Review: integration of the hunt bundle (collection branch, tip 3e337bcc8)

From mnema-talk v3.3.13, rendered on this site.

Reviewer: independent, read-only. Scope: what the coordinator did while merging about 40 reviewed fix and harness branches onto origin/main f58bc3fc6 (release 3.0.9): the coordinator's own commits, every merge whose result differs from the automatic merge, and interactions between fixes that no single fix review could see. The branch for H-L5-01 and H-L5-04 is not in this tip and was not reviewed. Own worktree at 3e337bcc8, removed at the end; nothing pushed, no branch created.

Verdict: OK to open the pull request

No blocker. One SHOULD-FIX (a documentation sentence, can go into the same pull request or later), and NOTEs. One defect from before this branch is written up below as a finding candidate. It is on 3.0.9 as well, and this branch neither causes nor worsens it.

Method

  • git log --first-parent origin/main..3e337bcc8: 52 first-parent commits. 10 are not merges: the eight the task names, the three answer-vetting commits (which are on the second parent of 90e3c71f6), and 82e51418f, the lane 5 store trace harness at the bottom of the branch (only tests and one workflow).
  • For every merge, the result was compared with git merge-tree --write-tree ^1 ^2. That redoes the automatic merge, so this finds every hand-resolved file, including a resolution that silently took one side. git show --cc alone misses that case. 12 merges differ from the automatic merge. In each of them, every line that either parent added since the merge base was checked against the result, and every line missing from the result was read and explained (list below).
  • Production files changed by more than one merge, and functions changed by more than one merge, were listed from the hunk headers and read.
  • Every lock statement the branch adds (blame against origin/main) was traced against the erasure's lock order. One cycle was tried on a real server.

A. The coordinator's own commits

a4258ed83, internal/automod/rule.go:268 (cleanDomain). A link-rule domain must now read back through hostOf as itself, or it is refused as "not a domain". Probes, before and after the fix: 0.00.. was accepted as 0.00 before (hostOf reads 0.00 as 0.0.0.0, so the entry could never match a link) and is refused now. 1.00, 0x7f.00, 192.168.000, https://0.00./x, 1.2.3 and 10.0.0.1 are refused before and after. example.com., EXAMPLE.com.., the Unicode form of xn--bcher-kva.example and a.1e are accepted before and after, with the same stored form. Reverting rule.go makes seed 0.00.. of FuzzCleanDomain fail. The fix is correct and refuses only names no link can match. Callers: the only caller is RuleInput.Normalize, used only by parseRule (internal/moderation/automod_rules.go:91) for create and update. Stored rules are loaded by automod.ListRules and compiled (internal/automod/compile.go:37) without being normalized again. So a rule stored earlier with such an entry still loads, and it matches nothing, as before. NOTE: updating such a rule with its full body (the admin dialog sends the whole config) now answers 400 "0.00" is not a domain until the admin removes that entry. This is the right outcome. No upgrade note is needed for an entry that never worked. NOTE: FuzzCleanDomain's property (hostOf(d) == d) is now exactly the check cleanDomain performs, so the fuzz target can no longer fail. It still guards a future removal of the check. The real oracle of this fix is FuzzLinkRules and the two seeds.

5d88a24bd (webhook settings test). It now counts the id in mentions instead of in the whole answer, because the mention-kinds fix also lists the member under direct_mentions. DirectMentions is documented, and built, as a subset of Mentions (internal/chat/core/message_types.go:58, internal/chat/messages/messages.go:151). So every "0" assertion is exactly as strict as before, and the "1" assertion measures what it always meant. It is not weaker. The test passes.

b61a553a7 (export fault cycle). buildExportInBackground now returns an error (!= nil replaces !). Drain now refuses every job after it starts (the drain fix), so the per-cut wait is DrainMail, which waits on the same WaitGroup (h.rec.wg, internal/auth/authn/account_mail.go:137) without closing job acceptance. Each build joins that group through TrackJob. The final step still uses Drain(30s). The test is equivalent: it waits for the same builds. Its only bound is now the test timeout instead of 30 s. It passes.

871dc3ab9 (presence burst). Stricter: it adds an assertion that the audit_presence drop counter equals the missing rows. It passes. The documentation sentence it changed is SHOULD-FIX 1 below.

14ebe8be1 (ci). It adds MNEMA_DRAIN_HAMMER=1 go test -race -run '^TestDrainHammer$' ./internal/auth/authn/. go test -list finds the test, and it ran green under -race (16 s).

