Files
OwnCord/Server/ws/livekit_webhook.go
J3vbandClaude Opus 5 39551de4a6 refactor(server): work off the complexity backlog — 62 findings to 0 (#1389)
* refactor(ws): split handleVoiceJoin into cohesive join-stage helpers

handleVoiceJoin was 130 statements / cyclomatic 59 / nestif 11, breaking all
three complexity budgets at once. Split along the stage boundaries the doc
comment already described: precheck, leave-current, persist, restore
moderator flags, grant token, complete. The publish-permission derivation
becomes its own helper because it is the one branch-heavy block inside the
token grant.

Pure move: every statement is preserved verbatim. The only edits are bare
`return`s becoming the typed returns of their new helper, `c.userID` becoming
the `userID` parameter inside voiceJoinPublishPerms, and voiceJoinComplete
re-reading `ch.VoiceMaxUsers` instead of receiving it — `ch` is never mutated,
so the value is identical.

Verified by normalising both revisions of the region to sorted, comment- and
whitespace-stripped statements and diffing: the only deltas are the ones
listed above.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor: collapse the three duplicated sibling pairs

dupl flagged three pairs of adjacent near-identical functions. Each pair is
now one parameterised implementation plus two thin, still-greppable wrappers.

- ws/voice_controls.go: handleVoiceMuteV2 / handleVoiceDeafenV2 share
  voiceSelfToggleV2; handleVoiceCameraV2 / handleVoiceScreenshareV2 share
  voiceStreamToggleV2. Camera and screenshare drawing from one
  voice_max_video budget (OC-0023) was a bug caused by exactly this
  duplication drifting, so one body is the point, not a side effect.
- db/mention_queries.go: ListMentionTargetsByRoles / ListMentionTargetsByUserIDs
  share listMentionTargets. The matched column is a closed named type
  (mentionTargetColumn) rather than a bare string, so the value interpolated
  into the SELECT cannot become caller-supplied.

Behaviour is unchanged: every rate-limit key, error code, error string, slog
message and slog key is preserved verbatim, including the two "failed to
update <kind> state" messages, which are now assembled the same way
enableVideoSlot already assembled them.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(api): extract readEmojiUpload from handleCreateEmoji

handleCreateEmoji was 101 lines against a 100-line budget. The upload-bytes
stage — pull the file out of the parsed form, cap its size, sniff its MIME
type and sniff its dimensions — is the one self-contained block in it, and it
already wrote its own refusals, so it moves out whole as readEmojiUpload.

The permission-before-parse ordering the doc comment calls out is unchanged;
so is every error string. file.Close() now runs when the helper returns
rather than when the handler does, which is strictly earlier and unobservable:
the bytes are already copied into raw and nothing else touches the handle.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor: extract one cohesive block from three single-budget offenders

Each of these was over exactly one budget, so each gets exactly one extraction
rather than a restructure:

- api/totp_handler.go handleVerifyTOTP (102 lines / 100): the block that
  resolves the user behind the partial-auth challenge and decrypts their TOTP
  secret becomes totpChallengeSecret. The ban-inside-the-partial-window check
  moves with it.
- service/message_reactions.go handleReaction (cyclop 21 / 20): the whole
  authorisation chain — channel lookup, archived gate, DM participant and
  block checks, non-DM permission check — becomes reactionAudience, which
  also returns the DM fan-out audience it already resolved. Check order is
  unchanged and load-bearing.
- db/admin_queries.go BackupToSafe (cyclop 21 / 20): the character allowlist
  loop and the SQL-comment rejection become validateBackupPathChars. That
  loop alone was most of the branch count.

No error string, no check and no ordering changed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(plugin): split InstallFromZip into staged install helpers

104 statements / cyclomatic 44 / nestif 12. Split along the stages the code
already had: installZipExtract (the per-entry write loop, with
installZipEntryDest holding the mode/symlink/zip-slip guard chain and
installZipWriteEntry the size-capped copy), installZipStagedManifest,
installZipPromote, and installZipReactivate for the :399 nested block.

