Fix review: H-L8-02, H-C-01, H-C-02, H-C-03
From mnema-talk v3.3.13, rendered on this site.
Reviewer: fix-reviewer (read-only). I wrote neither the findings nor the fixes.
Branches, fetched fresh, each checked out in its own temporary detached worktree (removed when done):
- A.
msch128/fix-h-l8-02: fix commit 87378bab3 on the lane 8 test commit 0b742e657. - B.
msch128/fix-h-c-01-02: fix commit 1c69fb153 on e4d472b15. - C.
msch128/fix-h-c-03: fix commit cea35580d on e4d472b15.
I reviewed only the diffs of the fix commits and read the surrounding code as needed. All jobs ran under nice -n 10, go test -p 1, one heavy job at a time. Docker for integration tests came from OrbStack. Every probe file was written into the temporary worktree, run, and deleted. Nothing was committed or pushed.
Summary:
| Fix | Finding | Verdict |
|---|---|---|
| A | H-L8-02: TURN allocation limiter saturated by forged sources | MERGE (notes only) |
| B | H-C-01: "@all @Role" leaves a muted role member at kind group | MERGE AFTER FIXES (one blocker, which belongs to H-C-02) |
| B | H-C-02: users-off webhook or bot "@all @name" names the member | MERGE AFTER FIXES (blocker: the web client still breaks the mute) |
| C | H-C-03 part A: event comment delete dialog shows Markdown source | MERGE |
A. H-L8-02: per-member allocation budget on the built-in TURN relay
Verdict: MERGE.
Does it remove the cause?
Yes. allowAllocation (internal/turnrelay/handlers.go:125-141) now keys allocLimit by the member id parsed from the credential username. Before, it was keyed by the source address. The cause the refuter named was a budget keyed by something the server cannot attribute, and that cause is gone. A member who forges any number of sources draws on one bucket only. The 4096-key table can then fill only with real, distinct members.
I read pion/turn v5.1.3 handleAllocateRequest and authenticateRequest (internal/server/turn.go, util.go) to check the order the author relies on:
- A request without MESSAGE-INTEGRITY gets a 401 with a nonce. No handler of ours runs.
- A bad or stale nonce gets a 401 or 438. No handler of ours runs.
- The AuthHandler (
authenticate) runs. It usesfailLimit.available, thenfailLimit.allowonly for a malformed or expired username. This is unchanged. - The integrity check runs. A forged signature ends here, and
OnAuthfires with verdict false. Our handler returns at once on false. - On success,
OnAuth(true)fires, then the 5-tuple retry check (step 2 of RFC 5766), then the attribute checks, then the QuotaHandler (step 7), and only then CreateAllocation.
So dropping the per-source allocation table opens no unauthenticated path. Nothing before the QuotaHandler ever touched allocLimit, and failLimit keeps its per-source-and-username shape. The finding's TCP question is settled the same way. TURN over TCP uses the same server-level QuotaHandler, so it gets the same per-member key, and limitedListener keeps its per-address connection caps.
Variants:
- Forged sources from one member: fixed (
TestOneMemberFromThousandsOfSourcesCannotKeepAnotherOut). - A banned member saturating with a still-valid credential: now refused in the quota handler before the budget is charged (
TestABannedMembersCredentialIsRefusedBeforeTheLimiter). - TCP clients sharing the table: fixed, because the table is now per member.
Own probes (not in the tests)
Two members behind one address each got their full burst of 60 from the same source. The office NAT case is not penalised.
One member's flood of 10,000 signed Allocates from 10,000 sources caused exactly 1 account lookup. The 8 s verdict cache and the single flight hold.
A check that returns an error (database down) fails closed. The request is refused, the bucket is not charged, and the error is not cached: 100 requests gave 100 lookups. See NOTE A-2.
A malformed username reaching the quota handler is refused. This path is unreachable after
authenticateanyway.Eviction profit with many real accounts (the lead's probe). K accounts each drain their bucket every second for 30 s against
newMemberLimiter(120, 60, 4096). I compared the attempts granted with the budget K accounts have without eviction, K × (60 + 2 × 29):K accounts granted budget without eviction ratio 4096 483,328 483,328 1.00 4097 501,368 483,446 1.04 5000 5,191,430 590,000 8.8 8192 13,429,696 966,656 13.9 See NOTE A-1.