4ae4fbe03 (docs + test helper). The orphan scan covers uploads/, avatars/, banners/, covers/ and exports/, and checks exports against data_exports.s3_key (internal/media/orphans.go:26-32, :74). The new OpenAPI and annotation sentences are true. make openapi-check is 0. The removed knownGap helper has no user left.

3903bc122 (benchmark). 120 allocation requests per minute per member (newMemberLimiter, internal/turnrelay/server.go:58, options.go:49) and 256 in total (options.go:46). True.

82b4ba901 (architecture). passPacer keeps one writer per SSRC (plus an RTX alias), and UnbindLocalStream removes both (internal/sfu/bwe.go:101, :163-170). True.

Answer vetting (e01ced711, 065d9b451, 0e5b74b47). internal/sfu/answer_candidates.go, answer_ssrcs_test.go and attach_viewer_test.go are byte-identical between 0e5b74b47, the tip that fixreview-H-L8-08 round 2 approved, and 3e337bcc8, and no commit touches them afterwards. ingress.go and ssrc_flood_test.go were changed later by other reviewed fixes (a15ea9e18, d77aa2893, 714225dd4). Every line the three commits added to ingress.go is still in the tree.

82e51418f (store trace harness, not in the task's list). Tests, fixtures and .github/workflows/store-trace.yml only. The workflow is pinned to full SHAs, has permissions: contents: read and a 60-minute timeout, and puts no untrusted input in run steps. NOTE: store-trace.yml is not in AGENTS.md's workflow tree. Neither are ci-macos-calls, codeql, load-test, onboarding and turn-probe on main, so this is not a regression.

B. Merges with a hand resolution

Hand-resolved files (result differs from the automatic merge): 3e337bcc8 (sim-sweep.yml, AGENTS.md), 3fcdb7097 (sfu.go, room_tracks.go, subscription_cycle_test.go), 50b495ffa (events.md, talks.md, upgrade.md, handler_validate.go), fa544c89a (SECURITY.md, room_attach.go, room_tracks.go, sfu.go, clock_step_test.go, source_replacement_test.go, subscription_cycle_test.go), 02395c7a8 (SECURITY.md, upgrade.md, erasure_in_flight_integration_test.go), 23dc6e381 (upgrade.md, webhooks.md), f305da991 (communityevents/handler.go), a5d17e226 (upgrade.md), 46f97679a (storeTrace.known.ts), 8cbc568c3 (two canary tests), 2f4342866 (moderation.md), 0c47488b8 (sim-sweep.yml, AGENTS.md). The combined diffs of 6577da7fb (message_mutations.go) and 17abe73dd are clean automatic merges (read below).

fa544c89a and 3fcdb7097 (resume budget on the steady clock). In the final tree the only writer of TrackInfo.resumeAt is stopSourceTracksLocked (internal/sfu/sources.go:104-117). It stores stopUntilLocked(..., layerNow()), which returns steadyNanos(now.Add(wait)) (sources.go:166). The only reader is resumeStoppedTrack (internal/sfu/room_attach.go:388-396), which compares with steadyNanos(layerNow()) and clears the value with a CAS. The budget (resumeBudget.at, a time.Time from layerNow()) is only used through now.Sub(b.at) (sources.go:135-137), which uses the monotonic reading. So neither the deadline nor the refill can be moved by a wall clock step. steadyNanos itself is steadyStartUnix + t.Sub(steadyStart) (layers.go:67). Nothing in the package still reads UnixNano for these values. The fa544c89a resolution (resumeAt in steadyNanos, stopUntil on layerNow) was then fully replaced by the per-source budget of bbfec1723 in 3fcdb7097, which took the fix side. That side already returns steadyNanos, because the fix branch was built on the steady-clock fix. The dropped TrackInfo.stopUntil, resumes and resumesAt have no users left (the build and vet are 0). Tests: TestClockStepBackResumesAStoppedCamera and TestClockStepForwardKeepsTheResumeHoldoff step the wall clock while keeping the monotonic reading. TestStopResumeBudget and TestResumeBudgetOutlivesCapturesAndRejoins use the room's budget. The SFU suite is green. SECURITY.md: both sentences are kept, in both merges. The budget text says "belongs to the publisher and source, so a new capture or a rejoin does not refill it", which is true: the entry of a publisher who left is deleted only once the budget is full again (sources.go:146-150).

