer" → "Audio Output"; button "Scan network" → "Scan renderers"; "Refresh" → "Reload".

[0.9.8] - 2026-06-29

Added

  • [core] UPnP renderer — DSD detection — RendererStatus.format is set to "DSD" when the current URI has a .dsf or .dff extension (MinimServer DSD tracks). PlayerState.format is populated for native renderers so the fullscreen player shows the DSD lock and hides the volume control automatically via the existing isDsd() check.
  • [core] RendererStatus.renderer_udn — new field on RendererStatus mirroring the value already injected into the SSE payload; the UI now has the UDN from the HTTP GET response without waiting for the first SSE event.
  • [core] UPnPRendererService.reachable / .uses_local_mpd properties — expose the two previously-private attributes as public properties; get_known() and _try_renderer_control() use these properties instead of accessing _reachable / _uses_local_mpd directly.
  • [core] GET /player/origins — new endpoint returning the canonical origin → label map (ORIGIN_LABELS). Clients call it at startup and merge the result into their static fallback, making the backend the single source of truth for origin display labels.
  • [core] signal path — enriched real-time chain display — _build_state() builds the full audio chain dynamically: Source → [Renderer] → [MPD] → [Connector] → DAC. Connector (USB / TOSLINK) inserted before the DAC. UPnP renderer prepended when active. Audio source (Qobuz, Tidal, Radio, Library…) prepended from origin. New uses_local_mpd: bool field in RendererStatus distinguishes upmpdcli (bridges to local MPD) from native network renderers (Marantz, Linn…) — for those, the signal path is Source → Renderer only.
  • [ui] fullscreen player — signal path replaces separate renderer badge — the signal path chain (e.g. • Qobuz → • music.#1 → • MPD → • USB → • Heed Abacus) is now built entirely by the backend and rendered as individual steps; the previously separate → music.#1 overlay badge has been removed.
  • [ui] mini player — renderer badge in source row — when a UPnP renderer is active, a → renderer_name badge appears in the source row; the USB/TOSLINK connector badge is hidden in that state (already visible in the fullscreen signal path).
  • [core] Qobuz — streaming proxy (GET /qobuz/stream/{track_id}) — stable, never-expiring URL for UPnP renderers; resolves HMAC-signed CDN URL just-in-time on each renderer fetch. Endpoint is public (no api_key required). Pass-through FLAC, no transcoding, same design as the existing Tidal proxy.
  • [core] UPnP renderer — NEXT / PREV — POST /upnp-renderer/next and /prev skip to the next or previous track in the active renderer queue. Returns 409 if no queue is active, already at boundary, or a transition is in progress. Qobuz proxy URLs are stable so PREV can re-request an already-played track without re-signing.
  • [ui] fullscreen player — NEXT / PREV routed to renderer — when a renderer queue is active, the NEXT / PREV transport buttons route through POST /player/control; can-next / can-prev are derived from queue position so buttons are correctly disabled at boundaries.
  • [core] UPnP renderer — multi-track queue playback — play_queue() sends a full album to the renderer via SetAVTransportURI + SetNextAVTransportURI (gapless pre-loading where supported). Tracks are chained automatically: end-of-track detection via NOTIFY polling, _advance_queue() re-resolves Qobuz HMAC-signed CDN URLs just-in-time. Supports Qobuz, Tidal and UPnP/MinimServer albums.
  • [core] UPnP renderer — RendererStatus queue fields — queue_position (0-based index), queue_total, queue_next_title, queue_next_artist, queue_next_album, queue_next_cover_token exposed in GET /upnp-renderer/status and renderer_status SSE event.
  • [ui] fullscreen player — "Up next" for renderer queue — when a renderer queue is active, the "Up next" strip below the controls shows the next track's title, artist and cover art, updated in real time via the renderer_status SSE event (no MPD queue call needed).
  • [core] UPnP renderer — RendererManager multi-renderer pool — replaces the singleton UPnPRendererService. Manages a dict[udn → UPnPRendererService]; exactly one renderer is active at a time (policy, not architecture). Config persisted in upnp_renderer_known.json; migrates silently from the legacy single-renderer format. Exposes GET /upnp-renderer/known, get_active_service(), connect(renderer), disconnect(udn), discover(). Auto-reconnect background task with exponential backoff (30 s → 5 min).
  • [core] UPnP renderer — UDN-based route architecture — all renderer routes reworked to /upnp-renderer/{udn}/action. Manager-level: GET /discover, GET /known. Per-renderer: GET|PUT|DELETE /{udn}/connection, GET /{udn}/status, POST /{udn}/play|next|prev|stop|pause|seek, PUT /{udn}/volume, POST /{udn}/notify. Route PUT /bypass removed.
  • [core] UPnP renderer — resume() method — resumes playback from PAUSED or STOPPED without loading a new URI. Called by the native-renderer transport fallback when Play is pressed and the renderer already has a track loaded.
  • [core] player — native renderer transport fallback — _try_renderer_control() dispatches toggle/next/prev/stop/seek to a non-local-MPD renderer (ARM UpMPd, Marantz, Linn…) when get_now_playing() returns nothing for that renderer. Wired in both control() paths: explicit source_id branch and auto-resolve branch.
  • [core] GET /player/outputs — unified output list: one entry per MPD audio_output config block (type: "mpd_output", real device name from MPD outputs command) + all known UPnP renderers (type: "upnp_renderer"), with live reachable and active flags. Falls back to a single "Local DAC" entry when MPD is unreachable.
  • [core] PUT /player/mpd-output/{output_id} — enables one MPD audio output exclusively (all others disabled) and disconnects any active UPnP renderer. output_id is the MPD outputid integer. Returns 503 when MPD is unreachable.
  • [core] player — native renderer state when local pipeline is idle — _build_native_renderer_state() provides title, artist, album, cover art, transport state, output_label, and signal_path when a non-local-MPD renderer is active but get_now_playing() returns nothing. Fixes "Nothing playing" and "No output selected" with a native ARM renderer.
  • [core] RendererStatus — cover_token and can_seek fields — cover_token exposes the current track's cover art token so the fullscreen player can render cover art for a native renderer; can_seek exposes AVTransport seek capability for the playback controls.
  • [core] Push notifications — temperature and service-down alerts — notify_temperature_alert() fires when CPU exceeds 85°C (15-min cooldown); notify_service_down() fires on active → failed systemd state transition (stale states pruned each cycle to prevent false positives on re-registration).
  • [core] radio → renderer routing — RadioService.play() now routes to the active UPnP renderer via play_queue() (AVTransport) when one is connected. Falls back to local MPD only when no renderer is active.
  • [core] DELETE /upnp-renderer/{udn} — permanently removes a renderer from the known list (disconnects if active, clears from upnp_renderer_known.json). Distinct from DELETE /{udn}/connection (disconnect only, keeps in list).
  • [ui] UPnP renderer card — swipe-to-delete — left-swipe gesture on an inactive renderer row reveals a red "Remove" zone; completing the swipe (≥ 140 px) calls DELETE /upnp-renderer/{udn} and removes the row. Uses the Pointer Events API (same pattern as ag-radio-card.js). The active output row is not swipeable.
  • [ui] PWA — App Shell precaching — vite-plugin-pwa (injectManifest) precaches all Vite-hashed JS/CSS/image assets at SW install (~1 MiB); first load after install is fully offline-capable on Chrome/Android.
  • [ui] PWA — differentiated cache strategies — cache-first for hashed Vite assets and version-pinned CDN (cdn.jsdelivr.net); stale-while-revalidate for Google Fonts and static images; network-first for HTML navigation.
  • [ui] PWA — install prompt (Android/Chrome) — compact banner when Chrome offers installation (beforeinstallprompt); dismissal persisted 30 days via localStorage; complete rewrite of pwa-install-prompt.js (English, CSS tokens, no duplicate SW handler).
  • [ui] PWA — offline UI — last known player state persisted to localStorage (5 s debounce); cold offline reload restores the mini-player instead of showing an empty screen; discrete Offline badge in the source row.
  • [ui] PWA — Link: rel=preload response headers — serve_https.py emits preload hints for critical JS/CSS chunks so the browser fetches them in parallel without waiting for the HTML parser.
  • [ui] PWA — apple-mobile-web-app-title — short label for the iOS home screen icon.
  • [ui] build — granular bundle splitting — main chunk reduced from 570 KB to 413 KB (−27%); 6 stable independent chunks: lit, icons, atoms, nowplaying, streaming, library-core.

Changed

  • [core] Steering — renderer pre-disconnect — switch_output() now disconnects a local renderer (uses_local_mpd=True, e.g. upmpdcli) before switching the ALSA device. Native/distant renderers (uses_local_mpd=False) are left running — their audio stack is independent of AG's ALSA configuration.
  • [core] GET /player/outputs — local_active during reconnect window — MPD outputs now show active: true when the persisted renderer is selected but not yet reachable (auto-reconnect still pending at startup), matching actual audio routing.
  • [core] UPnP NOTIFY after disconnect — orphaned NOTIFY requests (arriving after disconnect() but within the UPnP subscription TTL) now return HTTP 200 instead of 404, preventing upmpdcli from treating the subscription as lost.
  • [core] _try_renderer_control() — uses_local_mpd short-circuit — uses_local_mpd is now checked via the new property before calling get_status() (no await); get_status() is only called for the toggle action which needs transport_state.
  • [core] select_mpd_output — stale _active_udn guard — if _active_udn is set from persisted config but the service is not yet in _services (auto-reconnect pending), _active_udn is cleared instead of calling disconnect() (which would raise KeyError).
  • [ui] UPnP renderer card — reworked as multi-output selector — ag-upnp-renderer-card is now the unified output selector, loading from GET /player/outputs: one row per physical MPD audio output (USB, TOSLINK…) then all known UPnP renderers. Clicking a physical output calls PUT /player/mpd-output/{output_id} (enables exclusively, auto-disconnects renderer); clicking a renderer calls PUT /{udn}/connection. Active renderer shows Disconnect button + volume slider. States: Active (green) / Reconnecting (orange) / Idle (grey). SSE sync updates active/reachable in the list without full reload.
  • [core/ui] UPnP renderer — bypass mode removed — replaced by the output manager pattern: selecting Local DAC disconnects the renderer; re-selecting the renderer reconnects it. Endpoint PUT /upnp-renderer/bypass removed. Field bypassed removed from RendererStatus.
  • [ui] fullscreen/mini player — renderer references updated — bypass conditions removed; routes updated to UDN-based API.

