A reset peer reaches needs_attention via the auth-failure path (HMAC desynced by
the new incarnation) without passing through unreachable, so the 5s recovery
probe never saw it — detection waited up to a full 15-min health-check cycle
before 'Re-peer & heal' surfaced. Extract detectResetForPeer() and fire it
event-driven at the transition, plus a startup sweep for already-stuck peers.
15-min tick remains the backstop.
Finding 1: handleReactionAdd/Remove now drop reactions on a dead 1-on-1
(isDeadOneOnOne) — previously a survivor could react on a Deleted-User
thread and the relay fanned out to all peers via undefined target-origins.
Client Message.tsx withdraws add/toggle reaction affordances for dead DMs
(existing reactions still display read-only).
Finding 2: dmMembership purge test now runs with foreign_keys=ON (matches
prod) and asserts dm_members/dm_messages cascade cleanup on channel purge.
Finding 3: tombstone group-DM ownership transfer filters isDeleted=0 so
ownership can never move to a tombstoned member; covered by a new test.
A reset peer can reach needs_attention via the auth-failure path (HTTP up,
401/403 from a new incarnation crossing AUTH_FAILURE_THRESHOLD) without ever
passing through unreachable, so the unreachable-only recovery probe never
observes its epoch change and no reset journal is created — leaving a later
manual Re-peer with nothing to heal.
Add detectResetOnNeedsAttentionPeers() to the 15-minute health-check tick:
probe needs_attention peers with a non-null baseline (excluding those already
peer_reset_detected) and call markPeerReset on an observed epoch mismatch.
Detection only — never recovers a needs_attention peer to active; baseline
(peer_instance_id) and hmac_secret untouched.
Add healResetIncarnation (federationReset.ts): fires from onPeerActivated after
an authenticated re-peer to soft-tombstone the flagged pure S2S stubs of a reset
peer's dead incarnation, clearing stale friendships/DMs so the reported bug is
fixed. Two mandatory guards: a reason gate (allow-list of 8 genuine handshake
activation reasons; excludes health_check_recovery + startup_bootstrap so their
stale baseline can never silently resolve a journal without healing) and an
epoch comparison (dead_epoch === newEpoch => false alarm, no tombstone). Uses
tombstoneUser(uid, { purgeContent: false }); real federated accounts are left
flagged + intact for Phase 2. Runs outside any transaction. Wire into
onPeerActivated before the mutation-log re-sync.
Two mirror-image bugs from voice/DM-call transitions leaving stale state.
DM call → space channel (stuck "Connecting…"):
The last participant to leave a DM call for a space channel receives a
`dm_call_ended` echo (server empties the DM room on their `voice_join`).
The handlers called `disconnectFn()` unconditionally, tearing down the
space room they had just connected to. Route `dm_call_ended` /
`dm_call_rejected` / terminal `dm_call_undeliverable` through a new
`teardownDmCall()` that only disconnects LiveKit when not in a space
channel (`currentVoiceChannelId` null).
Space channel → DM call (still shown as "in" the voice channel):
1. Entering a DM call never cleared `currentVoiceChannelId`, so
`VoiceChannel` mapped the DM call's live LiveKit participants onto the
old space channel. Add `clearSpaceVoiceForDmCall()`, called in
`connect()` when `isDm`, restoring the invariant that a DM call has no
`currentVoiceChannelId`.
2. `dm_call_accepted` gated the caller's connect on `!isLiveKitConnected`,
so a caller already in a space channel was never connected to the DM
room. Gate on `wasOutgoingCall` only (connect() de-dupes same-room).
Tests: teardownDmCall.test.ts, clearSpaceVoiceForDmCall.test.ts.
Docs: docs/systems/voice.md.
Replace the placeholder Screenshots section with a hero voice/video shot
and a curated 6-shot grid (chat, screen-share controls, spaces, federated
DMs, social discovery, federation admin), linking to a new
docs/screenshots.md gallery of all 14 shots grouped by feature.
Images are optimized to 1600px WebP (~1.2MB total vs ~45MB raw); raw
originals in demo-pictures/ are gitignored.
