initialTopMostItemIndex was a plain index (top-aligned), so react-virtuoso painted with estimated row heights then corrected scrollTop after measuring the real (taller) dynamic bubbles — a visible jump on every chat switch. Use { index: 'LAST', align: 'end' } to pin the bottom edge instead, matching react-virtuoso's canonical chat pattern; the restore-to-saved-row path stays align: 'start'.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Move Supabase + LiveKit from the netralax.cloud VPS to a new netralax.de server. Adds the migration runbook (docs/), one-time move scripts (scripts/migrate/), and prod Caddy/LiveKit config templates (infra/). Repoints the desktop publish/changelog URLs and prod ops config to .de. JWT_SECRET + VAPID copied identically so already-installed clients keep working; the new server also serves the legacy .cloud hostnames.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Friend-DM "Nachricht nicht lesbar" recurred even after v0.21.3 because the
proactive sweep called shareConvKeyToUser, which reads the module-level
conv-key cache first. After a server-side cleanup the cache still held the
stale locally-bootstrapped key, so each side wrapped its own different key
for the peer and the bundles diverged anew.
Switch the sweep to rotate_conv_key when any peer's user-id bundle is
missing at the active version: a fresh symmetric key is generated, wrapped
for every member at their CURRENT pubkey, and the active version is bumped
under a row-level FOR UPDATE lock. Concurrent rotations are race-safe — the
loser sees "new version must be greater" and bails; the winner's bundles
propagate via realtime.
Realtime conversation_keys subscription now invalidates the cache for the
affected (conversationId, key_version) on any INSERT/UPDATE/DELETE — so
admin cleanups, peer rotations, or device wraps can no longer leave a
stale entry in this client's session cache.
queueInsert now fires the first event of a quiet period immediately and
only collapses follow-up bursts. BATCH_WINDOW_MS dropped 250 → 80 ms.
This closes the ~250 ms gap between the notification sound and the
message body appearing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lifts the `state.kind === 'connected'` guard from the soundboard hotkey
useEffect so OS-level shortcuts are always registered. Inside a call the
existing `playSoundboard` path routes audio into the LiveKit pipeline so
peers hear; outside a call the new `playSoundboardLocal` helper fetches
the blob via `getSoundBlob`, creates a short-lived object URL, and plays
through a fresh HTMLAudioElement on the system default output.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Audit of write-actions revealed that send (and edit via realtime UPDATE) are
already optimistic via local state insertion in `useConversationMessages`,
and friend nicknames are pure-local localStorage. Five user-write actions
were waiting on the ~100-200 ms server roundtrip + realtime echo:
* Toggle mute (`setConversationMutedUntil`)
* Toggle mentions-only (`setConversationMentionsOnly`)
* Toggle archive (`setConversationArchived`)
* Pin / unpin message (`pinMessage` / `unpinMessage`)
* Revoke device (`revokeDevice` RPC)
All five now flip local state synchronously and roll back on failure. The
existing realtime subscriptions reconcile canonically (no-op when the
optimistic patch already matches the server row), so this is purely a UX
latency improvement — no protocol or persistence changes.
Reactions (`toggleReaction` / `voteExclusive`) were intentionally skipped
this round: rollback semantics for the exclusive-vote path with multiple
sequential awaits are messy enough to warrant a dedicated pass.
byGalax