Fixed

  • [ui] mini player — cover art popover broken image — the detail popover (np-detail-cover) now applies the same _brokenCovers check as the thumbnail; images that failed to load (404 / network error) are skipped in the popover instead of showing a broken image icon. The @error handler is also added on the popover <img> to capture failures that occur before the thumbnail has had a chance to error.
  • [core] UPnP renderer — connect() UDN mismatch — POST /upnp-renderer/{udn}/connection was looking up the service by the path parameter instead of the request body udn, causing a KeyError (HTTP 500) when they differed.
  • [core] player — PUT /player/mpd-output/{output_id} — unknown output_id now returns 404 instead of silently disabling all outputs; existence check runs before the batch MPD command.
  • [core] UPnP renderer — queue index underflow — concurrent play_queue() + _advance_queue() could decrement _queue_idx to −1 on rollback; guard now requires _queue_idx > 0 before decrementing.
  • [core] library — Qobuz single-track to renderer — qobuz_queue() now calls play_queue([QueueEntry(…)]) instead of the removed play() method; cover token forwarded.
  • [core] sysinfo — temperature alert push spam — cooldown timestamp is advanced before the push attempt so a misconfigured push service cannot trigger repeated alerts on the same event.
  • [core] Qobuz proxy — shared HTTP session — GET /qobuz/stream/{track_id} reuses the shared aiohttp.ClientSession pool; per-request headers forwarded on .get() call, no session closed between requests.
  • [core] player — native renderer detection — _try_renderer_control() and _build_native_renderer_state() use the public status.uses_local_mpd field from RendererStatus instead of the private attribute.
  • [ui] renderer card — SSE unknown renderer — when a renderer_status SSE event arrives with a renderer_udn not yet in the known list, the card now triggers a reload instead of silently discarding the update.
  • [ui] library sources — CSS tokens — replaced hardcoded 10px with var(--spacing-xs), removed duplicate user-select: none rule, documented intentional #ffffff on error background.
  • [ui] UPnP renderer card — double "Active" badge via SSE path — _onStatusEvent did not clear active on the previously active renderer when a new one received connected: true. Fixed: when isNowActive=true, all non-target renderers in _known now force active: false.
  • [core] UPnP renderer — disconnect() removed renderer from known list — disconnect() filtered the renderer out of _services pool AND from _known, making it disappear from the card after disconnect. Fixed: only _active_udn is cleared; the renderer stays in _known so it can be re-selected without re-scanning.
  • [ui] UPnP renderer card — list not refreshed after backend restart — the backend publishes a renderer_status SSE event on startup before the frontend SSE stream reconnects, so the event was missed. Fixed: the card now listens to window.EventEmitter 'connection-status' and calls _load() on every SSE reconnection.
  • [core] player — 503 on Play with native (ARM) renderer — when the fullscreen player sent source_id explicitly, control() found nothing in get_now_playing() and returned 503 without reaching the renderer fallback. Fixed: _try_renderer_control() fallback now wired in both branches of control().
  • [core] UPnP renderer — _publish_status(force=True) didn't update the hash baseline — on forced publishes, _last_sse_hash was not updated because the hash computation lived inside the if not force: block. Hash is now always computed and _last_sse_hash always updated; only the early-return dedup check is gated on not force.
  • [core] UPnP renderer — _resubscribe_and_refresh() didn't restore reachability after SID recovery — _handle_update_result(success=True) was not called after a successful async_update() post-resubscribe; _reachable stayed False and the renderer card showed "offline" for up to 30 s after the SID mismatch was resolved.
  • [core] UPnP renderer — _retreat_queue pre-decrement unguarded against stop() race — stop() can clear _queue and reset _queue_idx = 0 between the public retreat_queue() guard check and the unconditional _queue_idx -= 1 inside _retreat_queue(). Guard added before the decrement.
  • [core] player — renderer included in signal path during startup reconnect window — _build_state() used only the connected flag; guard tightened to require connected AND reachable.
  • [core] library — asyncio.TimeoutError not caught in MPD queue-list endpoints — a stalled MPD during restart raised asyncio.TimeoutError which bypassed the (OSError, ConnectionRefused, BrokenPipe) empty-list handler and returned HTTP 500.
  • [ui] fullscreen player — "Up next" strip showed stale track after renderer disconnect — _nextTrack was only cleared when the renderer sent queue_next_title=null while still connected. Cleared explicitly on disconnect.
  • [ui] fullscreen player — cover error token not reset between consecutive album tracks — _coverErrorToken was only cleared on page reload or full source reconnect. Token is now reset on every track-title or source change.
  • [ui] fullscreen player — idle renderer badge missing when nothing is playing — the source row was conditioned on origin || source_name || hasSignal, all of which are falsy when the renderer is idle. Condition extended to include _rendererActive.
  • [ui] mini player — connector badge hidden for upmpdcli (local-MPD renderer) — the badge was unconditionally hidden when any renderer was active. upmpdcli routes audio through the local MPD stack, so the physical connector IS in the chain. Visibility now also passes when _rendererStatus?.uses_local_mpd is true.
  • [core] signal path — radio source shows "Radio" not the station name — origin_name for radio is content metadata (station name), not a chain step. Canonical label now comes from the ORIGIN_LABELS map keyed on origin.
  • [core] UPnP renderer — _retreat_queue rollback unsafe when stop() fires mid-retreat — mirror of the earlier _advance_queue fix: guard if self._queue: added before rollback.
  • [core] UPnP renderer — _uses_local_mpd stale after disconnect or failed reconnect — disconnect() now resets _uses_local_mpd = True; get_status() re-derives it from the saved location URL when _dmr is None via the extracted _is_local_renderer() helper.
  • [core] player — renderer status error swallowed silently in _build_state() — bare except Exception: pass replaced with logger.debug(..., exc_info=True).
  • [core] Roon — suppress ERROR logs when roonbridge service is stopped — connect() now gates on ActiveState via the existing D-Bus client before creating RoonApi; the daemon threads that emit ERROR: Connection is not (yet) ready! are never started.
  • [core] Roon — TCP pre-check before mDNS discovery — _sync_connect() now verifies the configured host is reachable (port 9330) before running RoonDiscovery.first(), which can block for several seconds when Roon Core is offline.
  • [core] Qobuz proxy — Range headers relayed end-to-end — GET /qobuz/stream/{track_id} now forwards the Range request header to the CDN and relays Content-Range, Content-Length and Accept-Ranges response headers with the correct status code (206 Partial Content).
  • [core] UPnP renderer — _check_queue_advance stacked create_task on rapid NOTIFYs — fixed by storing the task reference in _advance_task and skipping if the previous task is still running.
  • [core] UPnP renderer — _ext_stream_key missed /qobuz/stream/ proxy URLs — Qobuz proxy URLs (path /qobuz/stream/{id}) fell through to the raw URL fallback, causing a spurious AVTransport art fetch. Added a path-anchored check before the CDN ?eid= query-param check.
  • [core] UPnP renderer — _advance_queue rollback unsafe when stop() fires mid-advance — rollback now only applies when the queue is still populated.
  • [core] UPnP renderer — _publish_status(force=True) computed JSON hash unnecessarily — hash computation now skipped in the force path.
  • [core] UPnP renderer — renderer pinged every 30 s in SUBSCRIBE mode — ping interval extended to 60 s while the SSE heartbeat remains 30 s, halving outbound HTTP GETs to the renderer.
  • [core] Qobuz — ClientSession created per API call — get_stream_url(), handle_callback(), and validate_token() migrated to get_shared_session() from core.http.
  • [core] UPnP renderer — NOTIFY SID mismatch after AG restart — handle_notify() now detects 412, schedules _resubscribe_and_refresh() — a background task that unsubscribes, re-subscribes (fresh SID), and calls async_update() immediately so badge and transport state are accurate without waiting for the heartbeat.
  • [core] Qobuz — single-track play to renderer used expiring CDN URL — qobuz_queue() for a single track now uses the stable proxy URL like the multi-track album path.
  • [core] Qobuz proxy — CDN stall leaked aiohttp session indefinitely — ClientTimeout(total=None) meant a stalled CDN connection held the session open forever. Fixed: sock_read=300.
  • [core] Qobuz proxy — network errors returned HTTP 500 instead of 503 — except RuntimeError did not catch aiohttp.ClientConnectorError / asyncio.TimeoutError. Fixed: except Exception.
  • [core] UPnP renderer — NEXT / PREV returned ok:True when transition already in progress — public methods now raise RuntimeError("Queue transition in progress") → router returns 409.
  • [core] UPnP renderer — spurious URI-change detection after explicit play — _prev_track_uri was not updated after _advance_queue() started a track explicitly; the incoming PLAYING NOTIFY was misidentified as a natural transition, double-advancing the queue. Fixed by anchoring _prev_track_uri to the URI just played.
  • [core/ui] UPnP renderer — reachability tracking — renderer card now shows 🟠 Unreachable when the device goes offline (2 consecutive heartbeat failures) and recovers to 🟢 Connected automatically. Field reachable added to RendererStatus.
  • [core] UPnP renderer — auto-reconnect retry loop — on backend startup, auto-reconnect now retries with exponential backoff (30 s → 60 s → … → 5 min ceiling) instead of giving up after one attempt.
  • [ui] UPnP renderer — renderer_status SSE event missing from worker — sse-worker.js did not register a listener for renderer_status events, so the renderer card never received real-time status updates.
  • [core] library — MPD queue/albums return empty when MPD is stopped — GET /library/queue and GET /library/albums now return an empty list instead of HTTP 500 when MPD is unreachable.
  • [core] UPnP renderer — disconnect now stops playback — disconnect() calls AVTransport Stop before unsubscribing so the renderer stops playing immediately.
  • [ui] now-playing — radio cover art fallback — mini-player shows the music-note placeholder instead of the broken image icon when the cover proxy returns 404; fullscreen player uses a hidden probe <img> to detect silent CSS background-image 404s and shows the placeholder.
  • [ui] PWA — skipWaiting() removed from SW install event — prevented chunk 404s on update (new SW was taking control before the page reloaded with new asset hashes).
  • [ui] PWA — SW update reload loop — controllerchange triggers a single conditional reload guarded by sessionStorage; toast correctly says "Updating…".
  • [ui] PWA — SWR background refresh lifetime — fetchPromise now covered by event.waitUntil() so the SW is not killed before cache.put() completes.
  • [ui] PWA — isCDN too broad — replaced url.origin !== location.origin with explicit CDN_SWR / CDN_IMMUTABLE Sets; mutable cross-origin resources no longer routed into stale-while-revalidate.
  • [ui] PWA — dead icomoon.woff path — removed non-existent /fonts/icomoon.woff?1zo0jr from CACHE_URLS.
  • [ui] PWA manifest — broken screenshot references — removed screenshots field pointing to non-existent files (silent 404, degraded Chrome install dialog).
  • [ui] build — @lit/context chunk assignment — @lit/context and @lit/reactive-element now land in the stable lit chunk instead of main.
  • [ui] push notifications — push-manager.js — replaced raw fetch() with apiGet/apiPost for consistency; PushSubscription destructured explicitly via toJSON() before posting.

