release.yml already builds and publishes desktop installers for every platform
on a v* tag, so no new workflow was needed — it needed adapting to the fork.
electron-builder published to upstream's repository, so releases (and with them
the electron-updater feed, which had no source at all) went nowhere useful.
More consequential: postinstall ran electron-rebuild against uiohook-napi only.
electron-native-screenshare would have been compiled for Node's ABI rather than
Electron's and failed to load at runtime — and since the loader degrades
quietly by design, the symptom would have been screen sharing with no sound and
no error, which is the exact bug this module exists to fix.
Documented in docs/build-desktop.md, including why the build has to run on
GitHub (Gitea Actions provides no hosted runners, and the module needs MSVC)
and why the Windows runner stays pinned to windows-2022.
The originals were all 1.14s — long enough to be intrusive for events that
fire whenever anyone joins, leaves, or watches a stream.
Synthesised rather than sourced: no third-party file, so no licensing question
in a public repository. The parameters are measured, not guessed — envelope,
spectral peaks and decay taken from two reference sounds the instance owner
supplied. The timbre is fundamental plus octave at near-equal strength
(1.00 / 0.85 / 0.10 / 0.02) decaying to 1/e in 0.19s, with no reverb.
Call join rises C4→G4, leave falls D4→G3, matching the references' intervals
and their 100ms spacing. The stream pair reuses those resolution notes as
single tones at ~55% the volume: they fire far more often during a broadcast,
so they have to sit under the call sounds rather than beside them.
Stream start is the one event that happens once per broadcast, so it can be a
chord — with a low thump and a short air layer that both die inside 150ms,
adding weight without length. Stream end mirrors it descending, quieter, and
without the air, since brightness reads as arrival.
Generators and the measured parameters are kept in tools/sfx so these can be
retuned without redoing the analysis. Total size drops from 150KB to 41KB.
Electron's audio: 'loopback' captures the whole output mix, this app's own
playback included — so everyone else's voices went back out inside the share
and each listener heard themselves. Not acoustic echo but a digital copy of
the output, which is why headphones never helped, and why shareAudio already
defaulted to off in the desktop app.
Electron offers no way to exclude our own audio: the docs allow only
'loopback' or 'loopbackWithMute', and the handler discards the renderer's
constraints (restrictOwnAudio never arrives). electron-native-screenshare does
it at the OS level — WASAPI process loopback on Windows — capturing only the
shared window when its pid resolves, and otherwise everything except us.
The module hands raw PCM to the main process, so it crosses IPC and is
scheduled onto a running cursor in Web Audio to become a MediaStreamTrack,
published as ScreenShareAudio. Loading is optional and failure degrades to a
silent share rather than blocking the app or the screen share.
The browser path is untouched: Chrome honours restrictOwnAudio and has no echo.
Verified by typecheck (web and Electron main) and the web suite. The audio path
itself cannot be exercised here — no Windows, no Electron, no audio device.
Soundboard: naming a clip used window.prompt, which Electron does not
implement — it returned nothing, the flow aborted in silence, and adding a
sound worked in the browser while doing nothing at all in the desktop app.
Replaced with a two-step field inside the popover, identical in both.
Spotify, three separate defects behind the two symptoms reported:
Out of sync — a 20s poll stacked on the activity store's 5s debounce left
everyone else on the previous track for up to 25s. The next poll is now
scheduled just past the current track's end instead of on a fixed interval,
and a track change bypasses the debounce (it happens once every few minutes;
the debounce exists for chatty producers).
Vanishing — a paused track, and the silent gap Spotify reports between two
songs, both cleared the activity outright. Pausing is now carried as state
rather than absence, and an empty answer is tolerated for 25s before the
block comes down.
Progress bar — timestamps are computed with the server's clock and were drawn
against the viewer's, so any drift displaced the bar; and it kept advancing
after a pause until the next poll. The ready payload now carries server time
so each client can correct its own offset, and the bar freezes when paused.