The server's `playable` flag is computed Chromium-first, but HEVC
web-playability is browser-dependent: WebKit (Safari on macOS/iOS) decodes
HEVC via the OS while Chromium/Firefox/Electron can't. Treating the flag as
global wrongly showed Safari users the download fallback for files they can
play inline.
The client now treats `playable === false` as "needs a capability check": it
pre-renders the fallback only when the current browser also can't decode the
format, gated on a one-time canPlayType probe (BROWSER_SUPPORTS_HEVC). Capable
browsers attempt inline playback; the <video> onError handler remains the
safety net for genuine failures.
macOS screen recordings are HEVC inside a .mov container, which Chromium,
Firefox and stock Electron can't decode. The file uploaded fine and a
server-side ffmpeg poster was generated, but inline <video> playback failed
silently — stuck at 0:00 with no error, since AttachmentRenderer had no error
handling. Root cause: the system had no concept of web-playability.
Server detects, client degrades:
- mediaPlayable.ts: classifyVideoPlayable(mimetype, codec) — tri-state
(false = known-undecodable e.g. HEVC/ProRes, true = web codec in web
container, null = unknown/optimistic). Never widens `false` beyond codecs
that fail everywhere, so ffmpeg-less instances keep prior behaviour.
- probeMediaMeta now captures the video codec_name; the upload finish hook
stores the verdict in the new attachments.playable column (migration 0007).
- Flag propagated through every serializer: space messages, DMs, WS, and
federation relay (outbound + inbound) — federation-compatible.
- VideoAttachment component: playable===false renders a download card (poster
+ "Can't play here — download" + name/duration/size) with no dead-player
flash; otherwise plays inline with an onError fallback to the same card.
Specs updated: uploads.md, database.md, federation.md.
Voice presence (voiceStates/voiceUserStates/spaceVoiceStates) was only ever
delivered in the WS `ready` payload — i.e. at connect/reload. A user joining a
space mid-session got `member_joined` (no voice data) and a bare space object;
`GET /api/spaces/:id` (the channel-sidebar hydrator) carries no voice state
either. So members already sitting in a voice channel stayed invisible in the
new member's sidebar until a full page reload.
Fix at the systemic root: ConnectionManager.addUserSpace — the single chokepoint
every join path funnels through (invite, public join, join-request approval),
and which is NOT used on reconnect (that path uses setUserSpaces) — now pushes a
scoped `space_voice_state` snapshot to the joining user. The snapshot is built by
a new buildSpaceVoiceState(spaceId, userId) helper that is also the single source
of truth feeding buildReadyPayload (refactored to use it), so the connect-time
and join-time paths can never drift.
Robustness:
- Delivered over the same ordered WebSocket as voice_state_update deltas — no
REST snapshot-vs-event-stream race.
- VIEW_CHANNEL-filtered via computePermissions exactly like `ready`: a joiner is
never told who occupies a voice channel they cannot see.
- Client applies it scoped to the space (utils/voiceStateSync.applySpaceVoiceState):
merges occupants/statuses and rebuilds only that space's restriction keys,
never disturbing voice state in other spaces.
- Skipped when the space has no active voice and no restrictions (e.g. space
creation).
Tests: server helper behavior, the join push, and private-channel exclusion;
client scoped-apply. Specs updated (websocket.md, voice.md, spaces.md).
The sidebar's visibleChannels filter is keyed on the channelPermissions
Map. Creating a channel raced two state updates: the optimistic create
(added to channels with no permission entry) and the channel_created WS
event (the only thing that set the permission). When the optimistic add
won the race, the WS handler hit its dedup guard, skipped setChannels,
and set the permission by mutating the Map in place — no new reference,
so visibleChannels never recomputed and the channel stayed hidden until
loadSpace rebuilt the maps (i.e. leaving and returning to the space).
Centralize the logic in a new upsertChannel store action that replaces
channels and channelPermissions with fresh references, used by both the
create path and the channel_created handler. Also return the creator's
computed myPermissions (and isPrivate) from POST so the channel renders
immediately from the response, independent of WS timing.
Adds spaceStore.upsertChannel.test.ts covering the reference-identity
regression and the optimistic-reconcile path.