f305da991 (createEvent). The resolution keeps commitAfter(r, pending, err) from the automod side and drops that side's post-commit h.auditRecord(r, "event.create", ev), because the first parent moved the entry into createOrStartSeries. Paths in the final tree:

  • single event: Store.create(..., RecordTx(tx, eventEntry("event.create", ev))) (handler_series.go:36-39) writes exactly one entry, inside the transaction. A failed entry rolls the event back, so err != nil and the automod hit is not committed.
  • series: CreateSeries, then one auditRecord after it commits (handler_series.go:52). Exactly one entry. It is not inside a transaction, and doc/moderation.md:520-530 says so ("a new event (not a series)" is in-transaction, "series" is after the commit). The hit is committed when CreateSeries succeeded. A later GenerateSeries error is only logged, and the event exists then, so committing is right.
  • an error before creation (validation, frequency, until) returns before commitAfter commits anything. The in-transaction path does not call auditable. It needs no call: creation refuses a conversation's call Talk as its location (validateEventTalk, handler_validate.go:62-70), and RecordTx is nil-safe (audit.go:79). Correct.

Documents (02395c7a8, 50b495ffa, 23dc6e381, a5d17e226, 2f4342866). Every paragraph was compared with both parents:

  • SECURITY.md (02395c7a8): the erasure sentence of one side and the export orphan wording of the other ("the orphan cleanup removes only an archive no export row names") are both kept. This matches 4ae4fbe03 and orphans.go.
  • doc/upgrade.md: the sections "Nothing new may name an erased account", "Removals from a Talk in the audit log", "Calls in conversations no longer need View channels", "Conversations' calls leave the community's objects" and "A role mention next to @all breaks a mute" are each present once. No sentence of either parent is missing. The only lines dropped are conflict markers and blank-line placement.
  • doc/events.md and doc/talks.md (50b495ffa): each resolved paragraph is the union of both parents' sentences (automod screening plus "not at all for an event of a conversation's call Talk"; "was removed from it" plus "no community channel ... PATCH /api/talks/{id}" plus the admin-wide kick sentence). Nothing contradicts.
  • doc/webhooks.md (23dc6e381): "a web app from before 3.0.10" became "a web app that does not know direct_mentions yet". The version is not fixed yet, so this is the better wording.
  • doc/moderation.md (2f4342866): the voice.join row comes from the first parent (with 871dc3ab9's wording), and the new voice.disconnect row from the second. Both are kept.

8cbc568c3 and 46f97679a (emptied known-failure lists).

  • canaryKnownLeaks and canaryKnownDeepOnly (all H-L6-01; its fix d1e297554 is merged earlier in the branch) are empty. Ran VIS_SWEEP=1 go test -tags=integration -run TestNoCanaryReachesAnObserverOutsideItsClass ./internal/server: PASS in 1171 s, the deep mode the nightly job runs. Smoke is green in the full suite. So no removed row still leaks, and a leak that came back would fail.
  • storeTrace.known.ts drops the H-L5-02 and H-L5-03 rows (the fix's side) and clean reply-follow H-L5-05 (merged in 8b853104c before). The remaining rows are H-L5-01 and H-L5-04, both still open here. I recorded 20 new traces (STORE_TRACE_SEEDS=20 STORE_TRACE_FIRST=5000 go test -tags=integration -run 'TestStoreTrace$' ./internal/server, exit 0) and replayed them unsharded (STORE_TRACE_DIR=... vitest run src/stores/storeTrace.test.ts -t deep): passed. With 20 or more unsharded traces the test requires every nightly key to still break and no other key to break. So the list is right at this tip. (The lane 5 branch still to come will shrink it again.) The smoke replay is green in the web run below. NOTE: the replay prints happy-dom AbortError stacks at teardown, but the run is green.

0c47488b8 and 3e337bcc8 (sim-sweep.yml). The final jobs are sweep, availability, links, powers, clocks, cycle, explore, crash, sfu, trust and races. Each is present once. The text of each job block was compared with the parent it comes from: identical, except that sweep is the newer side's (with the pooled-availability upload, which availability needs), and that races has 14ebe8be1's step. actionlint .github/workflows/*.yml: exit 0. Commands: go test -list finds TestUnhandledSRTCPStreamsAreCapped, TestSubscriptionSwitchesRetainNothing, TestTrustChainSequences and TestDrainHammer. The races pattern matches 5, 1, 1, 3 and 21 tests in its five packages. For the web jobs, vitest list printed nothing for most env-gated files even with their variables set (a listing quirk; it did list groupExplore and groupCrash). So I checked instead that each file exists and that its describe.skipIf(!process.env.X) gate is the variable the job sets (SIM_SWEEP, SIM_POOL_DIR, LINK_SWEEP, SIM_POWERS, SIM_CLOCKS, SIM_CYCLE, SIM_EXPLORE, SIM_CRASH): all match. AGENTS.md: the sim-sweep line is the union of both descriptions, and visibility-sweep.yml is added. Its wording is a little long ("...; and ... ; and the SFU's resource bounds"), but nothing is lost.