Every zip-slip, symlink, entry-mode and uncompressed-size check is preserved
in the same order relative to the writes it guards. The 19 inline
`cleanup(); return` sites collapse to 4 in the orchestrator, one per stage,
because each helper now returns an error instead of unwinding itself — the
staging directory is still removed on exactly the same set of failures.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(api): split newWAFMiddleware into engine build and per-phase helpers

184 lines / cyclomatic 38, and the request-body block at :382 was the worst
nested site in the tree at nestif 17.

Engine construction moves out of the closure (wafInlineEngine, wafCRSEngine —
the Coraza directive string is lifted verbatim), and each request phase
becomes its own helper: wafInlineRequestHeaders, wafCRSRequestHeaders
(including the Host/Transfer-Encoding re-add for CRS 920280), wafFeedCRSBody
and wafInspectRequestBody, which is the old :382 block.

The three `handleWAFInterruption(w, it); return` sites inside the body block
become one: the helper now returns the interruption and the orchestrator
handles it. No statement runs between the two points on either side, so the
verdict is honoured identically — in particular a CRS body interruption still
returns without replacing r.Body.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(service): split SendMessage and lift EditMessage's access check

SendMessage was 79 statements / cyclomatic 35 with an 11-deep nested
attachment block at :101; EditMessage was one point over cyclop.

SendMessage becomes sendMessagePrecheck (permission and DM-block gates,
content sanitisation), sendMessageLinkAttachments (the :101 block: attachment
ownership, claim and link) and sendMessageDMSideEffects. EditMessage gets
editMessageCheckAccess and nothing else — one budget over earns one
extraction.

The sanitizeContent fixpoint and the attachment ownership check are unchanged,
as is the order of every gate. The DM side effects run behind
`isDM && !s.sendMessageDMSideEffects(...)`, so a non-DM never enters them;
inside, only the GetDMParticipantIDs failure returns false, matching the one
error the original early-returned on.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(admin): split handlePatchUser into per-field apply helpers

106 lines / cyclomatic 29, with the ban block at :154 nested 9 deep.

Each optional field of the partial edit becomes its own helper —
patchUserPrecheck, patchUserAuthorizeRole, patchUserApplyBan (the :154 block,
including the session disconnect and the broadcast) and patchUserApplyRole.
Each returns a bool meaning "keep going"; none of them writes a success
response, so the single response site in the orchestrator is unchanged.

Field application order, the permission-cache invalidation on a role change
and the disconnect-and-broadcast on a ban are all preserved, as are the three
fail-closed `mod == nil` guards, which now sit at the top of their own helper
and still fire on exactly the same conditions.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(admin): split handleSetup into first-run setup stages

143 lines / cyclomatic 30, with the optional-wizard block at :219 sitting
exactly on the nestif threshold.

Split into the stages the endpoint already had: request gating (rate limit and
origin check, which run before any auth exists on a fresh server), owner
account creation, and the wizard application that was the :219 block.

Every gate in front of the handler is a security control on an unauthenticated
endpoint; none moved relative to the work it protects. setup_wizard.go is
untouched.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor: split run() into named bootstrap and shutdown steps

131 statements / cyclomatic 57, with the executable-path fallback at :126
nested 9 deep.

The five anonymous `defer func(){...}()` blocks become named functions —
telemetryStop, runClosePlugins, runStopEventPersistence, runStopAuditWriter,
maintenanceStop — and the bootstrap stages move out likewise.

Every defer is still registered in run() itself, at the same point in the
sequence, so the LIFO teardown order is unchanged; that order is documented
in the surrounding comments and is load-bearing (the audit-writer stop must
follow database.Close's registration, the event-persistence stop must precede
it). runStopEventPersistence is now registered unconditionally with a nil
persister meaning "disabled", where the old code registered its defer inside
the enabled branch — a no-op occupying that slot cannot change the relative
order of the others.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(ws): split handleReconnect into resume stages

77 statements / cyclomatic 41, plus the replay block at :199 and, in
handleFreshConnect, the voice-state restore at :622.

handleReconnect becomes reconnectPrecheck, reconnectSelectReplay (with
reconnectVetColdTail for the cold-tier gap check), reconnectRegister and
reconnectWriteReplay. handleFreshConnect's stale-voice cleanup moves to its
own helper, where the `if h.livekit != nil` wrapper becomes a guard clause —
that block was the tail of its scope, so returning early and falling off the
end are the same.