Refactored

  • [core] mpd_client.fetch_outputs() — MPD audio output parsing (previously _fetch_mpd_outputs() inline in modules/player/router.py) moved to core/mpd_client.py as a public fetch_outputs(pipeline) function. Router now calls mpd_client.fetch_outputs(). The router no longer parses MPD wire format directly; the helper is independently testable.
  • [core] _try_renderer_control() — single dispatch block — removed the two-phase approach (separate if action == "toggle" to fetch transport state, then a second if/elif dispatch); merged into one try block. transport_state intermediate variable eliminated.
  • [core] LibraryService.is_mpd_source() — new public method replacing direct calls to the private _mpd_port() from modules/library/router.py. Two call-sites in the router updated.
  • [ui] fetchActiveRendererStatus() — the known → active UDN → status bootstrap chain, previously copy-pasted identically in ag-now-playing.js and ag-now-playing-fullscreen.js, extracted into library-store.js as a shared async helper.
  • [ui] fullscreen player — renderer transport routing — _control() now sends all transport actions (including next/prev) through POST /player/control. The UI-side shortcut that routed renderer queue next/prev directly to POST /upnp-renderer/{udn}/next|prev has been removed.
  • [core] PlayerService._get_active_renderer_svc() — new private helper centralises the get_container().get(RendererManager).get_active_service() lookup, previously repeated verbatim in three methods.
  • [core] PlayerService._renderer_queue_nav() — new static helper returns (can_next, can_prev) from a RendererStatus, replacing duplicated queue_total is not None and queue_position is not None guards.
  • [core] _save_config() dict comprehension — simplified from {**{k: v for k, v in r.items() if k != "active"}, ...} to {**r, "active": ...}.
  • [ui] Remove stale !bypassed guards — bypassed was removed from RendererStatus; all six !data.bypassed / !rs.bypassed guards in ag-now-playing.js and ag-now-playing-fullscreen.js replaced with the simpler connected check.
  • [core] _upsert_known() — renderer dict built once before the if/else branch instead of duplicated in each arm.
  • [core] _make_stream_on_play() — extracted shared static method replacing _make_qobuz_on_play / _make_tidal_on_play copy-paste closures in library/service.py.
  • [core] service monitoring — set construction — replaced set(dict) - set(dict) (two allocations per 10 s tick) with a list comprehension over the dict keys.
  • [ui] _rendererCanNext / _rendererCanPrev getters — replaced inline IIFEs in ag-now-playing-fullscreen.js template with named getters; return null when no renderer queue is active so the template falls back to player state.

[0.9.7] - 2026-06-26

Fixed

  • [ui] library — radio form font harmonization — type="url" inputs changed to type="text" + inputmode="url" so Safari/iOS no longer renders the value in monospace; ::placeholder inherits typography from the input element; image_url field now validates with pattern="https?://.+".
  • [ui] library — "Add custom station" button border — dashed border changed to solid.
  • [ui] library — genre/country select font size — increased from font-size-xs to font-size-sm for readability in the native dropdown.
  • [ui] sources — redundant source descriptions removed — Roon, Tidal, Qobuz source cards no longer repeat the source name as description; MPD keeps "Music Player Daemon"; Roon active source always shows at least the source name while zones are loading.
  • [ui] sources — Qobuz and Tidal merged under "Streaming Services" — reduces visual noise; service name shown in each molecule's own header.
  • [ui] sources — lib-settings-section CSS class — decoupled from lib-hqp-section so HQPlayer-specific style changes cannot bleed into other sections.

Added

  • [core] upnp_renderer — UPnP Control Point — AG can now send audio to any UPnP/DLNA MediaRenderer on the network. New module modules/upnp_renderer/: SSDP discovery, AVTransport control (SetAVTransportURI, Play, Stop, Pause, Seek, Volume), SUBSCRIBE/NOTIFY eventing via AG's own FastAPI callback, connection persistence across restarts. Auto-reconnects to the last renderer on startup.
  • [core] upnp_renderer — renderer dispatch in library queue — upnp_play(), qobuz_queue() and tidal_queue() route to the connected renderer when action='play'. MinimServer tracks use URI handoff (renderer pulls directly from CDN/MinimServer). Qobuz uses self-authenticated CDN URL (no proxy). Tidal uses the existing DASH→FLAC proxy with LAN-reachable IP.
  • [core] upnp_renderer — SSE renderer_status — live renderer state published on the dashboard SSE channel on every NOTIFY event, with a 30 s heartbeat fallback for renderers that don't SUBSCRIBE reliably.
  • [core] library — GET /library/upnp-browse and GET /library/search now use location parameter — replaces control_url; location is the device description URL used by async-upnp-client. Breaking change for UI clients.
  • [core] library — upnp_service.py rewritten with async-upnp-client — replaces hand-rolled SOAP/SSDP with async-upnp-client (python-didl-lite, defusedxml). Adds DmsDevice profile for ContentDirectory Browse/Search with proper DIDL-Lite parsing.
  • [core] /tidal/stream/ added to PUBLIC_PATH_PREFIXES — UPnP renderers can fetch the Tidal proxy without api_key in the URL (avoids key exposure in plaintext AVTransport traffic).
  • [core] core/utils/net.py — shared get_local_ip() helper (cached, replaces 3 duplicates).
  • [ui] ag-upnp-renderer-card — new molecule in the Sources panel: discovery, connect/disconnect, live playback status, Play/Pause/Stop/Volume controls.
  • [ui] renderer routing badge — → music.#1 badge in mini player and fullscreen player when a renderer is connected.
  • [ui] subscribeRendererStatus — shared SSE subscription in library-store.js (replaces 3 independent window.addEventListener calls).
  • [ui] library-api.js upnpPlay — routes MinimServer browser plays through /library/upnp-play which now dispatches to renderer when connected.

Changed

  • [core] async-upnp-client added to requirements.txt (pure Python, none-any wheels, Nuitka-compatible).
  • [core] upnp-browse param renamed — control_url → location in router and service.

[0.9.6] - 2026-06-25

Fixed

  • [core] audio_pipeline — cover art missing or wrong when playing via upmpdcli renderer — when an external UPnP control point (BubbleUPnP, Kazoo…) pushed music or radio to upmpdcli, AG fell back to MusicBrainz/iTunes which returned the wrong album cover or nothing. A new upmpdcli_cover module queries upmpdcli's AVTransport GetPositionInfo to retrieve the exact upnp:albumArtURI the controller sent (including radio station logos). Discovery is restricted to the local machine to avoid picking up network renderers (Marantz, etc.). Tidal and Qobuz proxy streams are excluded from the AVTransport path. Miss backoff (60 s TTL) and a provisional sentinel prevent SSDP storms on concurrent SSE ticks.
  • [core] hqplayer — WARNING flood at startup when HQPlayer not yet configured — get_status() attempted a TCP connection even when _host was None (not yet configured via UI), generating a warning on every poll. Guard added: returns HQPlayerStatus(available=False) immediately without attempting connection. Regression from 0409bb5.
  • [core] hqplayer — WARNING flood during DSD playback startup — _fetch_status() on TCP error did not cache the failure result, causing immediate retries on every poll cycle instead of respecting the 2 s TTL backoff. Failure path now caches HQPlayerStatus(available=False). Regression from c230b1b.
  • [core] hqplayer — naa_available field added to /hqplayer/connection — the connection response now includes whether the local networkaudiod service is active, queried via D-Bus (no subprocess). ServicesManager.is_service_active() added as the canonical D-Bus ActiveState helper.
  • [ui] hqplayer — connected state reflects full chain (HQPlayer + NAA) — the card now shows "Connected" only when both HQPlayer is reachable and networkaudiod is active. Shows "NAA offline" when HQPlayer is reachable but NAA is down. "Use as output" toggle hidden when NAA is offline. Toggle flag cleared automatically via updated() lifecycle when NAA availability drops, guarded against transient fetch failures (=== false not falsy). naa_available updated in real-time via service-metrics-sse SSE event (serviceId "hqplayer", no round-trip to /hqplayer/connection).
  • [core] services — is_service_active() handles absent ActiveState key — D-Bus response dict uses a filter comprehension; if both interfaces fail the key is absent. Changed props.get("ActiveState") == "active" to props.get("ActiveState", "inactive") == "active" to avoid false-inactive on empty response.
  • [core] hqplayer — _naa_available() logs D-Bus failures at DEBUG — bare except Exception: return False was silent; now logs the exception before returning False. Service name extracted to _NAA_SERVICE constant.
  • [ui] ag-tabs — bell animation missing on iOS — CSS transform animation applied directly on an <svg> element is not supported by iOS WebKit; the bell icon is now wrapped in a <span> that carries the animation.
  • [ui] ag-tabs — admin tab stat reverted to connected/total ratio — the previous change showed only the total account count, losing the connected context; reverted to num/den format (active sessions / total accounts).
  • [ui] beta badge missing in production — .env.production was absent from the build host; VITE_BETA=true was not substituted by Vite at build time, causing the badge to be tree-shaken out of the bundle.
  • [ui] ag-tabs — bell animation suppressed on iOS when OS Reduce Motion is active — the bell had no prefers-reduced-motion override; added explicit restore of the animation when the app animations toggle is ON, following the same pattern as status dots. Added will-change: transform for GPU compositing.
  • [ui] mini-player / fullscreen — active source not followed automatically when multiple sources play — both players now auto-follow the backend-active source by default; the user can override by tapping a dot or swiping, and the override is lifted when the chosen source stops playing.
  • [ui] mini-player — dot click did not suspend auto-follow — tapping a dot set _activeSourceIdx without setting _userSourceOverride, so the next SSE tick silently reverted the selection to the auto-followed source.
  • [ui] mini-player — prevShownId read from new items after clamp — the override-lift check identified the wrong source when the user's chosen source disappeared and the item list shrank, causing the override to stay active indefinitely. Fixed by capturing prevShownId before this._items is replaced.
  • [ui] fullscreen — player stuck on stopped source after override lift — when the override was lifted on the same SSE tick as state.playing=false (dead source's final tick), _followSource was never called because the state.playing && guard failed. Now immediately follows _sources.find(s => s.playing) when the override is lifted.
  • [ui] mini-player — apiGet missing from import — apiGet was called in _fetchAlbumTracks but only apiPost was imported, causing a ReferenceError when the album detail popover was opened.
  • [core] player — bitrate no longer displayed for ALAC, FLAC, TIDAL and Roon sources — a _LOSSLESS_FILE filter introduced in the Qobuz commit (78ef364) combined with a codec-detection fix (828ee91) silently hid the bitrate for all lossless and streaming sources. The filter is removed; MPD's reported bitrate is now shown for all MPD sources (ALAC, FLAC, WAV, streaming). For sources with no MPD bitrate (Roon, AirPlay), a PCM-equivalent is computed from bit-depth × sample-rate × 2 ch.
  • [core] player — TIDAL bitrate always 0 — MPD reports bitrate: 0 for the growing FLAC proxy stream (STREAMINFO total_samples=0 during progressive remux). The DASH manifest's bandwidth attribute is now parsed at stream-request time and injected into the format metadata via ext_stream_meta, providing the exact bitrate immediately.
  • [core] player — TIDAL format strip showed — for format and sample rate during warm-up — sample_rate_hz and codec from the DASH manifest were stored in ext_stream_meta but never consumed. audio_pipeline/service.py now builds a partial source_format ("FLAC | 96kHz") from the manifest when MPD's audio field is not yet available, so sample rate and codec appear immediately.
  • [core] audio_pipeline — ext_stream_meta LRU contract broken on merge=True — in-place dict update did not reposition the entry; merged entries could be evicted before genuinely older ones. Fixed with delete-then-reinsert.
  • [core] tidal — parse_dash_format broad except could silently drop bitrate_kbps — a ValueError on audioSamplingRate would discard a successfully parsed bitrate_kbps. Each field conversion now has its own try/except; failures are logged at DEBUG level and do not affect other fields.