Replace ripped Discord audio with self-authored Ogg Vorbis files (avoids
copyright on open-source launch) and fix three cue bugs:
- Anti-pop envelope: playSound now applies a 10ms fade-in (+ tail fade-out
for one-shots) so buffers no longer start at a non-zero amplitude. Kills
the pop on the looping call cues. Mirrors playTestTone.
- Deafen no longer plays mute+deafen at once: toggleDeafen flips isMuted and
isDeafened atomically, so SoundController saw both transitions. New tested
pure helper selectVoiceStateSound() suppresses the side-effect mute cue.
- Viewer-side watch feedback: stream_user_joined/left now play locally on the
watcher's own machine for explicit Watch/Stop actions, via
handleViewerWatchToggle. Direct local playback (not a watchingStreams diff)
keeps auto-teardown silent and works identically on Safari and Electron.
Docs: docs/systems/sounds.md updated (Ogg, dual-audience cue rows, mute/deafen
selection, viewer feedback, anti-pop envelope).
Eliminates the visible push-down/snap-back when the pagination loading
skeleton appears at the top of the message list. Achieved by rendering a
constant-height (~200px) top-of-list slot above messages whenever
hasMore===true; the skeleton's grey-bar contents toggle inside that slot
rather than the slot itself mounting/unmounting. Companion changes:
load-more trigger raised to fire before the slot enters the viewport with
an iOS Safari rubber-band guard, prepend scroll-restore math corrected to
handle non-zero prevScrollTop, useDelayedLoading threshold lowered to
50ms for pagination only.
Brings the parked mobile voice + screenshare polish from wip/mobile-polish
(commit dbb9b2c) into main. Conflicts in MainContent.tsx and mobile-ui.md
expected per the WIP commit message — resolved manually.
Each AvatarTile rendered at `size × size` with a 2px border under
`box-sizing: border-box`, giving a content area of `(size − 4) × (size − 4)`.
The inner `<Avatar size={size}>` exceeded the padding box, and the
`overflow-hidden + rounded-full` clip — centered on the wrapper — combined
with the avatar contents anchored at the padding-edge top-left to displace
photos and especially the centered initials gradient + letter toward the
lower-right of the visible disc, leaving a sliver of background opposite.
Compounding this, `Avatar`'s `inline-flex` root sat on the line-box text
baseline, so any inherited `line-height ≠ 1` (the DM list inherits the
row's line-height) drifted the avatar a further several px vertically.
Two corrections:
• Size the inner Avatar to `size − 2·TILE_BORDER_WIDTH` (matches the
padding box) — extracted as a constant so the dependency between
`border-2` and the inner size is visible.
• Center geometrically via `flex items-center justify-center` on the
wrapper, bypassing inline-flow placement so the Avatar is anchored
regardless of inherited type metrics.
Verified in the live app (DM sidebar, chat header, welcome header) and
with a 4-tile diamond at sizes 32 and 80: visible offset is now exactly
the border width on every tile, letters/photos sit dead-center.
Adds a regression test that pins both invariants (flex centering classes
present + inner Avatar style.width === tileSize − 4) so a future change
that re-introduces the bug fails fast. 12/12 AvatarStack tests, 365/365
web tests, typecheck clean.
design-system.md spec updated with the AvatarTile geometry contract.
Manual ownership transfers between two federated instances diverged because
`dm_channels.ownerHomeInstance` was stored as a BARE host (`orbit.ddns.net`)
for federated owners — via `transferGroupDmOwnership` copying `users.homeInstance`
verbatim — while `sourceInstance` always arrives as a full URL on the wire.
`processOwnershipTransferEvent` and `processMemberRemoveEvent` then compared the
two with strict equality and rejected legitimate inbound events as
`unauthorized_source`, keeping ownership permanently divergent across peers.
Live DB inspection on the two test instances confirmed both rows (nova + orbit)
had a BARE `owner_home_instance`, matching the bug report exactly.
Three compounding fixes:
1. Receiver authority checks now compare via `normalizeOriginForCompare` so
legacy bare-vs-full rows accept legitimate transfers (and kicks).