The parts that carry the invariants are moved verbatim: reconnectRegister
still takes h.seqMu, still calls registerNow inside that same critical
section (BUG-123 / OC-0206), still unlocks on every exit, and still emits the
"full" tier counter and telemetry on each of its three re-check failures.
handleReconnect's two-boolean contract is unchanged — the collapsed
`return false, false` sites are all fall-through-to-full-ready, and the
single `return true, false` is still the handshake-write-failure path whose
teardown already ran (OC-0051).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(server): fold in the adversarial review of the complexity refactors

Eleven skeptic passes over the refactor commits on this branch found no
blocker and no major — behaviour is preserved throughout. They did find
comment and accuracy defects worth correcting:

- db/mention_queries.go: the mentionTargetColumn rationale claimed the named
  type made the interpolated column "only ever one of the two constants". A
  Go named type is not closed, so that is a convention the type makes visible,
  not one it enforces. Reworded, gosec justification included.
- ws/voice_controls.go: the dupl collapse generalised away three specifics —
  that a server deafen is the moderator's to lift (now on the serverDeafen
  field), the concrete voice_states.camera / voice_states.screenshare column
  names, and the half of the OC-0023 rationale about neither stream kind
  hiding from the other's count. All three restored.
- ws/voice_join.go: `maxUsers := ch.VoiceMaxUsers` had been hoisted to the top
  of voiceJoinComplete, moving a read across the tail supersession guard. The
  read is inert, but it was the one statement in that commit whose position
  relative to a security guard changed; it now sits at its use, as before.
- ws/*_test.go: three test comments cited voice_join.go line numbers that the
  split invalidated. They now cite the helper by name instead.
- service/message_reactions.go: reactionAudience's doc claimed to enforce
  "every gate on reacting"; it enforces the channel-scoped ones, and the doc
  now says which gates stay with the caller.
- api/emoji_handler.go: the readEmojiUpload call reused the outer `ok` from
  the auth check by assignment; it gets its own readOK.
- admin/setup_handler.go: a moved comment kept a "the response above" deictic
  that no longer had a response above it.

No behaviour change. Build, vet, full tests and -race on five packages green.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(ws): clear the remaining complexity budgets across the hub

Eight files, thirteen findings. Each function is split at the stages it
already had; no branch is reordered, merged or inverted.

- handlers.go handleMessage (cyclop 28, 88 stmts): session re-check, frame
  decode and result application become handleMessageSessionRecheck,
  handleMessageDecode and handleMessageApply. The V2 constructor lookup ->
  DispatchV2 -> Result resolution order is untouched.
- serve_ready.go buildReady (cyclop 26, 61 stmts): the per-section fetches
  split out, readyChannelPayloads among them. Every visibility predicate is
  preserved verbatim — this is the payload that decides what a client may see.
- serve_pumps.go writePump (cyclop 31): writePumpWrite, writePumpDeliver,
  writePumpDrainChannel and writePumpDrainAndClose. Every channel receive
  stays in the same select statement, so scheduling is unchanged.
- hub_sweep.go sweepStaleVoiceStates (cyclop 22, 56 stmts): the staleness
  predicate, the hub-lock ordering and the position of the race hook are all
  as they were — handleVoiceJoin's BUG-088 ordering depends on them.
- hub_broadcast.go channelReadAudienceImpl and RefreshChannelVisibility
  (cyclop 22 each, 57 stmts): channelReadAudienceDM and
  refreshChannelVisibilityCanSend. The audience predicate is the OC-0090
  group-DM leak surface, so it is extracted, never simplified.
- livekit_webhook.go (nestif 13 and 14): webhookJoinedEnforceVoiceState,
  webhookLeftCleanupClient and webhookLeftFinishLeave. DB delete still
  precedes broadcast on every path.
- livekit_download.go EnsureLiveKitBinary (52 stmts): one extraction,
  ensureLiveKitStageBinary, keeping every archive path check intact.
- voice_moderation.go (nestif 8): voiceModDeafenRollback. The persisted
  server_muted flag remains the authority.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor(api): clear the remaining complexity budgets across the HTTP layer