Added

  • [core+lic] announcements — broadcast polling — license server admins can create broadcast announcements (type version/promo/alert/info, optional body/URL/expiry). Active announcements are delivered to AG instances via the existing 24 h /verify check-in. A delivery count is shown in the LS admin panel per announcement. AG admin tab shows a Lucide Bell icon (warning color, ring animation) when unread announcements exist; dismissal persists in localStorage.
  • [lic] ls-announcements — admin panel section — new ls-announcements molecule in the LS admin UI with a creation form (type, title, body, URL, expires_at) and a management table (activate/deactivate/delete, delivered count).
  • [ui] ag-announcement-banner — new light-DOM molecule displaying dismissable banners in the AG Admin tab. Fetches GET /license/online-status.
  • [ui] ag-tabs — Admin tab badge — animated Lucide Bell icon (--color-warning) on the Admin tab when unread announcements exist. Animation follows the global animations toggle. Admin tab stat now shows total account count instead of the meaningless active/total ratio.
  • [lic] campaign mailing — LS admins can compose and send HTML campaign emails to all active licence holders. Features: visual preview iframe (sandboxed, no test send needed), ls_base_url persisted in settings, concurrent SMTP (Semaphore(5)), campaign history log with recipient count.
  • [lic] GDPR unsubscribe — stateless HMAC-signed opt-out link injected in every campaign email (List-Unsubscribe header + footer). GET /ls/portal/unsubscribe sets email_opt_out=1 without login. UNSUBSCRIBE_SECRET separate from ADMIN_TOKEN so token rotation doesn't invalidate past links.
  • [lic] ls-mailing — admin panel section — new ls-mailing molecule: compose form, show/hide preview iframe, send to all button with confirmation, campaign history table.

