mirror of
https://github.com/J3vb/OwnCord.git
synced 2026-09-03 03:50:00 +03:00
* fix(identity): 1 defect(s) (OC-0151)
* fix(ws): 1 defect(s) (OC-0152)
* fix(admin): 1 defect(s) (OC-0153)
* fix(admin): 1 defect(s) (OC-0154)
* fix(voice): 2 defect(s) (OC-0155, OC-0167)
Replace distributeRoomKey's per-call offer counter with an instance-level
sliding-window budget shared by every voice_e2ee_offer send path.
- OC-0155: back-to-back rotations (the second run immediately by
drainPendingRotationOrArmTimer) each got a fresh pacing budget, so their
combined sends could exceed the server's single per-second cap.
- OC-0167: handleAnnounceInner's drain-time offer send bypassed pacing
entirely, letting a key holder joining a large ongoing call burst every
queued announce's offer unpaced.
The shared budget is reset in clearState() since the server's limit is
scoped per (sender, channel).
* fix(client): 1 defect(s) (OC-0156)
createPresenceSender dropped a queued custom_status when a later plain
status change superseded the pending retry. The retry now carries the
last committed custom_status forward.
* fix(client): 2 defect(s) (OC-0160, OC-0163)
OC-0160: exempt the handshake frames (ready, auth_ok) from the ws message
size limit and run the guard after parsing. A 'ready' frame grows unbounded
with member/channel/DM counts and carries no seq, so dropping it left the
client on empty stores with no error and no recovery path.
OC-0163: bracket a bare IPv6 host when building the wss:// URL so the
authority parses, and collapse bracketed/bare IPv6 literals to the same
cert_store_key so one server is not pinned (and user-confirmed) twice.
* fix(voice): 1 defect(s) (OC-0162)
updatePttKey armed the Rust poller when a PTT key was bound mid-call but
never applied the gate. The poller only emits 'ptt-state' on a press/release
transition, so an idle key produced no event and the already-published mic
stayed hot until the user's first physical press+release. Mirror the join-time
gate computation in updatePttKey, guarded on being in a call, polling actually
being live, and the mic not already being gated.
* fix(client): 1 defect(s) (OC-0164)
* fix(plugin): 1 defect(s) (OC-0165)
scanPluginDirectory now skips a malformed plugin subdirectory and joins its
error instead of aborting the whole scan, and LoadAll logs-and-continues so
one bad plugin directory cannot disable every other plugin.
* fix(ws): 1 defect(s) (OC-0166)
Route PresenceSelfEvent onto the owner's normal-priority queue instead of
letting it fall through to the UserTargetedEvent high-priority case, so a
user's own presence frames all share one FIFO and cannot be delivered out
of order relative to the visible presence_update path.
* fix(db): 1 defect(s) (OC-0168)
* fix(client): 1 defect(s) (OC-0169)
* fix(client): 1 defect(s) (OC-0171)
addMessage appended a broadcast at the tail even when trailing optimistic
rows were still unreconciled, so a message that committed while our own
send was in flight ended up ordered behind the row confirmSend later
stamped with a higher server id/timestamp. Insert before the trailing
unreconciled run instead.
* fix(voice): 1 defect(s) (OC-0172)
* fix(client): 1 defect(s) (OC-0174)
* fix(ws): 1 defect(s) (OC-0175)
* fix(client): 1 defect(s) (OC-0177)
* fix(client): 1 defect(s) (OC-0178)
* fix(voice): 1 defect(s) (OC-0179)
Undeafening no longer sends a voice_mute{muted:false} the server will
refuse while a moderator-imposed mute stands, matching the localServerMuted
guard already present in onMuteToggle.
* fix(client): 1 defect(s) (OC-0182)
* fix(plugin): 1 defect(s) (OC-0183)
* fix(client): 1 defect(s) (OC-0184)
Treat a trailing underscore as an emphasis delimiter, not part of the URL,
when scanning for the end of an autolinked URL.
* fix(client): 1 defect(s) (OC-0185)
Reveal .msg-actions-bar on .message:focus-within, not only on hover, so
keyboard users can see the per-message action buttons they Tab into
instead of activating them at opacity: 0.
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* fix(client): 1 defect(s) (OC-0186)
* fix(client): 1 defect(s) (OC-0187)
The Add Server modal validated addresses with its own narrower regex that
never gained IPv6 support when api.ts's validator did, so an IPv6 server
could be logged into but never saved as a profile. Extract the validator
into src/lib/hostValidation.ts and use it from both call sites.