Tray, native notifications and system audio in screen share were all found
already implemented and wired end to end; recorded in the roadmap rather than
built again.
The add-sound control was gated on channelPermissions, which carries the
per-channel bitfield; MANAGE_SPACE lives in the space-level one. The check
silently evaluated false for everybody, including owners, so the button never
rendered and there was no way to add a sound at all.
Also raise the clip cap to 2 MB — a 30-second clip at a high bitrate cleared
1 MB — and stop reporting every upload failure as 'too large', which sent
people to shrink a file that was not the problem.
Soundboard: the trigger travels over the WebSocket and every client in the
call plays the clip locally, instead of mixing it into the presser's
microphone or publishing a LiveKit track. No upstream bandwidth, no media
stack changes, and the clip is not degraded by voice processing.
Fan-out uses a new sendToRoomParticipants rather than sendToRoom: the latter
broadcasts a space room to the whole space, which is right for the presence
the sidebar shows and wrong for anything audible. The cooldown is enforced
server-side — a client-side one only slows down people not trying to abuse it,
and a soundboard is the easiest thing here to turn into a weapon. Playing is
open to anyone in the call; deciding what the buttons are needs MANAGE_SPACE.
Account menu: the name in the user bar had cursor-pointer and no handler, so
the interface was already promising a click that did nothing. Offers profile,
status and copy-id — not the Clips or account switching the reference design
shows, which would be dead UI here.
Call timer: startedAt comes from the server, so a late joiner sees the call's
age rather than their own arrival. Empty space rooms are destroyed already,
which is what makes the next call start from zero — no reset logic needed.
Voice stays get their own table rather than joining the audit log: that table
records points in time, a call is an interval, and pairing join/leave point
events would leave every query guessing at joins whose leave never arrived.
Sessions are opened and closed inside joinRoom/leaveCurrentRoom rather than at
the seven call sites that reach them, so no path can be missed, and
destroyRoom closes them too — it bypasses leaveCurrentRoom and would otherwise
leak open rows.
A restart leaves sessions open with an unknowable end time. They are closed at
startedAt, discarding that time rather than inventing it: crediting the gap
would hand someone hours they never spent, and the numbers are the point.
Mirrors the existing users.status sweep on boot.
Only closed sessions count, so a figure does not move on every refresh. Bars
scale to the leader, not the total — with five people every share of a total
looks identical. Statistics are readable by any member, since they are the
group's own numbers; the audit log, which names who did what, stays on
MANAGE_SPACE.
Records who changed what, and is the mechanism statistics will read — one
event table rather than two logs that drift apart.
The table is deliberately generic (action + target + JSON metadata) so a new
action needs no migration. Writes never throw: a kick must not fail because
its log entry could not be written, since the kick already happened.
Leaving is recorded as a different action from being removed. The same route
serves both, and a log that conflates them misleads exactly when it matters.
Actor is nullable with ON DELETE SET NULL: the event outlives the account, and
a log that vanished with its actor would be worthless. Reads are gated on
MANAGE_SPACE rather than a new permission bit, which would default to nobody
until every role was re-edited. Paging uses the snowflake id, stable even for
two events in the same millisecond, and an action this build does not know
still renders a row.
Favourites are stored server-side per user, so one made on the phone is there
on the desktop — the point of favouriting. The whole result is stored rather
than an id: the provider offers no lookup by id, so an id-only favourite could
not be rendered without re-finding it through search.
Category chips translate their label but not their query, which goes to a
provider that indexes in English.
The star sits beside the tile button rather than inside it: a button within a
button is invalid and swallows the click. Toggling is optimistic and reverts
on failure, and favourites skip both the loading skeleton and the infinite
scroll, which belong to provider-backed browsing only.
Server caps favourites per user and rejects non-http(s) URLs, which become
<img src> in everyone's picker.
OAuth Authorization Code flow, with tokens kept server-side: refreshing needs
the client secret, so the browser never holds a Spotify token — it asks this
instance what is playing and this instance calls Spotify.