[0.9.5] - 2026-06-22

Fixed

  • [core] packages — dry-run always reported success regardless of URL reachability — install() in ScriptInstaller, AptDebInstaller and AptRepoInstaller returned (True, "Dry-run completed successfully") unconditionally. Dry-run now sends a HEAD request to the download URL; an unreachable or 4xx/5xx response surfaces as (False, …). AptRepoInstaller additionally runs apt-get install --simulate to surface dependency conflicts when the package is already in the apt cache.
  • [core] audio_pipeline — HQPlayer stale cache persisted forever after stop — _refresh_hqplayer_cache never wrote None to _hqp_cache to guard against transient stopped states, but this also prevented clearing the cache when HQPlayer truly stopped (no track loaded). _get_hqplayer_item now raises _HQPlayerTrulyStopped (private sentinel) on confirmed stop; the cache is explicitly invalidated in that case while network errors keep the stale value.
  • [core] license — per-request aiohttp.ClientSession in license router — check_key, activate_license, get_public_config and the deactivate notify in delete_license each created a new ClientSession (TCP handshake + TLS) per call. Replaced with core.http.get_shared_session(), the same pool already used by library, tidal and qobuz.
  • [core] audio_pipeline — subprocess blocks event loop on NI data fetch — _get_local_ni_data() called subprocess.run("ip"/"iw") synchronously from a coroutine, stalling the event loop for up to 3s on every cache miss (TTL was 10s). Calls are now pre-warmed via asyncio.to_thread before pipeline construction; TTL raised to 60s.
  • [core] audio_pipeline — HQPlayer volume returned negative integers — formula 100 * (1 + volume_db / 60) produced negative values for volume_db < -60 (e.g. -140 for muted). Now clamped to [0, 100]; None when volume_db is unknown.
  • [core] audio_pipeline — cpu_percent() always returned 0.0 — a new psutil.Process instance was created on every cycle; the first call always returns 0.0 without a prior snapshot. Instances are now reused via _psutil_procs.
  • [core] audio_pipeline — incoherent double /proc/asound read — _get_alsa_latency and _get_alsa_buffer_fill each read status and hw_params independently, yielding different hardware snapshots. A shared _read_alsa_pcm_state helper performs a single coherent read.
  • [core] audio_pipeline — ALSA pointer wraparound incorrect on ARM64 — appl_ptr - hw_ptr < 0 was corrected with + 2**32; on ARM64 snd_pcm_uframes_t is 64-bit, causing negative or absurd latencies after ~6h. Fixed with + 2**64.
  • [core] audio_pipeline — _pid_identify_cache unbounded — raw dict with no TTL or size limit; replaced with TTLDictCache(300).
  • [core] audio_pipeline — M4A codec mismatch between modules — _enrich_with_topology mapped .m4a → ALAC while service._query_mpd mapped M4A → AAC for the same file. _detect_mpd_format is now the single source of truth; service.py imports it directly.
  • [core] audio_pipeline — _mpd_command duplicated core.mpd_client.mpd() — redundant implementation removed from now_playing.py; all calls go through core.mpd_client.
  • [core] audio_pipeline — _get_mpd_now_playing used direct Unix socket — replaced with core.mpd_client.mpd_batch (TCP), consistent with all other MPD consumers. currentsong + status sent in a single connection via command_list_ok_begin.
  • [core] audio_pipeline — blocking file I/O in async endpoint — open(topology_path) in async def get_topology() blocked the event loop; replaced with asyncio.to_thread(topology_path.read_bytes).
  • [core] audio_pipeline — dead variable + deprecated API — now = asyncio.get_event_loop().time() removed (now never used; get_event_loop() deprecated since Python 3.10).
  • [core] audio_hw — blocking I/O in async event loop — read_text(), exists(), iterdir() called directly from coroutines, stalling the event loop on every scan. All filesystem access moved to _scan_card_dir() (synchronous method) called via asyncio.to_thread; per-card details now fetched in parallel via asyncio.gather.
  • [core] audio_hw — exception path poisoned the 60 s cache — a transient I/O error (e.g. USB hotplug race) fell through to _cache.set(), serving an empty or partial device list for the full TTL. The error path now returns without caching.
  • [core] audio_hw — subdevice availability always reported as 1/1 — the if sub_info_file.exists() block contained only pass; subdevices_total and subdevices_available were always hardcoded to 1 regardless of actual device state. sub0/info is now parsed for subdevices_count and subdevices_avail.
  • [core] audio_hw — early return did not cache the negative result — when /proc/asound/cards is absent the empty result was returned without _cache.set(), causing a /proc stat on every subsequent call with no back-off.
  • [core] audio_hw — non-deterministic device ordering — iterdir() returns entries in filesystem order; a sorted() call now guarantees stable ordering by pcmNp name.
  • [core] hqplayer — play_uri / play_library_item silently failed on every play attempt — _send_batch passed allow_empty=False to all commands including <Play/>, which closes the connection without a body; _read_xml_response blocked for 3 s then raised HQPlayerError. Every HQPlayer playback call failed after loading the queue. Fixed by passing allow_empty=True for the Play command.
  • [core] hqplayer — resolve_file_path rejected all files when music_root='/' — path-traversal guard used startswith(root + os.sep) which fails for root '/'; replaced with Path.relative_to().
  • [core] hqplayer — active_rate=0 collapsed to None — rate if rate else None treated 0 Hz (DAC idle) as "no rate data"; fixed to rate if rate is not None else None.
  • [core] library — asyncio.get_event_loop() deprecated in upnp_service — replaced with asyncio.get_running_loop() (correct API inside async context, avoids DeprecationWarning on Python 3.13).
  • [core] library — _probe_port made two HTTP GETs per candidate URL — body was fetched once, then _fetch_device fetched the same URL again; _parse_device_text() helper now reuses the first response.
  • [core] library — _mpd_album_cache was an unbounded raw dict — replaced with TTLDictCache(60) + prune_to_size(100) to prevent unbounded RAM growth on large libraries.
  • [core] library — get_ext_stream_meta called twice per HTTP track in mpd_queue_list — result stored once and reused for both display enrichment and cover_token construction.
  • [core] library — Roon _subkey race in roon_queue_list — concurrent run_in_executor calls both read-modify-write sock._subkey without a lock; replaced with time.time_ns() as the subscription key.
  • [core] license — _ping leaked the aiohttp TCP connection — await session.post(...) without async with left the ClientResponse unclosed; fixed with async with session.post(...) as _resp: pass.
  • [core] license — online _fetch treated 4xx (revoked, not found) as "unreachable" — resp.raise_for_status() raised for 4xx and 5xx alike; now 5xx → "unreachable", 4xx → structured status from the LS (e.g. "revoked").
  • [core] license — start_checker leaked old background loop on double call — previous asyncio.Task was not cancelled before creating a new one; fixed with cancel guard.
  • [core] license — trial file non-atomic write → false "tampering detected" on crash — open(path, 'w'); json.dump(...) truncates on power loss; replaced with atomic_write_json (write-to-temp + rename).
  • [core] license — trial only written to first location on creation — any(...) short-circuited after first success, leaving the fallback location empty; tampering detection now requires all-locations consistency. Fixed with explicit list comprehension.
  • [core] license — require_full_license silently bypassed when service not initialised — if _service is None: return allowed all protected endpoints without a license check; now returns HTTP 503.
  • [core] packages — www.lesbonscomptes.com domain did not match after www. stripping — _validate_download_url strips www. before comparison; the set entry had the prefix, causing intermittent failures when set iteration returned that domain first. Changed to lesbonscomptes.com.
  • [core] config_validation — blocking subprocess and filesystem I/O in Pydantic validators — subprocess.run(['systemctl', ...]) and Path.exists() ran synchronously inside field validators on the async event loop, blocking for up to N×5 s per validation request. System-state checks (unit existence, file existence) moved to an async service method using asyncio.gather + asyncio.to_thread; validators now only check structure and path safety.
  • [core] config_validation — depends_on accepted duplicates silently — validate_no_duplicates covered start and stop but not depends_on; ["a","a"] was accepted without error. Fixed by extending the validator to all three list fields.
  • [core] auth — create_user accepted whitespace-only passwords — update_user rejected passwords where .strip() == \"\" but create_user did not; \" \" (6 spaces) was hashed and stored. Fixed via a model_validator on CreateUserRequest.
  • [core] auth — timing oracle on disabled accounts — authenticate_user returned immediately for disabled accounts (~1 ms) vs. burning a dummy bcrypt hash for nonexistent accounts (~250 ms), enabling username enumeration. Disabled-account path now also runs verify_password(password, _DUMMY_HASH) to equalise timing.
  • [core] auth — WebAuthn concurrent flows clobbered each other's challenge — _pending was keyed by username only; a second begin_* call for the same user overwrote the first challenge regardless of flow type (register vs. authenticate). Key changed to \"username:kind\" so both flows can coexist independently.
  • [core] core/users.py — _load_users read cache outside the lock — _cache and _cache_mtime were read without holding _lock while _save_users wrote them under _lock; concurrent async calls could observe a stale cache. Lock upgraded to RLock; _load_users now holds the lock for the full mtime-check + read sequence.
  • [core] audio_app_config — _run_command timeout covered only process creation — asyncio.wait_for was applied to create_subprocess_exec only; communicate() had no timeout and could block the event loop indefinitely on a hung sudo systemctl restart or stalled NFS mount. The timeout now wraps communicate() and kills the process on expiry.
  • [core] audio_app_config — restore_backup broken in production — sudo cp /var/backups/.../* /etc/... was not listed in the sudoers allowlist; every restore failed silently. Added the missing entries to the installer.
  • [core] audio_app_config — backup files world-readable — sudo chmod 600 /var/backups/audiogravity/*/* was not in sudoers; the chmod silently failed, leaving backup files (which may contain MPD passwords or Shairport auth keys) world-readable. Return value now checked; sudoers entry added.
  • [core] audio_app_config — total_count in backup list wrong under limit — BackupListResponse.total_count was computed from the already-sliced list; callers received total_count == limit instead of the real file count. Total is now captured before slicing.
  • [core] audio_app_config — debounce_task leaked on shutdown — the reload task created by the package-event listener was not stored on self and therefore not cancelled by cleanup(). It now fires _load_service_map on a partially-torn-down service on shutdown.
  • [core] audio_app_config — blocking open() in async _load_service_map — audio-config.json was read synchronously on every package-install reload event; replaced with asyncio.to_thread.
  • [core] audio_app_config — dead symlink check in _validate_path — path.is_symlink() was called after Path.resolve(), which always returns False; comment claimed a defence that did not exist. Removed; resolve() already handles symlink traversal via is_relative_to.
  • [core] audio_app_config — shairport serialiser produced unindented nested blocks — _dict_to_libconfig applied indent_str to leaf values but not to block headers or delimiters, producing malformed libconfig on save; round-trip parse→save→parse was broken for any config with nested sections.
  • [core] audio_app_config — _map_device_to_name O(n×m) scan — iterated all cards then all devices to find a card by ID; replaced with audio_hw_service.get_card_by_id() which is O(1) via the card index.
  • [core] test_version.py + test_push.py — paths broken after monorepo split — parents[2] pointed to /home/ad (old monorepo root); fixed to parents[1] (repo root). Script path updated to scripts/generate_vapid_keys.py.
  • [core] packages — _validate_destination_path path-traversal bypass — bare startswith(allowed_path) matched /tmp/audiogravity-packages-evil; fixed with separator-aware check (startswith(root + "/") or == root).
  • [core] packages — yes_proc leaked on install timeout — when use_yes=True and communicate() timed out, only the main process was killed; the yes subprocess kept running indefinitely. Fixed with yes_proc.kill() in the timeout handler.
  • [core] packages — installation_logs O(n) eviction — list.pop(0) shifted 499 entries per log line; replaced with deque(maxlen=500) for O(1) eviction.
  • [core] packages — background version-check task not tracked — asyncio.create_task result discarded; cleanup() method added to cancel it on shutdown.
  • [core] performance — subprocess.run blocking in async governor methods — _write_cpu_file, save_governor_config, and create_systemd_service blocked the event loop; wrapped in asyncio.to_thread.
  • [core] performance — cancel_test blocked event loop on process.wait() — subprocess.Popen.wait(timeout=2) ran synchronously inside an async def; replaced with await asyncio.to_thread(process.wait, timeout=2).
  • [core] performance — stddev computed by expanding histogram to 1 M-entry list — statistics.pstdev([lat]*count …) allocated ~28 MB per test on Pi; replaced with weighted variance over histogram buckets (O(k), k ≤ 200).
  • [core] performance — cleanup() published service_started on shutdown — wrong event_type; corrected to service_stopped.
  • [core] performance — logger.warn() removed in Python 3.13 — replaced with logger.warning().
  • [core] profiles — critical-stop error recorded but service start not gated — French error string only; the if/else already gates the start correctly (finding was a false positive on the logic, only the French string was fixed).
  • [core] profiles — bare except: pass in _listen_for_service_events — silently discarded all ServiceState parse errors; replaced with except (ValueError, KeyError) as exc: logger.debug(...).
  • [core] push — ec.generate_private_key monkey-patch applied on every instantiation — double-wrapping risk; guarded with getattr(..., '__name__') != '_patched_generate' so the patch is applied at most once per process.
  • [core] services — _get_cgroup_path called subprocess.run without asyncio.to_thread — blocked the event loop on every service monitoring poll; wrapped with to_thread at all three call sites.
  • [core] steering — _verify_alsa_device_exists, _verify_audio_flow, _get_active_alsa_devices blocked event loop — synchronous /proc/asound reads in async methods; extracted into _*_sync static methods called via asyncio.to_thread.
  • [core] sysinfo — CPUInfo with stale current_freq cached permanently — the full object including the dynamic frequency was cached on first call; now only immutable fields are cached and current_freq is re-read on every call.
  • [core] sysinfo — "Moins d'une minute" returned as API uptime_duration — French string in a user-facing JSON field; replaced with "Less than a minute".
  • [core] sysinfo — asyncio.get_event_loop() deprecated in log streamer and WebSocket terminal — replaced with asyncio.get_running_loop().
  • [core] sysinfo — thermal zone reads blocking in async _get_temperature_info — synchronous open() calls inside async method; extracted to _read_zone_temps sync helper called via asyncio.to_thread.
  • [core] radio — missing await on add_custom_station — the endpoint returned a coroutine object, making custom station creation permanently broken; await added.
  • [core] tidal — submit_redirect failure not surfaced — the router ignored the bool return value and always responded HTTP 200; now raises HTTP 400 on failure.
  • [core] tidal — out.read() blocking in async stream generator — every audio chunk read called out.read() synchronously on the event loop; replaced with await asyncio.to_thread(out.read, _CHUNK).
  • [core] tidal — PKCE code_verifier logged on token-exchange failure — full error response body written to logs; now only error and error_description fields are logged.
  • [core] tidal — concurrent start_pkce() calls silently corrupted each other's verifier — logged a warning when a flow is already pending; callers should not call start_pkce twice without completing the flow.
  • [core] qobuz — _extract_secrets silently produced wrong secrets when bundle regex found no matches — now raises RuntimeError when no timezone has a complete seed+info+extras triplet.
  • [core] qobuz — concurrent bundle cache misses fetched the 200 KB JS twice — asyncio.Lock added with double-checked locking.
  • [core] radio — _discover_mirrors fire-and-forget task not tracked — _discover_task stored on instance; stop() added to cancel it; exceptions now surface via add_done_callback.
  • [core] radio — _by_url dict grew unbounded from search hits — capped at 1000 entries with eviction of non-saved entries.
  • [core] player — _get_active_alsa_card blocked event loop on /proc/asound scan during DSD poll — synchronous filesystem walk extracted to _scan_active_alsa_card_sync and called via asyncio.to_thread.
  • [core] sse_manager — shutdown() called .cancel() on a List[Task] instead of each task — AttributeError on every graceful shutdown; monitoring loops kept running after the event loop closed.
  • [core] utils/files — os.fsync() in atomic_write_json blocked event loop 50–300 ms on SD card flush — removed; os.replace() atomic rename is sufficient for config-file integrity.
  • [core] dbus_client — get_dbus_client() singleton had no concurrency guard — two concurrent callers both created separate D-Bus connections; asyncio.Lock added.
  • [core] roon_client — get_roon_client() singleton had no concurrency guard — asyncio.Lock with double-checked locking added.
  • [core] http — _get_shared_session() not coroutine-safe — two concurrent first callers both created ClientSession; asyncio.Lock added.
  • [core] dbus_client — bare except: pass swallowed asyncio.CancelledError — replaced with except (TypeError, ValueError): pass.
  • [core] ttl_cache — TTLCache.get() always missed for a cached None value — _value is not None guard prevented None from being representable as a cache hit; replaced with _UNSET sentinel.
  • [core] JWT and auth error messages in French — "Token invalide ou expiré", "Accès réservé aux administrateurs", "Les invités ne sont pas autorisés" translated to English (returned verbatim in JSON API responses).
  • [core] www.lesbonscomptes.com domain entry did not survive www. stripping — changed to lesbonscomptes.com fixing an intermittent test failure.
  • [lic] XSS in portal upgrade and download forms — filename from the server Content-Disposition header and newKey from user input were injected into .innerHTML without escaping; a crafted value could exfiltrate the admin session token. HTML-escaping helper _esc() added to both components.
  • [lic] version_scope stripped on license resend, bulk-resend, and transfer — generate_lic() was called without the version_scope parameter, silently converting v1-scoped licenses to all-versions licenses and bypassing the upgrade paywall.
  • [lic] All-versions lifetime licenses incorrectly rejected on AG v2 — payload.get('version_scope', '1') defaulted absent scope to "1", causing version_expired for licenses issued before scoping was introduced. Default changed to None (no check when field is absent).
  • [lic] Admin token comparison was not constant-time — != string comparison allowed character-by-character brute-force timing attack; replaced with hmac.compare_digest.
  • [lic] Blocking SMTP calls on the async event loop — all mailer.send_*() in admin handlers now use await asyncio.to_thread(...).
  • [lic] Blocking file I/O in import_db — shutil.copy2 and Path.write_bytes wrapped with asyncio.to_thread.
  • [lic] purge_audit read db.total_changes (cumulative) before commit() — now reads cur.rowcount after the DELETE for an accurate count.
  • [lic] /deactivate unprotected when VERIFY_KEY is empty — @limiter.limit("10/minute") added.
  • [lic] Revoked license could be deactivated, corrupting audit trail — AND revoked_at IS NULL added to the deactivate_order UPDATE condition.
  • [ui] XSS in license status acquisition steps — unsafeHTML(_acquisitionStepsHtml()) replaced with a safe Lit html template; _priceDisplay (from backend license_price) is now a text node.
  • [ui] XSS in license status portal URL — _portalUrl from the backend now validated as https?:// before embedding in InfoModal HTML string; javascript: and data: URLs are rejected.
  • [ui] XSS in package update confirm dialog — pkg.label, pkg.installed_version, pkg.available_version (from package registry) are now HTML-escaped with escapeHtml() before injection into showConfirm HTML string.
  • [ui] JSON.parse on corrupted localStorage crashes auth init — auth.js:initAuth now wraps JSON.parse(userStr) in try/catch; a corrupted jwt_user value calls clearAuth() and returns false instead of throwing.
  • [ui] showToast without window. prefix throws ReferenceError — ag-config-editor.js:_handleCancel was calling bare showToast; fixed to window.showToast.
  • [ui] CodeMirror instance leaks on every config editor open/close — disconnectedCallback added to destroy the CodeMirror instance and prevent DOM accumulation over long admin sessions.
  • [ui] @lit/context imported from CDN (no SRI, supply chain risk) — replaced with local npm package @lit/context (installed as proper dependency).
  • [ui] addToHistory used as implicit global in ag-admin-page — added to named imports from common.js; was silently broken in module-isolated test environments.
  • [ui] this.loading not reactive in ag-config-page — loading: { type: Boolean } added to static properties; the loading guard now prevents concurrent service config requests.
  • [ui] Stale radio search results overwrote current ones — _loadSearch now uses AbortController; results from cancelled requests are discarded.
  • [ui] version.test.js path broken after monorepo split — ROOT corrected to audiogravity.ui/ repo root; VERSION resolved from sibling audiogravity.ops/VERSION; frontend/ prefix removed from file paths.