* fix(client): 1 defect(s) (OC-0189)
DM sidebar rows dropped mention counts entirely and the header total
excluded muted conversations outright, so a direct mention in a muted DM
was invisible. Render a mention badge that outranks the plain unread
badge, and count a muted channel's mentionCount toward the header total.
* fix(client): 1 defect(s) (OC-0190)
* fix(client): 1 defect(s) (OC-0191)
* fix(client): 2 defect(s) (OC-0157, OC-0176)
* fix(client): 1 defect(s) (OC-0161)
confirmTotp answers 401 for a wrong enrollment code while the session is still valid; firing the global onUnauthorized sink signed the user out and deleted their stored credential. Opt that one call out via a skipUnauthorized flag on doFetch.
* fix(admin): 1 defect(s) (OC-0173)
* fix(identity): 1 defect(s) (OC-0180)
* fix(admin): archived channel PATCH skips voice eviction and fan-out (OC-0158)
handlePatchChannel commits the AdminUpdateChannel write, then re-reads the
channel to drive voice eviction and the visibility fan-out. When that
post-commit re-read failed, the handler returned early: the archive was
durable but connected clients were never told and voice members were never
evicted, leaving users talking in a channel that no longer exists for them.
Drive the post-commit work off the values already in hand rather than
abandoning it when the re-read fails.
Adds SetPatchChannelPostCommitHook so the test can land a cancellation in
that exact window deterministically instead of racing wall-clock timing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* fix(admin): role changes commit with no client ever notified (OC-0170)
broadcastRoles derived its context from the inbound *http.Request, so the
roles_update fan-out was tied to the request lifetime. A role create,
update, or delete could commit to the database and then broadcast nothing
once that request context was done, leaving every connected client on a
stale role list until the next full resync.
Decouple the fan-out from the request context so the broadcast follows the
commit rather than the caller.
Adds BroadcastRolesForTest to reach broadcastRoles from the external test
package.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* fix(client): username rename stomps the profile card header (OC-0188)
The account profile card's header is a resolveDisplayName() slot, but the
username-rename save path wrote the raw username straight into it. A user
with a display name set would see the header switch from their display
name to their new username after a rename, disagreeing with every other
surface that renders the same identity.
Resolve the header through the same display-name path the initial render
uses, so a rename updates the username field without touching the header.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* fix(client): settings overlay never focuses when mounted already-open (OC-0181)
mount() synced initial state — including the show() that calls
focusDialog() — before appending root to the container. .focus() on a
still-detached subtree is a silent no-op, so a caller that mounts while
uiStore.settingsOpen is already true (ConnectPage's lazy first-open path)
got a visible overlay whose focus trap never captured focus: keyboard
users landed outside the dialog with Tab escaping to the page behind it.
Attach root before syncing initial state so focusDialog() runs against a
connected subtree.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* chore: satisfy the CI gates for this fix batch
The fix batch's own commits left three CI gates red. Nothing here changes
behaviour; every edit is a lint, type, or formatting correction to code
this batch introduced.
golangci-lint:
- OC-0153 and OC-0173 replaced the last two uses of admin's setupSanitizer,
and OC-0151 the last use of api's sanitizer, leaving both package-level
bluemonday vars unused. Remove them along with the now-unused imports,
and reword the comments that named them so they still explain why the
fixpoint sanitizer is the right one without pointing at deleted symbols.
- Modernize the new handshake-deadline test's loop to range-over-int.
tsc --noEmit:
- jsdom ships no types and @types/jsdom is not a dependency, so declare the
surface the new admin-panel test uses, following src/types/jitsi-rnnoise.d.ts.
- Narrow the last-call lookup instead of indexing under
noUncheckedIndexedAccess, with an explicit failure message.
- membersStore.setState replaces whole state, so the presence-sender mocks
must supply typingUsers.
prettier: reformat the five files this batch touched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* chore(ledger): record the 2026-08-19 hunt and its fixes
Adds the 41 findings confirmed by the 2026-08-19 hunt and marks the 40
fixed on this branch, each with its commit, the test that pins it, and
revertProof "pass".
"pass" means an independent check, not the fixing agent's self-report:
every commit had its source diff reverted against the working tree, its
own test re-run and required to FAIL, then the source restored and the
test required to PASS. Commits whose tests live inline in Rust
#[cfg(test)] blocks were proven the same way at hunk level, splicing the
pre-fix source onto the post-fix test module.