The callback arrives as a plain browser redirect with no Authorization header,
so the OAuth state carries the user id signed with the instance secret and is
compared in constant time; without that, anyone could bind their Spotify
account to another user.
Activities are now tracked per producer. pushActivities replaced the whole
list, so the desktop game detector and Spotify would erase each other — losing
exactly the case this is for, a game and Spotify at once.
Polling backs off when the tab is hidden and keeps the last known track on a
network error rather than reporting 'stopped listening'. A rejected refresh
token (access revoked on Spotify's side) drops the row so the UI stops
claiming a live connection.
Scope is read-only: user-read-currently-playing and user-read-playback-state.
Per the fork's language rule, the new UI ships in en and pt-BR, and this
round also translates the privacy panel.
Nothing in the project was translatable — every string sat inline in English.
en.ts is the source dictionary and its type is derived from it, so a typo or a
missing key fails typecheck instead of rendering the raw key at runtime.
pt-BR.ts is deliberately Partial: translation proceeds one system per update
and anything absent falls back to English, so a half-migrated interface is
never broken, only partly English.
Locale is persisted, guessed from the browser on first run, and kept in sync
with <html lang> through a subscription — persisted state rehydrates after
first paint, so a one-off assignment would miss it.
Translates the voice input panel (including the mic test shipped earlier
today) and the profile card as this round's system. Language options are
labelled in the active language, so a wrong pick can always be undone.
The activity selector sat below `if (!isOpen || !user) return null`. `user`
arrives asynchronously, so the hook ran on some renders and not others; React
counts hooks per render and tore the tree down with error #310 as soon as a
profile finished loading.
Move it above the guard and let the selector tolerate a null user. Typecheck
and the suite both passed with the bug in place — TypeScript cannot see hook
order and nothing renders this modal across the null-to-loaded transition.
The git server runs as its own stack in /opt/gitea and is reached by container
name over the app's internal network, publishing no host port of its own. Only
the reverse-proxy entry belongs here, where the Caddyfile lives.
The activity pipeline was already complete end to end — Activity type, store,
WS broadcast, server validation, presence relay, and an ActivityCard used by
four list surfaces — but the profile card rendered none of it, which is the
'Listening to Spotify' block the design calls for.
Add ProfileActivity: richer than ActivityCard because the card has room for
artwork, track and artist, so it reads details/state/assets. All optional, so
it degrades to the bare name that today's process-based detector supplies.
Also scheme-check activity image assets server-side. activity.url was already
restricted to http(s) but assets.largeImage/smallImage were only length-checked
— an asymmetry that was harmless while nothing rendered them, and is not once
they become <img src>: a client could point them at a host it controls and
harvest the IP of everyone opening that profile.
The level meter only measured a stream a call had already opened, so settings
offered no way to check a mic before joining — the panel said as much.
Add startMicTest/stopMicTest on AudioManager: the processed input bus is
routed to the master output through a dedicated gain node, so the loopback can
be disconnected precisely. Settings had deliberately never opened the mic
itself; a mic test cannot honour that, so the test hands the mic back when it
stops.
Releasing needs two independent guards, because the user may join a call
mid-test: AudioManager only stops the exact stream it opened (identity check,
not a flag), and the caller must consent — the UI reads the call state, which
AudioManager cannot, as it does not import stores. Unmounting mid-test tears
the loopback down too.
The profile popout already existed and was reachable from eleven places —
messages, mentions, avatars, member list, DMs, activity panel — but no voice
surface opened it, so clicking someone during a call did nothing.
Wire it into the voice user rows (VoiceChannel's sidebar list) and the name
label on grid tiles, whose avatar was already a ProfileAvatar; the name beside
it not reacting read as the click failing.
Left mobile alone deliberately: MobileSpacesScreen already opens the profile
from its row wrapper, and MobileVoiceJoinSheet would layer a history-pushed
full-screen profile inside a bottom sheet, which cannot be verified here.