Added

  • [core] audio_hw — force_refresh query parameter on GET /audio-hw/devices — ?force_refresh=true bypasses the 60 s cache and triggers an immediate rescan, useful after a USB hotplug event.
  • [core] audio_hw — total_cards as a computed field — converted to Pydantic v2 @computed_field; can no longer diverge from len(cards).
  • [core] requirements-dev.txt — test dependencies (pytest, pytest-asyncio, httpx) separated from requirements.txt in audiogravity.core.
  • [core] audio_pipeline — TestDetectMpdFormat, TestAlsaPcmState (latency, 2**64 wraparound, closed device), TestHqplayerVolume (clamping, None), TestPidIdentifyCacheType — 64 tests passing.
  • [site] THIRD_PARTY_NOTICES.md — full attribution for all open-source components: frontend (Lit, Cytoscape, dagre, Lucide, Chart.js, CodeMirror, Inter, JetBrains Mono), backend (FastAPI, Pydantic, pywebpush/MPL-2.0, cryptography, roonapi, 15+ libraries), runtime binaries (ffmpeg/LGPL, cyclictest/GPL-2, iperf3/BSD-3, smartctl/GPL-2), external API (Radio Browser), and optionally-installable audio software (MPD, upmpdcli, shairport-sync, HQPlayer NAA, Roon Bridge/Server).
  • [site] index.html — announce bar updated to v0.9.4; Features section: title includes Tidal, Unified Transport card adds stream origin badge, High-resolution library card mentions UPnP search playable, MQA removed from signal-path description; Compare table: new row for native Tidal & Qobuz streaming (OAuth2/PKCE), UPnP row updated to mention search + directly playable results; footer version v0.9.4 and "Open Source" link in Legal.
  • [site] README.md — reference to THIRD_PARTY_NOTICES.md in license paragraph and Documentation section.

[0.9.4] - 2026-06-20

Fixed

  • [core] Pydantic v2 deprecation — .dict() → .model_dump() — services/service.py was using the deprecated .dict() method; replaced with .model_dump().
  • [core] RuntimeWarning coroutines never awaited — two test helpers left coroutines unawaited when patching asyncio.wait_for; both fixed by closing coroutines before raising or restructuring the mock.
  • [ui] Code duplication removed across components — loadConnection(host, fetchFn, tag) helper extracted to utils-lit.js, used by Qobuz/Tidal/HQPlayer output molecules (3×10 lines removed). queueWithFeedback(queueFn, label) helper added to library-api.js, used by ag-library-search and ag-library-browse. _formatRelativeTime() removed from ag-service-card and ag-profile-card in favour of the existing formatTimestamp() from utils-lit.js. 16 unit tests added (formatTimestamp, loadConnection, queueWithFeedback, escapeHtml XSS, jitterChart destroy, push URL format, password trim).
  • [ui] XSS — escapeHtml imported directly in ag-admin-page — removed window.escapeHtml ? ... : username ternary; escapeHtml now imported as an ES6 module so username is always escaped in the delete-user confirmation dialog.
  • [ui] XSS — unsafeHTML(this.label) removed from ag-metric-detail — label prop now bound with Lit's auto-escaping (${this.label}); unused unsafeHTML import removed.
  • [ui] Memory leak — ag-network-test jitter Chart.js not destroyed — this._jitterChart.destroy() added to disconnectedCallback().
  • [ui] Push unsubscribe used wrong HTTP method — changed from POST with JSON body to DELETE with query param (apiDelete('/push/unsubscribe?endpoint=...')), matching the backend router.
  • [ui] Password trim in user modal — this._password now .trim()-ed before validation; whitespace-only passwords are caught by the existing length < 6 check instead of reaching the backend.
  • [core] core — D-Bus proxy cache eviction on error — stale unit proxies are now removed from _unit_proxy_cache on call_get_all failure so the next call rebuilds a fresh proxy instead of reusing a dead one indefinitely.
  • [core] core — Roon disconnect timeout — _roon_api.stop() wrapped in asyncio.wait_for(timeout=5) to prevent indefinite hang; state is always cleaned up in finally.
  • [core] core — JWT decode error log sanitised — exception logged as type(e).__name__ only, preventing any token fragment from appearing in logs.
  • [core] core — EventBus QueueFull now logged — slow subscriber drop logged at DEBUG level for diagnostics; logging module added to events.py.
  • [core] sysinfo — smartctl FileNotFoundError crashes monitor — caught and degraded gracefully; disk temperature returns None when smartctl is absent. SYSLOG_IDENTIFIER fixed to _SYSLOG_IDENTIFIER= (journalctl match syntax). Invalid grep_pattern regex now returns 400. lscpu absence caught. Admin PTY shell logs an audit entry.
  • [core] push — webpush() without timeout — timeout=10 added to prevent indefinite blocking on unreachable endpoints. Endpoint now validated as HTTPS URL. endpoint query param annotated with Query(). logger.error on successful unsubscribe changed to logger.info.
  • [core] config_validation — appconfigfile path traversal — paths now validated against /etc and /usr/local/etc whitelist (mirrors audio_app_config). systemd_unit bounded to max_length=255. Substring match replaced with line-start check. ValueError in validation now returns 400 instead of 500.
  • [core] Steering — ALSA device injection via config write — alsa_device now validated against ^hw:\d+,\d+$ before any config file substitution; get_steerability() no longer hardcodes upmpdcli/roonbridge as always available; _verify_alsa_device_exists() checks /proc/asound/cardX/pcmYp (playback subdevice), not just the card directory.
  • [core] Performance — cyclictest IndexError on trailing token — bounds check (i + 1 < len(tokens)) added to the cyclictest line parser.
  • [core] Performance — $CONFIG_FILE unquoted in boot script — config path now passed as sys.argv[1] and shell-quoted ("$CONFIG_FILE"), preventing word-splitting on paths with spaces.
  • [core] Performance — duration_seconds always 0 — test start captured with time.monotonic(); actual elapsed time stored in LatencyTestResult.
  • [core] Performance — bare except: clauses — replaced with except (ValueError, IndexError) and except (subprocess.TimeoutExpired, OSError).
  • [core] Profiles — asyncio.gather() without timeout — both stop and start phases wrapped in asyncio.wait_for(timeout=30).
  • [core] Profiles — stopped_count counted FAILED as stopped — FAILED state now goes to failed_count; only INACTIVE/DEAD states increment stopped_count.
  • [core] Profiles — export written to world-readable /tmp — export path moved to settings.audiogravity_home with chmod(0o600).
  • [core] License — server response handling hardened — /check and /activate now verify HTTP status before calling resp.json(); unexpected Pydantic shapes return 502; lic_content validated as JSON before writing to disk; X-Verify-Key header extracted to _verify_headers() helper; _portal_base() validates URL structure.
  • [core] Services — enum comparison bug — validate_service_properties was comparing CPUSchedulingPolicy/IOSchedulingClass enum members to plain strings (always-false conditions); corrected to compare against enum values.
  • [core] Services — D-Bus call unguarded — get_unit_file_state() now wrapped in asyncio.wait_for(timeout=2s); timeout logs a warning and falls back to subprocess.
  • [core] Services — stale cgroup FD not evicted — fd.seek(0) failure now evicts the dead FD from the LRU cache instead of leaving it for the next call.
  • [core] Services — service_name path param unvalidated — all service_name Path params now carry pattern=r"^[a-zA-Z0-9._@-]+(?:\.service)?$" → 422 on invalid input.
  • [core] audio_app_config — path traversal via symlink — _validate_path now calls resolve() first, then checks is_relative_to() on the resolved path; a symlink pointing outside /etc is rejected even if the link itself lives inside /etc.
  • [core] audio_app_config — restart duplication — _restart_service() extracted; update_config and restore_backup both delegate to it.
  • [core] audio_app_config — model_validator incorrect — @field_validator('content','data') replaced with @model_validator(mode='after') for correct single-evaluation semantics.
  • [core] Packages — shell injection via installer config — install_script_args, check_command, uninstall_commands and version_check_command now use shlex.split + create_subprocess_exec instead of shell=True f-strings. version_check_url validated against ALLOWED_DOWNLOAD_DOMAINS. gpg_key_path and sources_list_path validated by _validate_destination_path() before any sudo cp. _validate_package_name() added in apt_repo.py. 9 security tests added.
  • [core] Player _poll_loop NameError on get_now_playing() failure — items now initialised to [] before the try block, preventing NameError when get_now_playing() raises.
  • [core] Player DSD state inconsistent after partial gather() failure — _dsd_active reset to False on exception so volumes are not left partially forced.
  • [core] Auth POST /users returned 200 instead of 201 — corrected to HTTP_201_CREATED.
  • [core] Auth PATCH /users/{u} accepted whitespace-only password — passwords that are blank after .strip() are now rejected with 400.
  • [core] Auth update_user race condition post-update — update_user() now returns Optional[User] directly, eliminating the separate get_user() call after write.
  • [core] Auth WebAuthn userHandle exception was silent — bare except replaced with logger.warning() for diagnosis; _b64url_to_bytes moved to module top.
  • [core] users.py write race condition — _save_users now holds a threading.Lock for thread-safe concurrent writes.
  • [core] JWT missing jti claim — uuid.uuid4() added to every token payload, enabling future revocation support.