2. New `canonicalizeHomeInstance` helper in `federationAuth.ts`; every write
site that persists `ownerHomeInstance` (`transferGroupDmOwnership`, group DM
creation, lazy federation in member-add, `processMemberAddEvent` bootstrap,
`processOwnershipTransferEvent` receiver storage) routes through it. Full URL
is the canonical storage form, matching how `sourceInstance` arrives.
3. `dm_owner_updated` WS event extended with optional `newOwnerHomeUserId` and
`newOwnerHomeInstance` fields. Client `updateDmOwner` writes them when
present and leaves existing values untouched otherwise (legacy-server safe).
Without this, `getOwnerInstanceForDm` returned the previous owner's home
after a successful WS broadcast, routing the next owner-only op to the wrong
instance.
Coverage: new `federation.ownershipTransfer.test.ts` (7 receiver tests including
the headline bare-vs-full regression and the dedup replay guard); new bare-vs-full
case in `federation.kick.test.ts`; two new client-side cases in
`groupDm.ownerRouting.test.ts` covering both the extended-payload write path and
the legacy-server passthrough. Tests: 1053 server + 364 web, all green.
Specs updated: `dm-system.md` historical bugs + frontend handler table + WS
state-change events table; `federation.md` `ownership_transfer` receiver flow;
`websocket.md` event-fields table.
Transferring ownership or kicking a member surfaced "Target user is not a
member of this DM channel" whenever the target was a federated user.
Root cause: the client passed `canonical.id` from `useCanonicalUserView`,
which returns the user's HOME id when the home view is in the userViews
cache. After owner-routing the request to the owner instance, that
instance's `dm_members.userId` (its own local replicated id) never
matched the home id, so `isDmMember` returned false. The same failure
mode applied across any cross-instance scenario where the
channel-serving instance and the owner-serving instance disagree on the
local replicated user id for the same federated user.
Fix: both endpoints now accept federated identification, mirroring the
existing pattern on `POST /api/dm/:id/members`:
- `POST /api/dm/:id/transfer` body: `{ newOwnerId? } | { homeUserId, homeInstance }`.
Federated args win when both are supplied (strictly more specific).
- `DELETE /api/dm/:id/members/:targetUserId` reads optional
`?homeInstance=<origin>` query; when present, the URL segment is
treated as a homeUserId and resolved via `resolveOrCreateReplicatedUser`.
Client `api.dm.kickMember` and `api.dm.transferOwnership` gain an
optional `federated` argument; `DmRosterPanel` and `MobileGroupDmInfo`
pass it whenever the target has `homeUserId` + `homeInstance` populated.
Adds 5 server tests (3 transfer + 2 kick) covering federated targets,
the federated-wins-over-local precedence rule, and federated-non-member
rejection. Updates 2 client routing tests and 2 DmRosterPanel test
assertions for the new signature. Updates `docs/systems/dm-system.md`
and `docs/systems/api.md`.
Server: 965 tests pass (was 960). Web: 362 tests pass (was 360).
The old 2x2 grid for 3+ members produced a cramped, misaligned look
(circles cut off inside their own border, no overlap, just four small
disks in a grid). The 2-member overlap aesthetic — equal-size tiles
that overlap diagonally — was the visual gold standard but only
existed for that one count.
Extend the same "huddle of overlapping faces" aesthetic to all
member counts:
• 3 members → equilateral triangle of 62%-size tiles (top, bl, br)
• 4 members → diamond of 58%-size tiles (top, right, bottom, left)
• 5-10 members → diamond with `+N` overflow occupying the bottom
slot (z-index boosted above neighbors so the digit is never clipped)
Tiles are positioned radially around the box center on a circle of
radius (S − T) / 2 so the farthest edges of each tile graze the
bounding rect — no clipping, no wasted whitespace. Z-index descends
clockwise from the top slot so each tile tucks slightly under its
clockwise neighbor, mirroring the 2-member z-stack.
The 0/1/2-member cases and the icon-override branch are unchanged.
Props interface unchanged — no call-site updates needed.
Tests updated to assert the new layout markers (`triangle`, `diamond`)
and to verify the geometry (top tile in triangle has smallest `top`;
overflow tile in diamond has largest `top`). 11/11 AvatarStack tests
pass; 360/360 web tests pass; typecheck clean.
design-system.md spec updated with the new layout table + geometry
section.
Jannis Braun