- router.go NewRouter (cyclop 28, 84 stmts): split by wiring concern into
  routerTOTPKey, routerHealthDeps, routerMiddleware, routerUploadRoutes,
  routerPluginWiring, routerVoiceRoutes and routerMetricsRoutes. Middleware
  ORDER is a security property (auth before handler, WAF before body parse,
  rate limit before work) and is unchanged; the returned cleanup func still
  closes over and releases everything it did before.
- auth_handler.go handleRegister (133 lines) and handleLogin (cyclop 21,
  152 lines): registerPolicyGate, registerReadRequest, loginReadRequest and
  loginAuthenticate. The always-compare posture, every rate-limit key, every
  counter reset and the ban-check-versus-password-compare order are all
  preserved — including loginUserFailureThreshold staying unscaled by
  scaledAuthLimit, which is deliberate and commented.
- upload_handler.go handleServeFile (cyclop 31, 128 lines): serveFileResolve
  and serveFileAuthorize. Every header this sets — Content-Disposition
  included, which is what stops a stored file being served as active content —
  is still set with the same value in the same circumstances.
- profile_handler.go handleUploadAvatar (120 lines): avatarUploadReadImage,
  mirroring readEmojiUpload in shape but with the avatar caps and MIME set.
  The two deliberately do not share a helper.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* refactor: clear the last complexity budgets in db and admin

- db/account.go DeleteAccount (cyclop 28, 55 stmts): grouped by subsystem into
  deleteAccountAdminGuard, deleteAccountDMChannels and
  deleteAccountCloseDMChannels, each taking the same transaction. The
  transaction boundary, the delete ORDER (which foreign keys depend on) and
  the rollback path are unchanged.
- admin/logstream.go handleLogStream (cyclop 24): logStreamAuthorize. Flush
  cadence, heartbeat and disconnect detection untouched.
- admin/setup_wizard.go validateWizard (cyclop 23): grouped by section into
  wizardValidateIdentity, wizardValidateNetwork and wizardValidateMedia. Every
  message and bound is unchanged — this is the first input-validation boundary
  on a fresh server, before any auth exists.

With this the tree is at zero: golangci-lint run reports 0 issues against the
budgets set in #1384 (funlen 100/50, cyclop 20, nestif 8, dupl 150), with no
//nolint and no exclusion added anywhere.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 20:39:45 +02:00