OC-0159 is recorded as a duplicate of OC-0152: the flow-reconnect and
flow-message lenses independently found the same unbounded handshake
write and proposed the same helper over the same call sites.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
* test(e2e): make the voice-roster join fixture self-consistent
The voice-widget join test emitted a voice_state for user_id 4 claiming
username "newvoiceuser", but id 4 is "member2" in MOCK_MEMBERS_MULTI_ROLE.
A real server never sends a voice_state whose username disagrees with the
member record for that id, and the same file's VOICE_STATE_EVENT already
pairs id 1 with "testuser" correctly — this one event was the outlier.
The contradiction was invisible while the roster rendered the payload's
raw username. OC-0177 makes it resolve identity through membersStore so a
nickname shows the same in voice as everywhere else, at which point the
fixture's own inconsistency surfaced as a failure.
Send id 4's real username and assert on it. The test still covers what it
did before — a genuine join by a user not previously in voice, asserted by
name and by roster count.
Verified against the app unchanged: with the old fixture the spec fails
1/5 (matching CI), with this one it passes 5/5.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M6gVN2JM5wrduhkNaFCxdK
---------
Co-authored-by: Claude <noreply@anthropic.com>
133 lines
6.6 KiB
Go
133 lines
6.6 KiB
Go
package ws
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
"log/slog"
|
|
)
|
|
|
|
// EmitEvents routes typed events to the appropriate broadcast methods.
|
|
// Called from readPump goroutines after a V2 handler returns. ctx carries the
|
|
// dispatching connection's cancellation for the routes that hit the database.
|
|
//
|
|
// CRITICAL ordering: SequencedDMEvent MUST be checked before ChannelEvent
|
|
// because DM events implement both interfaces. The SequencedDMEvent path
|
|
// calls sendSequencedToUsers which preserves the seqMu serialization guarantee.
|
|
//
|
|
// VoiceChannelEvent MUST be checked before ExcludeSenderEvent because voice
|
|
// events implement a superset of ExcludeSender semantics but target by voice
|
|
// channel membership rather than channel focus.
|
|
func (h *Hub) EmitEvents(ctx context.Context, events []Event) {
|
|
for _, ev := range events {
|
|
switch e := ev.(type) {
|
|
case SequencedDMEvent:
|
|
// Normal priority: sequenced frames must share the per-client FIFO
|
|
// so the max-seq ack watermark never passes an undelivered event.
|
|
h.sendSequencedToUsers(e.ChannelID(), e.ParticipantIDs(), e.Payload())
|
|
case VoiceChannelGuardedEvent:
|
|
h.sendToUserIfInVoiceChannel(e.VoiceChannelID(), e.TargetUserID(), e.Payload())
|
|
case VoiceChannelEvent:
|
|
h.sendToVoiceChannelExcept(e.VoiceChannelID(), e.ExcludeUserID(), e.Payload())
|
|
case ExcludeSenderEvent:
|
|
// An invisible user's public presence half rides this branch;
|
|
// like the visible case below, it must invalidate any queued
|
|
// coalescer entry so a stale connect-time presence can't flush
|
|
// after (and overwrite) this fresher user-chosen status. The
|
|
// drop and the broadcast run atomically under presenceMu (see
|
|
// dropQueuedPresenceAndBroadcast, OC-0005) so a flush racing in
|
|
// at the same moment can never enqueue its stale frame after
|
|
// this fresher one.
|
|
if po, isPresence := ev.(PresenceOthersEvent); isPresence {
|
|
h.dropQueuedPresenceAndBroadcast(po.excludeUserID, func() {
|
|
// Normal priority, excluding the owner — NOT
|
|
// broadcastExcludeLow. Every other source of this same
|
|
// user's presence (connect/disconnect via BroadcastToAll,
|
|
// and the visible presence_update path below) already
|
|
// shares the normal-priority queue; putting this one on
|
|
// the low-priority queue instead split one user's
|
|
// presence across two per-client FIFOs with different
|
|
// durability (low silently drops on overflow instead of
|
|
// disconnecting, so no replay ever repairs the loss) and
|
|
// different drain order (writePump always drains normal
|
|
// strictly before low, so a newer frame on one queue can
|
|
// be delivered before an older frame still sitting on the
|
|
// other) — exactly the hazard OC-0214 fixed for the
|
|
// visible case below (OC-0003).