The composer's GIF button drew a filled rounded rect with the letters knocked
out, which reads as a solid square rather than a picker. Invert it: stroked
outline with filled letters, reusing the original glyph paths scaled to centre.
Banners already accept absolute URLs on both ends (server isValidAssetUrl
allows http(s); the profile render branches on banner.startsWith('http')), so
the picker stores the remote URL directly with no upload path. Previews can now
hold either a blob: or an https: URL, so revoking is guarded — calling
revokeObjectURL on a remote URL is a silent no-op that would hide a mistake.
The channel name under 'Voice Connected' was a plain div. Making it navigate
needed more than an onClick: voiceStore never recorded which space the call
was in, and spaceStore.channels only holds the space currently being viewed —
so after navigating away the call's channel was unresolvable, which is also
why the label degraded to a generic 'Voice Channel'.
Capture space and channel name at join time (the only moment they are
reliable) and use them for both the label and the jump. Covers space calls
and DM calls.
better-sqlite3@12.11.1 ships prebuilt binaries for ABI 127/137/141/147 only;
Node 20 is ABI 115, so prebuild-install falls back to node-gyp, which fails on
node:20-slim for lack of python3/make/g++.
Add the toolchain to the builder stage, and in the runtime stage install, use
and purge it inside a single layer so the final image ships no compiler.
The CLA Assistant only records a signature when the exact phrase is posted as a comment on the PR, so ticking the checkbox alone left the check red with no explanation (see #39). Spell the required comment out next to the checkbox, using the exact phrase from custom-pr-sign-comment in .github/workflows/cla.yml.
Avatar opened the profile popout whenever it received a user prop. Since user is how every avatar gets its gradient, colour and status dot, all 22 call sites became profile triggers by accident — including the picture inside the profile card itself, which re-anchored the card to that picture on every click and walked it across the screen (120px right, 36px down, until it pinned at the viewport clamp).
Avatar is now presentational. A new ProfileAvatar carries the open-the-profile behaviour at the five call sites that actually want it. The card's own picture escalates to the full profile modal instead of reopening the card.
The card also places itself off its measured size via the shared computeFloatingPosition engine, replacing six call sites that each hand-computed coordinates against a guessed 460px card height.
Closes#37
The CLA Assistant appends each signature as a direct commit to the branch named in 'branch:'. That was main, which the 'Require CI on main' ruleset rejects ('Repository rule violations found'), so signatures were never recorded and the check stayed red however often a contributor signed — a deadlock for every outside contribution.
Point the store at the cla-signatures branch. The ruleset targets the default branch only, so the bot can append there without granting any actor a bypass on main. That branch is seeded with the existing signature and carries its own ruleset blocking deletion and non-fast-forward pushes.
better-sqlite3 11.x removes its environment cleanup hook from Statement::~Statement() after the Node environment is torn down. Node 24.19.0 asserts on the null environment and aborts the worker, so vitest reported 'Worker exited unexpectedly' and exited non-zero on a fully green test run — failing the required check on every PR.
11.x also ships no prebuild for Node 24, so CI compiled it from source on every run. 12.x has prebuilds for that ABI and the V8 13.9 shims 11.x lacks. drizzle-orm declares better-sqlite3 >=7, so the major bump is in range.
Supersedes #26.
Completes the design record on main: Plan B's plan and the federation spec were
already here; this adds the umbrella security spec (source of truth for the
remaining container/web/desktop/remediation workstreams) and Plan A's plan.
Final whole-branch review (opus) fixes:
- docker-publish.yml: upload-sarif was if:always() but not continue-on-error, so a
Trivy SARIF-emit flake would fail the job and SKIP the multi-arch publish. Made it
non-blocking so a scanner hiccup never blocks a release.
- deployment.md: seed-admin-rotated.txt is root-owned (written via docker exec, which
bypasses the gosu drop) — reverted an over-correction. Corrected the canonical
runtime-stage build description (no toolchain; non-root gosu). First-boot chown note.
- restore.sh: comment ownership root -> uid 1000.
devsyncwrld
TheZwiss