332 lines
13 KiB
Go

package ws
import (
"context"
"fmt"
"log/slog"
"net/http"
"strconv"
"strings"
"github.com/livekit/protocol/auth"
"github.com/livekit/protocol/livekit"
"github.com/livekit/protocol/webhook"
)
// webhookMaxBodyBytes bounds the webhook request body to prevent unbounded
// reads from an unauthenticated caller.
const webhookMaxBodyBytes = 64 * 1024
// NewLiveKitWebhookHandler returns an HTTP handler that processes LiveKit
// webhook events. It synchronises LiveKit room state back into OwnCord's
// voice_states DB — primarily for crash recovery when a participant
// disconnects from LiveKit without sending a WS voice_leave.
//
// Speaker detection is handled client-side via LiveKit's
// RoomEvent.ActiveSpeakersChanged (lower latency than webhooks).
func (h *Hub) NewLiveKitWebhookHandler(apiKey, apiSecret string) http.HandlerFunc {
// The SDK receiver verifies the token signature AND that the token's sha256
// claim matches the request body hash, binding verification to the body so
// a captured token cannot be replayed against a forged payload.
provider := auth.NewSimpleKeyProvider(apiKey, apiSecret)
return func(w http.ResponseWriter, r *http.Request) {
// Check Authorization header BEFORE reading the body to avoid
// allocating memory for unauthenticated requests.
if r.Header.Get("Authorization") == "" {
slog.Warn("livekit webhook: missing Authorization header")
http.Error(w, "unauthorized", http.StatusUnauthorized)
return
}
// Bound the body before the SDK reads it (ReceiveWebhookEvent uses an
// unbounded io.ReadAll internally).
r.Body = http.MaxBytesReader(w, r.Body, webhookMaxBodyBytes)
// ReceiveWebhookEvent verifies the JWT signature, the token's body-hash
// claim, and the exp/nbf claims, then parses the payload. This replaces
// the previous manual ParseAPIToken/Verify sequence, which was not bound
// to the request body (forgery/replay).
event, err := webhook.ReceiveWebhookEvent(r, provider)
if err != nil {
slog.Warn("livekit webhook: verification failed", "error", err)
http.Error(w, "unauthorized", http.StatusUnauthorized)
return
}
slog.Info("livekit webhook received",
"event", event.Event,
"room", event.GetRoom().GetName(),
"participant", event.GetParticipant().GetIdentity(),
)
switch event.Event {
case "participant_joined":
h.handleWebhookParticipantJoined(r.Context(), event)
case "participant_left":
h.handleWebhookParticipantLeft(r.Context(), event)
default:
slog.Debug("livekit webhook: unhandled event", "event", event.Event)
}
w.WriteHeader(http.StatusOK)
}
}
// parseParticipantIdentity extracts a user ID and optional join token from a
// LiveKit participant identity formatted as "user-{id}" or
// "user-{id}:{joinToken}".
func parseParticipantIdentity(identity string) (int64, string, error) {
if !strings.HasPrefix(identity, "user-") {
return 0, "", fmt.Errorf("invalid identity format: %s", identity)
}
body := identity[5:]
idPart, joinToken, _ := strings.Cut(body, ":")
userID, err := strconv.ParseInt(idPart, 10, 64)
if err != nil {
return 0, "", err
}
return userID, joinToken, nil
}
// parseRoomChannelID extracts a channel ID from a LiveKit room name
// formatted as "channel-{id}".
func parseRoomChannelID(roomName string) (int64, error) {
if !strings.HasPrefix(roomName, "channel-") {
return 0, fmt.Errorf("invalid room name format: %s", roomName)
}
return strconv.ParseInt(roomName[8:], 10, 64)
}
func (h *Hub) handleWebhookParticipantJoined(ctx context.Context, event *livekit.WebhookEvent) {
// Detach from the triggering HTTP request before doing any cleanup work,
// mirroring every sibling teardown path (readPump's defer and
// unregisterFailedHandshake use context.WithoutCancel in serve.go /
// serve_pumps.go, rollbackVoiceJoin uses it in voice_join.go, the hub
// sweeps use context.Background in hub_sweep.go). Without this, a webhook
// sender (LiveKit) that hangs up mid-request cancels r.Context(), and the
// rogue-participant GetVoiceState/RemoveParticipant calls below would
// either wrongly skip (treating a cancelled read as a transient error, ok)
// or fail outright instead of completing the eviction.
ctx = context.WithoutCancel(ctx)
p := event.GetParticipant()
room := event.GetRoom()
if p == nil || room == nil {
return
}
userID, joinToken, err := parseParticipantIdentity(p.Identity)
if err != nil {
slog.Warn("livekit webhook: participant_joined bad identity",
"identity", p.Identity, "error", err)