Revert check
I put back the parent's handlers.go and server.go:
TestOneMemberFromThousandsOfSourcesCannotKeepAnotherOutFAILS: "round 0: another member's first allocation failed: ... error 486".TestABannedMembersCredentialIsRefusedBeforeTheLimiterFAILS: "a banned member's signed request passed the quota handler".
I then set evict = false in newMemberLimiter, and TestAFullMemberTableAdmitsANewMemberAndStaysBounded FAILS. TestAFullFailureTableRefusesNobody also passes before the fix. It pins unchanged behaviour, as the commit message says. The hunt's inverse oracles were replaced by normal tests, and the old TestRateLimiter (refuse at the cap for newRateLimiter) is kept.
Findings
- NOTE A-1: a cohort of more than 4096 real accounts resets its own budgets.
limit.go:69-88makeRoom evicts the bucket with the most tokens. When every bucket is drawn down, an evicted member comes back with a full burst of 60.- From 4097 accounts on, a group cycling through the table gets more attempts than its accounts' budgets add up to: 8.8 times with 5000 accounts, 13.9 times with 8192 (probe 5).
- Not a blocker:
- Registration is invite-only, and each account needs voice rights. Thousands of accounts on one community is an insider with admin-visible scale.
- What the attacker gains is allocation attempts. Real allocations stay capped at 4 per member and 256 in total.
- No other member is ever refused because of the eviction, which was the property the finding was about.
- Each eviction costs a 4096-entry scan under the limiter mutex, which runs inline in pion's read loop. That is microseconds.
- The documentation ("the fullest one is forgotten rather than a new member refused") is accurate. The commit message's "the forgotten bucket lost the least" is true per eviction but not over a cycle.
- Possible hardening: give an evicted key back only what it had, for example a small "recently evicted" tombstone. Optional.
- NOTE A-2: a failed account check now runs ahead of the rate limit.
handlers.go:130asksmayJoinbeforeallocLimit. Errors are not cached (handlers.go:83-90), and the check has a 2 s timeout (internal/server/router_build.go:296).- While the database is down or slow, every signed Allocate from a member waits for a fresh lookup of up to 2 s inside pion's single read loop for that listener. Before the fix, this happened only for requests within the per-source budget.
- This is not a regression in practice:
- Before the fix, a member within budget (60 burst, then 2/s) could already keep the loop waiting on lookups of up to 2 s each.
- A spoofing member was not limited at all.
- The verdict is cached for 8 s whenever the database answers.
- The commit message's "Cost on the packet path" does not mention the error case.
- NOTE A-3: OnAuth starts one goroutine per verified request. This is pre-existing code, unchanged by this fix (
handlers.go:107-114,events().OnAuth).- Probe: one lookup hanging for 2 s, then 20,000 calls of the
OnAuthhandler with verdict true for one member. - Result: +20,000 goroutines, back to +0 after the lookup returned.
- So the count is bounded by the signed-request rate × the lookup duration (at most the 2 s timeout). It grows only while a lookup is in flight, which is normally milliseconds every 8 s, or always while the database fails.
- Every request of an allocation that passes the integrity check fires
OnAuth: Refresh, CreatePermission and ChannelBind as well as Allocate. A member flooding signed CreatePermission or Refresh requests during a database outage therefore adds about rate × 2 s goroutines. Send indications carry no integrity and do not count. - It is bounded, so it is not a growth path without limit, and it needs a valid credential.
- Cheap hardening for a follow-up: under
s.mu, check whether a verdict is fresh or a flight exists, and spawn only when neither holds. Then at most one goroutine per member at a time. - I recommend a follow-up rather than blocking this fix.
- Probe: one lookup hanging for 2 s, then 20,000 calls of the
- NOTE A-4: established allocations of a banned member.
- They are dropped at the first definite "no" from
mayJoin(DropMember,handlers.go:96-101). That happens on the member's next signed request after the 8 s verdict expires: a permission refresh, a channel bind refresh or an allocation refresh, which browsers send every few minutes. - Nothing calls
DropMemberat the moment of the ban (grep: the only caller ishandlers.go). - This is unchanged by the fix, and the documentation does not claim more.
- They are dropped at the first definite "no" from
- NOTE A-5: benchmark write-up.
doc/benchmark.md:388-392describes the relay limit ("accepts 120 allocations per minute") as shaping the load-test setup. That was the per-address limit, and it is now per member. It is a historical measurement, so the change is optional: say "per member" or "at the time, per address".