|
|
h.BroadcastToAllExcept(po.excludeUserID, e.Payload())
|
|
})
|
|
} else {
|
|
// Low priority: typing indicators are ephemeral.
|
|
h.broadcastExcludeLow(e.ChannelID(), e.ExcludeUserID(), e.Payload())
|
|
}
|
|
case PresenceSelfEvent:
|
|
// Normal priority, NOT the UserTargetedEvent default below (which
|
|
// PresenceSelfEvent also satisfies — this case must stay ordered
|
|
// before it so the type switch picks this one). Every other
|
|
// source of this same user's own presence — the visible
|
|
// presence_update path (PresenceEvent -> BroadcastToAll) and the
|
|
// connect/disconnect coalescer's private half
|
|
// (BroadcastPresence -> h.SendToUser) — already shares the
|
|
// normal-priority queue. Routing this one through
|
|
// h.SendToUserHigh instead split one user's own presence across
|
|
// two per-client FIFOs with different drain order: writePump
|
|
// always drains high strictly before normal, so a newer
|
|
// invisible self-frame on high could be delivered before an
|
|
// older visible-status frame still sitting on normal, leaving
|
|
// the owner's own client on a stale status — the same hazard
|
|
// OC-0003/OC-0214 fixed for the "others" half of presence.
|
|
h.SendToUser(e.TargetUserID(), e.Payload())
|
|
case UserTargetedEvent:
|
|
// High priority: targeted events (DM opens, mentions).
|
|
// dm_channel_open is unsequenced and targeted, so replay can never
|
|
// deliver it — an addressee mid-reconnect would be left with an
|
|
// unreachable DM channel. Bump the visibility watermark so any
|
|
// client resuming from a seq at or before the open takes the
|
|
// full-ready path (whose payload includes DM channels).
|
|
if _, isOpen := ev.(DMChannelOpenEvent); isOpen {
|
|
// Ratcheted upward only: see bumpVisibilityWatermark on Hub.
|
|
// A plain Store(Load(&h.seq)) here (as on the other two
|
|
// writers) let a writer that read an older h.seq overwrite a
|
|
// concurrently stored higher watermark, silently regressing
|
|
// mustFullResync's boundary.
|
|
h.bumpVisibilityWatermark()
|
|
}
|
|
h.SendToUserHigh(e.TargetUserID(), e.Payload())
|
|
case ChannelEvent:
|
|
h.BroadcastToChannel(e.ChannelID(), e.Payload())
|
|
case VoiceVisibilityEvent:
|
|
// Server-wide, but never to a client that cannot read the channel.
|
|
// ctx is threaded from the dispatching connection so the audience
|
|
// lookup dies with it rather than outliving the request.
|
|
h.broadcastVoiceEvent(ctx, e.VisibleChannelID(), e.Payload())
|
|
case BroadcastAllEvent:
|
|
// Normal priority for everything, including presence: connect and
|
|
// disconnect presence for the same user already go out via
|
|
// hub.BroadcastToAll (serve.go, serve_pumps.go, hub_broadcast.go).
|
|
// Splitting handler-driven presence onto the low-priority queue
|
|
// put it in a different per-client FIFO than those, so writePump
|
|
// (which always drains normal strictly before low) could deliver
|
|
// a newer connect/disconnect frame before an older presence_update
|
|
// still sitting in the low queue — leaving the observer's final
|
|
// view of that user's status stale. Routing everything through
|
|
// BroadcastToAll keeps every source of one user's presence in a
|
|
// single ordered, seq-stamped, replayable stream (OC-0214).
|
|
if pe, isPresence := ev.(PresenceEvent); isPresence {
|
|
// A user-chosen presence also bypasses the connect/disconnect
|
|
// coalescer; drop any entry still queued for this user and
|
|
// broadcast atomically under presenceMu (see
|
|
// dropQueuedPresenceAndBroadcast, OC-0005), or the pending
|
|
// flush (up to 300ms later) could race in between the drop
|
|
// and the broadcast and overwrite this fresher status with
|
|
// the stale connect-time one.
|
|
h.dropQueuedPresenceAndBroadcast(pe.userID, func() {
|
|
h.BroadcastToAll(e.Payload())
|
|
})
|
|
} else {
|
|
h.BroadcastToAll(e.Payload())
|
|
}
|
|
default:
|
|
slog.Warn("EmitEvents: unknown event type", "type", fmt.Sprintf("%T", ev))
|
|
}
|
|
}
|
|
}
|