50b495ffa code files. handler_validate.go imports auth, automod and talks, and all three are used. In channel_permissions.go, checkChannelMove combines H-L6-02's removal of targetCategoryVisible with the call-Talk fix's snap.IsDMTalk(c.ID) refusal. Both are kept, and the order (hidden check first) is right.

Other combined diffs. In 6577da7fb (message_mutations.go changeReaction), core.ReactionForm normalization comes before the applyReaction refactor of the other fix, so every path of the toggle sees the normalized emoji. In 02395c7a8 and 8cbc568c3 (test worlds), the renames canaryWorld and erasureWorld are consistent, and the tests compile and pass.

C. Interactions between fixes

Functions changed by more than one fix: audit.filter.restrictTo (H-L6-01, then H-L6-03 built on it), channels.selectExpired (auto-delete pin fix, then the call Talk fix), checkChannelMove, changeReaction, createEvent and auditRecord, SetAnswer and removeStaleSenders (H-L8-07 and H-L8-08), resumeStoppedTrack and stopSourceTracksLocked. Each was read in its final form. No function lost one fix's behaviour to the other.

Locks. The new locks are: the reporter row FOR NO KEY UPDATE (reports.go:198), the message FOR UPDATE on delete (message_mutations.go:173), FOR KEY SHARE on the pin's root (pins.go:184), the group conversation FOR UPDATE (dm_members.go:19), the webhook channel FOR NO KEY UPDATE (webhook/store.go:233), the advisory locks (account tokens, owner seed, role templates), and the 0103 triggers (FOR KEY SHARE on users). Only the erasure takes users rows in a conflicting mode (FOR UPDATE, account_erase.go:85). FOR KEY SHARE and FOR NO KEY UPDATE do not conflict, so the report path cannot cycle with the 0103 triggers. The advisory locks are taken first in their transactions and are each held by one code path only. The pin path takes a key-share lock on the root, and the sweep skips locked rows (SKIP LOCKED), so no wait arises there. No cycle is created by combining two of this branch's fixes.

One cycle that is already on 3.0.9 (finding candidate, not caused by this branch): the erasure takes the member's users row first and then empties their messages (internal/chat/account/account_erase.go:42). The member's own edit from a second device locks the message first (FOR NO KEY UPDATE, message_mutations.go:65-70) and then inserts into message_edits (edit_history.go:34), which needs the users row (the foreign key check on 3.0.9, the 0103 guard on this branch). PostgreSQL resolves the cycle by aborting the transaction whose deadlock check runs while the cycle exists. When the edit's wait starts more than deadlock_timeout (1 s) before the erasure reaches its messages step, that transaction is the erasure. Reproduced on a real server, through the real routes, with the erasure held by a pgx tracer right before its messages step:

  • held for 300 ms: the edit answers 500 (deadlock detected), and the erasure answers 204.
  • held for 2.5 s: the erasure answers 500 (erase "UPDATE messages SET content = '' ...": deadlock detected (SQLSTATE 40P01)), and the edit goes through. The same probe on f58bc3fc6 (3.0.9) gives the same two outcomes. Nothing leaks: the erasure rolls back, and a retry succeeds. But the H-L10-05 review's statement "the erasure always wins" (fixreview-H-L10-05.md:232-234) holds only when the writer starts waiting less than 1 s before the erasure closes the cycle. A large account, whose earlier erase steps take longer than that, can see its deletion fail once. The probe is at reviews/review-bundle-integration-deadlock-probe.go.txt (put it in internal/server and run it with -tags=integration). (The probe is not in the repository; the finding it led to is H-L10-06, doc/security/findings/2026-11-hunt/h-l10-06.md.) Suggested fix, in a later hunt branch: retry the erasure transaction on 40P01, or lock the member's messages (SELECT ... FROM messages WHERE user_id = $1 FOR UPDATE) right after the users row. The second is still racy for messages the member is about to edit, so a retry is simpler. Severity: availability, low (S4). (Fixed later as H-L10-06: the erasure transaction is run up to three times when the database aborts it as a deadlock; after that the request answers 503 "the server is busy" and the member stays signed in; fix review: h-l10-06.md.)