return
}
channelID, err := parseRoomChannelID(room.Name)
if err != nil {
slog.Warn("livekit webhook: participant_joined bad room",
"room", room.Name, "error", err)
return
}
slog.Info("livekit webhook: participant joined",
"user_id", userID,
"channel_id", channelID,
"room", room.Name)
// Validate that the participant has a matching voice_states row (BUG-127).
// A replayed token from a previous session will not have a matching row,
// so we remove the rogue participant from LiveKit.
if h.db != nil {
h.webhookJoinedEnforceVoiceState(ctx, userID, channelID, joinToken)
}
}
// webhookJoinedEnforceVoiceState is the voice_states reconciliation stage of
// handleWebhookParticipantJoined: it matches the joining participant against
// their DB row and removes them from the SFU when the row is missing, points
// at another channel, or carries a different join token. Callers guarantee
// h.db != nil.
func (h *Hub) webhookJoinedEnforceVoiceState(ctx context.Context, userID, channelID int64, joinToken string) {
state, stateErr := h.db.GetVoiceState(ctx, userID)
if stateErr != nil {
// A transient read failure (I/O error, lock contention, a
// maintenance window) is not proof of a rogue participant —
// treating it as one would eject a legitimate participant from
// the SFU on a single bad read. Mirrors sweepStaleVoiceStates'
// hasChannelPermChecked guard: skip and let the participant be;
// a later webhook retry or sweep tick resolves it.
slog.Error("livekit webhook: GetVoiceState failed, skipping rogue-participant check",
"error", stateErr, "user_id", userID, "channel_id", channelID)
return
}
if state == nil || state.ChannelID != channelID {
slog.Warn("livekit webhook: rogue participant_joined — no matching voice state, removing",
"user_id", userID, "channel_id", channelID)
if h.livekit != nil {
if rmErr := h.livekit.RemoveParticipant(ctx, channelID, userID, joinToken); rmErr != nil {
slog.Error("livekit webhook: failed to remove rogue participant",
"error", rmErr, "user_id", userID, "channel_id", channelID)
}
}
return
}
// Verify join token matches to prevent token replay from old sessions.
if joinToken != "" && state.JoinedAt != joinToken {
slog.Warn("livekit webhook: stale join token on participant_joined, removing",
"user_id", userID, "channel_id", channelID,
"expected_token", state.JoinedAt, "got_token", joinToken)
if h.livekit != nil {
if rmErr := h.livekit.RemoveParticipant(ctx, channelID, userID, joinToken); rmErr != nil {
slog.Error("livekit webhook: failed to remove stale participant",
"error", rmErr, "user_id", userID, "channel_id", channelID)
}
}
return
}
}
func (h *Hub) handleWebhookParticipantLeft(ctx context.Context, event *livekit.WebhookEvent) {
// Detach from the triggering HTTP request before doing any cleanup work
// (OC-0018), mirroring every sibling teardown path (readPump's defer and
// unregisterFailedHandshake use context.WithoutCancel in serve.go /
// serve_pumps.go, rollbackVoiceJoin uses it in voice_join.go, the hub
// sweeps use context.Background in hub_sweep.go). Without this, a webhook
// sender (LiveKit) that hangs up mid-request cancels r.Context(), which
// makes channelReadAudience's GetChannel call fail and fail closed to an
// empty audience (hub_broadcast.go) — silently dropping the voice_leave
// for anyone who has READ_MESSAGES on the channel but is not currently in
// the room. Unlike the DB row, no sweep ever re-emits that missed
// broadcast. The same cancellation would also make both
// LeaveVoiceChannelIfMatch branches below fail on their synchronous first
// attempt.
ctx = context.WithoutCancel(ctx)
p := event.GetParticipant()
room := event.GetRoom()
if p == nil || room == nil {
return
}
userID, joinToken, err := parseParticipantIdentity(p.Identity)
if err != nil {
slog.Warn("livekit webhook: participant_left bad identity",
"identity", p.Identity, "error", err)
return
}
channelID, err := parseRoomChannelID(room.Name)
if err != nil {
slog.Warn("livekit webhook: participant_left bad room",
"room", room.Name, "error", err)
return
}
slog.Info("livekit webhook: participant left",
"user_id", userID,
"channel_id", channelID)
// Clean up voice state if the user disconnected from LiveKit
// without sending a WS voice_leave (e.g. crash, network loss, F5 reload).
h.mu.RLock()
c, exists := h.clients[userID]
h.mu.RUnlock()
if exists {
h.webhookLeftCleanupClient(ctx, c, userID, channelID, joinToken)