Added

  • [core] HQPlayer POST /hqplayer/stop — endpoint to stop playback (was documented but not implemented); also adds stop() service method and has_dsp_config public property. 13 unit tests added covering stop, Literal validation, has_dsp_config, _read_xml_response helper, and play request validation.

Fixed

  • [core] Radio PUT /radio/{uuid} broken — edit_station() was awaited incorrectly (missing await), causing PUT to always return a coroutine instead of the result.
  • [core] Radio added_at timezone-unaware — datetime.utcnow() (deprecated Python 3.12) replaced with datetime.now(timezone.utc).
  • [core] Audio pipeline DSD stream not detected — audio_str="dsd64:2" caused int("dsd64") → ValueError silently swallowed; DSD multiplier now parsed correctly (dsd_mult × 44100, 1-bit).
  • [core] Audio pipeline cgroup v1 PID→service mapping — split("::", 1) only matched cgroup v2; now uses fields[-1] to support both formats.
  • [core] Audio pipeline dbus-send without timeout — both subprocess communicate() calls now wrapped with asyncio.wait_for(timeout=3.0) + proc.kill() on expiry.
  • [core] Audio pipeline MPD URI control-char filter — any(ord(c) < 0x20) replaces the newline-only check.
  • [core] Library UPnP queue ignores output_id — upnp_search_queue now honours req.output_id like Qobuz/Tidal.
  • [core] Library SOAP Browse object_id not XML-escaped — _sax.escape(object_id) applied before template formatting.
  • [core] Library Qobuz album/playlist queue duplicated — extracted _queue_qobuz_tracks() helper (~25 lines removed).
  • [core] Library _tidal_cover crashes on dict uuid — isinstance(uuid, str) guard added.
  • [core] Qobuz private_key leaked in query string — _exchange_code changed from GET+params to POST+body.
  • [core] Qobuz token refresh race condition — asyncio.Lock added to TidalOAuthService.get_access_token with double-check pattern.
  • [core] Qobuz half-authenticated state — handle_callback returns False when _find_working_secret() returns None instead of saving an app_secret=None state.
  • [core] Qobuz user_id could be string "None" — str(user.get("id")) when id is JSON null now correctly returns None.
  • [core] Tidal stale token not cleared on network error — _access_token set to None when refresh fails and token is past expiry.
  • [core] HQPlayer item_type/action unvalidated — changed from str to Literal["track","album","artist"] / Literal["play","add"] in HQPlayerPlayLibraryRequest.
  • [core] HQPlayer XML parsing duplicated — extracted _read_xml_response() static helper used by both _send() and _send_batch().
  • [core] get_shared_session() exported — LibraryService._http_session() now delegates to the app-wide shared session (closed on graceful shutdown), eliminating socket leak.

Added

  • [ui] Config editor — blank configuration hint — when a service config file has all sections empty (package defaults active), the form view now shows an info banner explaining the state and offering a direct link to Expert Mode to view and uncomment the full file.
  • [core] Tidal/Qobuz credential rotation detection — 401/403 responses from playbackinfopostpaywall (Tidal) and track/getFileUrl (Qobuz) after token refresh now raise typed exceptions (TidalClientRotatedError, QobuzClientRotatedError) logged at ERROR level with a remediation hint. The Tidal stream endpoint returns HTTP 503 instead of 404 on rotation. 10 unit tests cover both services.
  • [core+ui] Stream origin badge in the players — the mini and fullscreen Now Playing views now show where the audio comes from (Tidal, Qobuz, a UPnP/DLNA server by name, radio, a local file, Roon, AirPlay…) with a logo + label, instead of an undifferentiated "MPD". The backend tags each playing source with its provider, independent of transport.
  • [site] FAQ — connecting Qobuz or Tidal — one entry covering both flows: Qobuz OAuth (auto-redirect, no paste) and Tidal PKCE (copy/paste the redirect URL), noting the required subscription tier for each.
  • [core] UPnP search queue — clicking "+" or play on a UPnP/DLNA search result now works: track stream URLs are passed directly to MPD, album items are resolved via ContentDirectory Browse. Titles and cover art are pre-registered during search so the queue display and Now Playing badge are enriched immediately. 17 unit tests cover track/album/routing/metadata-registration cases.

Fixed

  • [core] _ext_stream_key anchored to URL path/query — Tidal detection now checks the URL path for /tidal/stream/ and Qobuz detection checks the query string for eid=, instead of raw substring matching. Prevents a UPnP URL whose path contains /tidal/stream/ or eid= from being misclassified. 8 regression tests added.
  • [ui] Software tab description — removed the "dry-run simulation" claim from the Pro feature list; the toggle is admin-only and not visible to regular Pro users.

Changed

  • [ui] Library search lists UPnP/DLNA servers — known media servers (e.g. MinimServer) now appear as sources in Library search, alongside MPD, Roon, Qobuz and Tidal (previously missing).
  • [ui] UPnP/DLNA text search from the Search tab — selecting a UPnP/DLNA server badge while in the Search tab now stays in search mode and runs a text query against the server (previously always redirected to the UPnP browser).
  • [ui] Settings: single version + logout moved in — the panel now shows one unified product version (front and back share it) with the Swagger API link, and the Logout button moved from the top bar into the Settings footer.
  • [ui] Top-bar icons — the settings button now uses a gear icon and the mobile navigation button a hamburger (swapped for clarity).
  • [ui] Source connection status uses the shared status dot — Qobuz, Tidal and HQPlayer now render the same status indicator as the other sources.

Fixed

  • [ui] Qobuz sign-in flow — the Qobuz OAuth page now opens in a popup (the AG UI stays underneath) that closes itself once authentication succeeds, so you land back in Audiogravity automatically instead of having to close the tab and reopen the app.
  • [ui] Missing-cover icon in the fullscreen player — the "no artwork" placeholder icon was rendered far too small; it now fills the cover area again.
  • [core] Tidal playback skipped every track — the seek cache pre-created the remux output file, so ffmpeg (run without -y) refused to overwrite it and exited having written 0 bytes; MPD received an empty stream and skipped to the next track. ffmpeg now overwrites the placeholder, and a 0-byte remux is never cached.
  • [ui] Bundle analysis report no longer deployed — the Vite stats.html bundle treemap was copied into the build output and shipped to production (reachable at /stats.html). It is now opt-in only (vite build --mode analyze) and written outside the deployed output, so it never ships.

Removed

  • [ui] Standalone styleguide page — the unused styleguide.html design-system gallery was removed (Storybook is the live component reference).