H-L6-03 together with the erasure's audit scrub (NOTE, by design of both). legacyTalkRemoval (internal/audit/filter.go:116) classes an old per-Talk user.talk_kick row by details.channel_id. The erasure's scrubAudit (account_erase.go:39) sets details = '{}' on every entry about, by or mentioning the erased member. After that scrub, such a row falls to the user.talk_kick auditors row (classes.go:67). Every auditor then sees "member removed at time T, by a moderator (name blank or given)". It no longer names the Talk, so it says what an admin-wide kick says. No Talk leaks.

Tests passing for the wrong reason: none found beyond the FuzzCleanDomain note. The webhook, export-cycle and burst adaptations measure what they did before (section A). The canary and trace lists were checked in deep mode.

SHOULD-FIX

  1. doc/moderation.md:556: "the server log writes each one at the default log level". The line is slog.Warn("audit log: presence queue full, event dropped", "channel", channel) (internal/server/audit_presence.go:43). Warn is shown at the default level, so that part is true. But the line names only the Talk: not the member, and not whether it was a join or a leave. So the log does not keep "each one" in the sense the sentence gives the reader. Fix either side: add "user", user, "joined", joined to the log line (one line of code; nothing new to protect, the server log already carries member ids), or say "the server log notes each drop with its Talk".

What I ran (on 3e337bcc8, nice -n 10, all exit codes read)

  • go build ./... 0; go vet ./... 0; go vet -tags=integration ./... 0.
  • make complexity-check 0, make docs-check 0, make openapi-check 0, make contracts-check 0, node scripts/check-english-only.mjs 0, actionlint .github/workflows/*.yml 0.
  • Full suite go test -tags=integration -p 1 -count=1 ./...: exit 0, 85 packages ok, internal/repo included (internal/server 292 s, internal/sfu 91 s). Spot check that the integration tests really ran (three tests with -v).
  • Lane 7 deep sweep: RACE_SWEEP=1 go test -tags=integration -p=1 -count=1 -timeout 50m -run '<races pattern>' on the five packages: exit 0. MNEMA_DRAIN_HAMMER=1 go test -race -run '^TestDrainHammer$' ./internal/auth/authn/: exit 0.
  • VIS_SWEEP=1 canary sweep: PASS (1171 s).
  • Store trace deep: 20 recorded traces, replay passed.
  • web: npm ci, then vitest run src/lib/media src/stores --maxWorkers=3: 118 files passed, 6 skipped; 1221 tests passed, 5 expected failures, 88 skipped. Exit 0.
  • Probes: cleanDomain with 16 inputs, with and without a4258ed83 (removed afterwards). The erasure/edit deadlock on this tip and on f58bc3fc6 (removed from both trees; kept as a .txt under reviews/).

Not checked

The nightly web sweeps themselves (sim, powers, clocks, cycle, explore, crash, links: hours each). Only their wiring was checked. The SFU deep probes (SFU_SSRC_FLOOD_DEEP, SFU_SUB_SLOPE) were not run. The trust chain sweep was not run. The web suite outside src/lib/media and src/stores was not run (the coordinator ran all of it on an earlier tip). The fixes themselves were not reviewed again; each has its own review.

Delta after 3e337bcc8 (tip c54712470)

Nine first-parent commits, reviewed in a fresh worktree of the tip, which was removed afterwards. Nothing pushed.

Verdict: OK after one fix

MUST-FIX 1: doc/security/reviews/2026-11-hunt/h-l8-08.md must not be in the pull request. H-L8-08 is the fix this branch keeps local because one request was enough (LEDGER: "handled with the S1 procedure anyway"). The brief, section 8, says that for such a finding nothing that shows the attack may go into a pushed branch or pull request before the fix is released. The fix's own code, tests and commit messages are neutral: they say an answer is "refused", and never that anything crashed. The report is not neutral:

  • its title is "answer SDP ends the server process";
  • it names the library's failing line (rtpreceiver.go:544, nil pointer in the library's goroutine);
  • it lists the answer shapes that reach it (a second SSRC in a section, three whitespace variants of a=ssrc-group, a repeated mid). Fix: drop h-l8-08.md and its README row (doc/security/reviews/README.md, the row starting [h-l8-08]) from this branch, and add both in a docs commit after the release that carries the fix. The other copied reports name H-L8-08 only by id, without detail (bundle-integration.md:104 and :257), so they can stay.