} else if h.db != nil {
// Client already disconnected from WS — use channel-conditional delete
// to avoid wiping a newer row if the user reconnected and rejoined.
deleted, dbErr := h.db.LeaveVoiceChannelIfMatch(ctx, userID, channelID, joinToken)
if dbErr != nil {
slog.Error("livekit webhook: LeaveVoiceChannelIfMatch failed (client gone)",
"error", dbErr, "user_id", userID, "channel_id", channelID)
} else if deleted {
h.broadcastVoiceEvent(ctx, channelID, buildVoiceLeave(channelID, userID))
}
}
}
// webhookLeftCleanupClient is the still-connected-client stage of
// handleWebhookParticipantLeft: it compare-and-clears the client's voice
// fields for this exact join instance, then either finishes the leave or, when
// the client has already moved on, clears the stale DB row.
func (h *Hub) webhookLeftCleanupClient(ctx context.Context, c *Client, userID, channelID int64, joinToken string) {
// Atomic compare-and-clear under c.voiceMu, replacing the previous
// read-then-read-then-clear: two independent unlocked getVoiceState
// snapshots followed by an unconditional clearVoiceState is not a
// guard at all — no lock spans the second read and the clear, so a
// voice_join committed on the readPump goroutine in between (a
// channel switch, or a same-channel rejoin with a fresh token) is
// wiped out from under the new session, dropping its VoiceTopic
// subscription along with it. client.go's clearVoiceStateIfMatch
// only compares the channel, not the token, so it would still be
// fooled by a same-channel rejoin — this compares both, inlined here
// via direct field access (same package as client.go) under the
// client's own voiceMu.
c.voiceMu.Lock()
matched := c.voiceChID == channelID && c.voiceJoinToken != "" && c.voiceJoinToken == joinToken
if matched {
c.voiceChID = 0
c.voiceJoinToken = ""
c.e2eePubKey = ""
c.e2eeSignature = ""
}
c.voiceMu.Unlock()
if matched {
h.webhookLeftFinishLeave(ctx, c, userID, channelID, joinToken)
} else if h.db != nil {
// Client has voiceChID=0 or moved to a different channel (e.g.
// after F5 reload), or this webhook is for an older join instance.
deleted, dbErr := h.db.LeaveVoiceChannelIfMatch(ctx, userID, channelID, joinToken)
if dbErr != nil {
slog.Error("livekit webhook: LeaveVoiceChannelIfMatch failed (stale DB row)",
"error", dbErr, "user_id", userID, "channel_id", channelID)
} else if deleted {
h.broadcastVoiceEvent(ctx, channelID, buildVoiceLeave(channelID, userID))
slog.Info("livekit webhook: cleaned stale DB voice row after reconnect",
"user_id", userID, "channel_id", channelID)
}
}
}
// webhookLeftFinishLeave is the tear-down stage of webhookLeftCleanupClient,
// reached once the client's voice fields matched this join instance and were
// cleared: drop the voice subscription, clear the DB row, move the E2EE key
// holder on, and broadcast the leave.
func (h *Hub) webhookLeftFinishLeave(ctx context.Context, c *Client, userID, channelID int64, joinToken string) {
h.pubsub.Unsubscribe(c, VoiceTopic(channelID))
if h.db != nil {
if err := leaveVoiceChannelWithRetry(ctx, h, userID, channelID, joinToken); err != nil {
slog.Error("livekit webhook: LeaveVoiceChannel exhausted retries",
"error", err, "user_id", userID, "channel_id", channelID)
}
}
// This participant is out of voice, so the E2EE key holder may
// need to move. Without this the map keeps naming the departed
// user and the real lowest-uid participant's rekey offers are
// rejected with NOT_KEY_HOLDER. Safe here: no locks are held.
h.updateKeyHolder(channelID)
// The leaver's own client state was just cleared above, so
// broadcastVoiceEvent's still-in-the-room union can no longer see
// them — without broadcastVoiceEventWithLeaver's extra term, a
// participant without READ_MESSAGES on this channel (voice
// membership needs only CONNECT_VOICE) never learns the server
// already tore down their call. Mirrors finishVoiceLeave and
// CleanupVoiceForChannel, which add the leaver for the same reason.
h.broadcastVoiceEventWithLeaver(ctx, channelID, buildVoiceLeave(channelID, userID), userID)
slog.Info("livekit webhook: cleaned up stale voice state",
"user_id", userID,
"channel_id", channelID)
}
// MountWebhookRoute is a helper for the router to mount the webhook endpoint.
func MountWebhookRoute(h *Hub, apiKey, apiSecret string) http.HandlerFunc {
return h.NewLiveKitWebhookHandler(apiKey, apiSecret)
}