Three fixes for multi-instance call state consistency:
1. Client dm_call_accepted handler only auto-connects to LiveKit if
the user was the caller (outgoingCall was set). Other instances of
the same user just clear ringing state without connecting.
2. Server processDmCallAcceptEvent remote path skips duplicate
broadcast when FederatedCallEntry is already active (prevents
state conflicts from host fan-out arriving after local accept).
3. Ready payload handler clears stuck incomingCall when restoring
an already-active call after page refresh.
When findOrCreateDmChannel creates or finds a local channel for a
federatedId that has an active FederatedCallEntry with null dmChannelId,
update the entry. Prevents stale null references for Path B calls.
Three bugs that combined to corrupt DM identities during initial sync:
1. Sync endpoint omitted federatedId for group DMs, causing the receiver
to treat them as 1-on-1 DMs and compute wrong pair hashes — creating
phantom channels that collide with real ones.
2. backfillHomeUserId unconditionally overwrote existing homeUserIds,
so a single wrong match would permanently corrupt a user's identity
and cascade to all subsequent lookups.
3. Migration duplicate-stub Criterion 1 ("shared 1-on-1 DM membership")
incorrectly merged different users from the same domain who were
simply having a conversation, destroying one user's identity.
The relay broadcast loop skipped members whose homeInstance matched the
source instance, assuming they already received the message on their home
server. This broke delivery for federated users (e.g. youruser@nova on
orbit) who are actively connected to the remote instance.
Client-side dedup in addRealtimeMessage already handles double delivery
via sourceMessageId cross-matching, making the server-side skip both
unnecessary and harmful.
Queue dm_close relay in DELETE /api/dm/:id after broadcasting locally,
and queue dm_reopen relay inside the myDm.closed === 1 branch of POST
/api/dm so peers are notified only on actual reopen, not new DM creation.
Receive read_state_update events from peers, translate message
coordinates to local IDs via sourceInstance/sourceMessageId
mapping, update read_states with timestamp-only LWW.
profileUpdatedAt ?? createdAt treated freshly registered users as having
"newer" profiles than users with real edit history, because createdAt is
always NOW at registration time. This broke federation profile sync:
the client correctly pushed home → remote, but the remote server's LWW
guard rejected the write (stored createdAt > incoming profileUpdatedAt).
A null profileUpdatedAt means "never edited" — that's timestamp 0, not
the user's creation time.
After deleting a federated identity, the server-side user_federation_registry
and users.replicated_instances were not cleaned up, causing "already connected"
errors when trying to re-federate. The deletion endpoint now authoritatively
removes both the registry row and the replicatedInstances entry, and bumps the
LWW timestamp to prevent stale client syncs from re-inserting them.
Also extends the endpoint to accept mode 'leave' (skip S2S, just clean up),
and enables the "Select instances..." scope option in DeleteIdentityDialog.
Switch from manual space-ID collection to collectDeletionBroadcastTargets and
add user_updated broadcast so clients patch their caches when a federated user
is deleted via S2S. Force-disconnect moved after broadcasts so other tabs
receive events before the connection is torn down.
After a prior deletion + re-federation, multiple user records share
the same homeUserId (one deleted, one live). The unfiltered .get()
returned the older deleted record, causing the idempotency check to
short-circuit and miss the live record entirely.
When a federated user's identity has been tombstoned (isDeleted=1),
findFederatedUser filters them out, causing resolveOrCreateReplicatedUser
to silently create a new stub — a "zombie" resurrection. This guard checks
for a deleted row before creating a stub and returns null instead.
All 11 call sites across federation.ts and dm.ts have been updated with
appropriate null guards: federation relay handlers reject or skip the event
(participant_not_found / accepted no-op), while dm.ts routes convert null
to undefined so the existing 404 path handles it.
Adds persistent federation registry storage endpoints under
/api/users/@me/federation-registry. GET returns all registry entries
with the stored updatedAt timestamp. PUT replaces the full registry via
LWW (409 on stale write) in an atomic transaction.
processDmTypingStartEvent and processDmTypingStopEvent handle
typing indicator relay from peers. Uses federatedId for channel
lookup, resolveLocalUser for ephemeral identity (no stub creation).
Also clears typing indicator in processCreateEvent when a relayed
message arrives — belt-and-suspenders for dropped relay packets.
sendTypingRelay() mirrors sendCallRelay() — direct POST to peers,
no outbox, no retry. Uses federatedId for cross-instance channel
identification. Wired into handleDmTypingStart() for typing_start
and broadcastDmMessage() for typing_stop.
Broadcasts dm_typing_stop to DM members before dm_message_created,
so the typing indicator clears immediately when a message arrives
instead of lingering for up to 3 seconds after delivery.
Extend the add-member endpoint to resolve federated identity via
resolveOrCreateReplicatedUser() when homeUserId+homeInstance are provided,
falling back to the existing local userId lookup.
POST /api/dm created channels with federatedId=NULL, so when the S2S
reply arrived, processCreateEvent couldn't find the channel and created
a duplicate. Now computes the deterministic SHA256 hash at creation time
when either participant is federated.
Client-federation users (e.g., youruser@nova logged into orbit)
send DMs on the remote server. The S2S relay forwards these back to the
author's home instance, but verifyAttribution rejected them because the
author's homeInstance didn't match the sourceInstance.
Now also accepts when the author's home matches the receiving instance
(getOurOrigin()), covering the homeward relay case.
The outbox worker interval was reduced from 10s to 1s in FED-009,
so a busy sender can now hit 60 req/min during sustained traffic.
90 gives 50% headroom.
- Resolve homeUserId from DB in sendFederatedCallStart/End (not raw userId)
- Clear existing timeout in createFederatedCall before overwriting
- Clear federatedCallToken/Url in leaveVoice and handleForceDisconnect
- Remove unnecessary `as any` cast in relay processor
- Fix race window: store pendingHmacSecret AFTER remote peer confirms,
not before (admin endpoint + auto-rotation worker)
- Add hex validation on newSecret at /peer/rotate endpoint
- Use pending-secret-aware signing in initial sync worker
- Add test for corrupt state (pendingHmacSecret set, secretRotationAt null)
Jannis Braun