No other report holds anything private. A scan of every copied report and the README for local paths, user names, host names, IP addresses other than documentation and loopback ranges, e-mail addresses and AI attribution found:

  • generic paths only (~/go/pkg/mod, /tmp/probe.txt);
  • "OrbStack", a tool name;
  • one "the NAS" with no name or address (h-l8-07.md:236). The README table lists 27 reports, the same 27 files the two commits add, with matching links.

00add7986: lane 5 merge, hand resolution in lockMessageForDelete

The result was compared with a rerun automatic merge. Only three files differ, and each was read:

  • message_mutations.go: the final function locks the message (FOR UPDATE), then calls ownerWords in a statement of its own, then newestOfThread. ownerWords (moderation.go:340) holds the H-L7-07 query of the first parent, with the same text and the same arguments. The lane 5 side had folded the Owner check back into the locking statement as a subquery. That is exactly the form H-L7-07 fixed (a subquery of the statement that waited for the lock does not see the reply that held it). The coordinator correctly kept the H-L7-07 form.
  • newestOfThread (recount_inputs.go:30, max(created_at) over the message and its replies) runs after the lock, in its own statement, so it sees every reply committed before the lock. Once the root is locked no new reply can join. The recount is told this value (handler_messages.go:261), so its input is right.
  • storeTrace.known.ts: emptied from the lane 5 side. Runs:
  • TestAModeratorsDeleteThatWaitedForTheOwnersReplyKeepsIt passed 5 of 5 runs.
  • Revert check: with the ownerWords call removed, the test fails (204, the Owner's reply not kept, bulk delete answers deleted 0).
  • Deep race sweep (RACE_SWEEP=1, the races pattern, five packages): all ok.
  • Empty trace list: the smoke replay (the checked-in traces, every position) passes. A deep run with 20 newly recorded traces (seeds 7000 to 7019, unsharded) passes, so no key breaks any more.

dee2930d6

A clean automatic merge, identical to the rerun. It brings in 1f207e6cb, the store split, which fixreview-H-L5-01-04 round 2 covered.

fc8649030: previewText moved

previewText is now in web/src/shared/lib/markdown.ts. It computes markdownToPlain(content).replace(/\s+/g, ' ').trim(). That is the same function as before, because the old stripMarkdown returns exactly markdownToPlain(content). chat/lib/replies.ts re-exports it, so the five chat importers are unchanged and the event comments import it from shared. npm run lint and typecheck both exit 0.

097ab3878b and 3109606d0

  • 97ab3878b: traceFrame(ev) is a type conversion between identical structs. No behaviour change.
  • 3109606d0: overwrite: true on upload-artifact for a shard's pool. actionlint passes. A rerun replaces the earlier upload of the same shard instead of failing on a name conflict.

078f567c9: TestGeneratorRefusesAMemberWhoMayNotJoin

It is a real behaviour test.

  • Mutation check: with the generator's own mayJoin refusal disabled (generator.go:38), the test fails: "a member who may not join got a relay socket".
  • Positive control (my probe, removed afterwards): the same call with a check that allows the member gets a socket. So the refusals are not caused by a malformed request.
  • It covers refusal and lookup failure and checks that no allocation is held afterwards.

28f359fc2: SHOULD-FIX 1 closed

The new sentence names the hub's voice join and voice leave lines: internal/ws/voice_join.go:98 Info with user and channel, and internal/ws/voice_presence.go:88 Info with user and channel. They log at the default level and name the member and the Talk. True.

Ran on c54712470 (nice -n 10, -p 1, exit codes read)

  • go build 0; go vet 0; go vet -tags=integration 0.
  • make staticcheck 0; make complexity-check 0; make docs-check 0.
  • go test ./internal/repo/ 0; English check 0; actionlint 0.
  • go test -tags=integration ./internal/chat/... ./internal/turnrelay/ 0.
  • Full internal/server integration suite 0 (297 s).
  • web: npm ci, npm run lint 0, npm run typecheck 0.
  • vitest run src/stores src/chat src/events --maxWorkers=3: 211 files, 1984 passed, 2 skipped, exit 0.

Something wrong on this page? Tell us.