[0.9.3] - 2026-06-14

Added

  • [core+ui] Tidal streaming source — connect a Tidal HiFi account (PKCE login) and browse Favorites / New Releases / Playlists, plus search; playback streams lossless FLAC through a local DASH→FLAC ffmpeg proxy (Tidal delivers FLAC as segmented DASH, not a direct URL). Requires ffmpeg (added to the backend installer).
  • [core+ui] Tidal Charts & Editorial browse — two new browse pills: Charts (TIDAL's Top Hits, Viral / Rap / R&B / Pop Hits…) and Editorial (Popular, Trending, TIDAL Rising, Podcasts…), extracted from Tidal's home page and rendered as playlists; tapping one opens its tracks like any playlist.
  • [core] Tidal in-track seek — the proxy remuxes each track to a seekable FLAC file (served as it is produced, so start-up stays fast) and keeps it in a small, disk-backed cache (current track + a couple of recent, wiped at startup). Replays and reopens are served with HTTP Range, so you can seek within Tidal tracks. The very first play of a track was not seekable in this release; it is since 0.9.34.
  • [ui] Top bar — mobile navigation toggle — a left-side button opens the vertical tab menu on mobile, mirroring the settings burger on the right.
  • [ui] Top bar — Library shortcut — a one-tap button (left of Logout) jumps straight to the Library tab; licence gating is preserved (a locked tab opens the licence modal instead).
  • [ui] Roon source logo — the RN text placeholder for Roon sources is replaced by a Roon logo rendered as a currentColor mask, so it stays visible in light, dark and the active (selected) state.

Changed

  • [ui] DRY-RUN toggle is now admin-only — the Audio Software DRY-RUN simulation is scoped to administrators (a command-preview / catalog-validation tool, not a dependency-resolving simulation); its size now matches the settings toggles.

Fixed

  • [ui] Login page version label — shown as v0.9.2 (lowercase prefix, no space).
  • [core] AirPlay now-playing on ARM — shairport-sync track metadata leaked the raw <dbus_fast…Variant…> repr in the pipeline now-playing; the a{sv} values are now unwrapped (portable, no-op on x86).
  • [ui] Fullscreen player volume swipe — adjusting the volume slider no longer triggers the player's multi-source swipe.
  • [ui] Library horizontal scroll — scrolling a horizontal row (pills, album shelves) no longer switches tabs; the swipe-to-switch handler now yields to any horizontally-scrollable ancestor (CSS class or inline overflow), not just inline ones.
  • [ui] Fullscreen player pull-down dismiss — swiping down to scroll back up in a long tracklist no longer dismisses the player; the dismiss gesture only fires when the scroll area is already at the top.

Removed

  • [ui] Obsolete top-bar buttons — the documentation (audiogravity.app) button and the admin "Components" (bundle stats) button were removed.
  • [ui] Tab-bar now-playing — the vertical tab sidebar no longer displays the current stream (redundant with the mini player).

[0.9.2] - 2026-06-09

Focus: Qobuz Hi-Res streaming integration + DSD volume safety.

Added

  • [core] Qobuz OAuth2 module (modules/qobuz/): full OAuth2 authentication replacing the deprecated username/password login. Bundle credential extraction from the Qobuz web player JS, OAuth URL generation, browser-based login, code-to-token exchange, working secret discovery, and persistent config storage.
  • [core] Qobuz search, browse, and playback (library/service.py): header-based API auth (X-App-Id/X-User-Auth-Token), correct API signature format (float timestamp, ASCII+UTF-8 byte concatenation), 3 parallel single-type search calls, album queueing, cover art on all browse/search methods.
  • [core] Qobuz catalog browsing — featured albums (album/getFeatured), editorial playlists (playlist/getFeatured, playlist/get). Three new endpoints: GET /library/qobuz-featured, GET /library/qobuz-playlists, GET /library/qobuz-playlist-tracks. Playlist queueing support.
  • [core] External stream metadata registry (now_playing.py): stable key derivation from eid query parameter for Qobuz streams — correct title, artist, album, and cover art in now-playing and queue despite ephemeral signed URLs.
  • [core] Qobuz as virtual source — src_qobuz injected in the player when connected; bitrate displayed for Qobuz FLAC streams.
  • [ui] Qobuz connection card (ag-qobuz-output.js): OAuth2 connect/disconnect molecule for the Sources view, with polling and events.
  • [ui] Qobuz catalog pills (ag-library-browse.js): browse pills switch to Favorites / New Releases / Selection / Playlists when Qobuz is selected.
  • [ui] Qobuz icon on source cards (optimized 56×56 webp).
  • [core] Qobuz OAuth2 unit tests (24 tests): bundle extraction, service persistence, OAuth flow, Pydantic models, router endpoints.
  • [core] Qobuz catalog browsing unit tests (22 tests): featured albums, playlists, playlist tracks, search, cover helper, library router endpoints.
  • [core] DSD volume unit tests (10 tests).
  • [ops] ./dev.sh test --report generates TEST_REPORT.md — markdown summary with per-suite pass/fail/skip counts, durations, and failure detail.

Changed

  • [core] Qobuz config cleanup — removed qobuz_app_id, qobuz_app_secret, qobuz_username, qobuz_password settings; OAuth2 manages credentials via UI.
  • [ops] build-backend-package.sh — qobuz added to ENABLED_MODULES template and upgrade migration loop; QOBUZ_FORMAT_ID=27 in .env template.
  • [ui] Topbar doc button opens audiogravity.app; internal doc links removed from settings panel.
  • [docs] Single source of truth — CHANGELOG.md and RELEASE_NOTES.md consolidated in the landing repository; AG-repo copies marked obsolete in CLAUDE.md.

Fixed

  • [core] DSD volume protection (player/service.py): 6 bugs causing volume to snap to 100% during non-DSD playback — stale HQPlayer cache, race conditions, incorrect restore target.
  • [core] HQPlayer stale track — _current_track cleared after 30s stopped.

[0.9.1] - 2026-06-07

Changes since 0.9.0. Range: 74a7897 (0.9.0 baseline — Merge feat/hqplayer-integration into master, 2026-05-30) → release HEAD. The earlier v0.9.1 git tag (b3ab0c2) was never a release and is superseded by this one.

Focus: ARM/Debian (aarch64) portability. Audiogravity now installs and runs on aarch64 (Raspberry Pi 4 / Debian) alongside x86/DietPi — with the absolute constraint of never regressing the existing x86/DietPi target. This cycle also hardens the install/packaging path, configures the production environment at install time, fixes the standalone prod server (SSE + WebSocket), and adds test coverage. Period: 2026-05-30 → 2026-06-07.

Added

  • ARM64 (aarch64) as a first-class target — backend, frontend, license server and audio-software installs validated end-to-end on a Raspberry Pi 4 (Debian), with x86/DietPi kept non-regressed.
  • [ops] upmpdcli on ARM64 — scripts/build-upmpdcli-arm64.sh builds the libnpupnp → libupnpp → upmpdcli chain natively for aarch64 and publishes a checksum-verified .deb bundle on audiogravity.app. The package registry installs it via a per-arch arch_fallback (no upstream arm64 package exists anywhere); the official source still wins where it covers the arch.
  • [ops] WebAuthn passkeys at install — --public-url sets WEBAUTHN_ORIGIN and derives WEBAUTHN_RP_ID; forwarded by the bootstrap and all-in-one installers and documented in the README (passkeys require a real HTTPS domain, not a bare IP).
  • [ops] Push contact — --vapid-email sets the VAPID sub; forwarded by the installers.
  • [ui] HQPlayer manual IP entry — connect to a HQPlayer on a different (but routable) subnet that the local /24 discovery scan can't reach.
  • [ui] Standalone server WebSocket proxy — the prod Python server now proxies the admin terminal WebSocket (/sysinfo/terminal/ws) via raw tunneling.
  • [site] Page — the "Releases" link now points to the public GitHub repo; install section clarified (works on the LAN by default, passkeys/push moved to an optional block that requires a public domain); portable local preview dev.sh (auto-detected LAN IP).
  • [ops] Deterministic dependencies — requirements.lock (68 packages pinned, wheels verified for x86_64 and aarch64, Python 3.13.5) + runtime/build split (requirements-dev.lock); license server dev requirements (requirements-dev.txt).
  • [core] — VAPID key loading, PayPal IPN webhook flow (8 tests), and scoped pytest warning filters for third-party deprecations.

Changed

  • [core] Auth — replaced passlib with direct bcrypt calls.
  • [ui] Prod server consolidation — the frontend package now bundles the canonical serve_https.py (line-by-line SSE streaming + gzip) instead of a divergent, broken inline copy; the systemd unit invokes it with args + WorkingDirectory.
  • [ops] Production .env configured at install — license server URL/key, VAPID, and ENABLED_MODULES (incl. hqplayer).
  • [ops] build-backend.sh self-bootstraps — installs system prerequisites and python3; the generated installer uses non-interactive apt; deps synced via pip install -r with wheels only.

Fixed

  • [ui] Prod SSE stuck on "Connecting…" — the packaged proxy buffered the event stream (shutil.copyfileobj); now streamed so the UI connects.
  • [ui] Broken-pipe traceback noise in the standalone server when a client drops a keep-alive connection (e.g. after closing an SSE tab).
  • [core] Runtime bugs surfaced on ARM — CPU model (lscpu fallback), governor field (current_governor), cpu-governor.service boot script (stdlib json), and ORJSONResponse deprecation warnings.
  • [core] Service config paths resolved across OS layouts (/etc vs /usr/local/etc) and reloaded when software is installed/removed at runtime.
  • [core] hqplayer module missing from the packaged .env ENABLED_MODULES (its endpoints returned 404).
  • [core] aarch64 startup crash — caused by passlib vs bcrypt>=4.1.
  • [ops] Audio-software install made arch-aware — Roon per-arch URLs ({roon_arch}), download allowlist, stdin for interactive vendor installers, and complete uninstall; upmpdcli .deb ships its systemd unit, creates its system user, and stops/disables cleanly on removal.
  • [ops] sudoers completed (including Roon uninstall) and hardened (sudo tee instead of sudo sh -c).
  • [ops] Nuitka / packaging — explicit includes (psutil, distro, orjson, pydantic, ntplib), SHA256SUMS merge + checksum verification in installers, missing transitive deps (pyparsing, ifaddr), gh auto-install, license.pub lookup, and miscellaneous installer robustness fixes.
  • [ops] Environment scripts — no longer hardcode /home/dietpi paths or LAN IPs (runtime detection); seed a default admin user; provision dev sudoers/polkit; resolve visudo by absolute path on DietPi.

Removed

  • [ops] obsolete backend/update.sh (incompatible with the Nuitka-binary production model).

Scope note: this release consolidates the arm-debian-compat work (main repo) plus this landing repository. The earlier v0.9.1 tag (b3ab0c2) was never a release — delete it (local + remote) and re-tag the release commit as v0.9.1.