Gates
go test -race -count=1 ./internal/turnrelay/: ok (15.1 s).go test -tags=integration -run 'TURN|Turn' ./internal/server/: ok. This includesTestBuiltinTURNFollowsAccountState,TestWebRTCConfigNeedsSessionAndCarriesTURNandTestBuiltinTURNHealth.- gofmt clean.
go vetwith and without-tags=integration: clean. make complexity-check: clean.go test ./internal/repo/: ok.check-english-only: ok.check-doc-links: ok.- No allowlist, limit or test weakened. Commit type is
fix. No public API changed (Limitskeeps its fields, and only the comment changed). - The documentation in operations.md and troubleshooting-voice.md matches the new behaviour. The "486 Allocation Quota Reached" for a refused banned member is reported as
forbidden, as documented.
B. H-C-01 and H-C-02: mention kinds by the strongest allowed reason
Verdict: MERGE AFTER FIXES.
The server half is correct for every human member, and nothing that was right before changed. But H-C-02's documented effect, a muted member reached through a plain-text name, still happens in the web client: live badge, desktop notification and sound. After a reload, the client now disagrees with the server. A bot named next to @all is still notified.
What the fix does
addRoleMentions(mentions_roles.go:64-99) raises a member who is already in the set at kind group to kind role. A name (kind 0) stays.allowedKind(mentions.go:107-115) demotes kind 0 to group underNoUsers, unless the member is inAllowUsers.- The order inside resolveMentionSet: names and @all first, then allowedMentions, which demotes, then addRoleMentions, which raises. So a demoted member who holds a mentioned role ends at kind role. That is right.
Every reader of the kind, checked against the author's table
- readstate (
readstate.go:54-56):mention_countcountskind <> group, or any kind unlesssuppress_everyone. With the fix:- A muted role member counts "@all @Dev". This is the H-C-01 fix.
- A member who suppresses @all and @here now also counts the role. This matches the client's rule for suppress_everyone (
messageMentionsMeDirectlyreadsmention_roles), so a former mismatch is fixed as a side effect.
- push
namedMentions(recipients.go:174-197):kind <> group.TestPushFollowsTheStoredMentionKindscovers both directions. - thread follow (
thread_follow.go:41-55,thread_follow_sync.go:79):kind <> group.- A role member reached by "@all @Dev" in a reply now follows the thread, as for "@Dev" alone. This is consistent.
- A member named next to @all by a users-off webhook no longer follows. This is consistent with the policy.
- inbox (
search/inbox.go) and thementionslist (messages.go:136-149) ignore the kind, so nothing changes there. - cap (
mentions_cap.go):capRecipientscuts only underNoEveryone. Then there is no @all or @here, so no group kind exists to raise or to create. Demotion needs a member that mentionSkipped kept, which under NoEveryone means an AllowUsers member, and those keep kind 0. So the 50-recipient cut and its kind ordering are unchanged.capNames(100 names) runs before all of this.
- web client (
web/src/stores/chat/notifications.ts:37-46,events.ts:144-146): a muted channel counts a message when the server'smentionslist contains me and (the text names me or a role of mine is inmention_roles). See BLOCKER B-1.
Combinations that were right before and are still right (own integration probe, plus the author's table, which passes):
- An ordinary "@all" in an unmuted channel: everyone at kind group, badge 1.
- A name mention by a member: kind 0.
- "@here @Dev" with an offline role member: kind role. The member is reached by the role alone.
- "@here @all @Dev @NAMEDMUTED" from the owner (my own text, upper-case name):
- plain members: group
- role holders: role
- the named member: name
- the other named members (not in this text): group
- the offline role member: role
- the offline named member: group
- Edit "@all @Dev" to "@all only": the role members go back to group, so edits recompute.
Revert check
I put back the parent's mentions.go and mentions_roles.go, and all of these FAIL:
TestAllowedMentionsDemoteAPlainTextName(unit): "bob 0 (want group), eve 0".TestMentionKindsOfEveryRouteAndGroupTestEditedMentionsGetTheSameKindsTestPushFollowsTheStoredMentionKindsTestAllPlusRoleMentionBreaksTheRoleMembersMuteTestWebhookWithoutUserMentionsDoesNotNameAMemberAfterAll
The two recorded inverse assertions are now normal tests.
Findings
- BLOCKER B-1 (H-C-02 is not fixed for the web client).
Scenario: a webhook with
allow_user_mentionsoff andallow_everyoneon posts "@all @max" to a channel max has muted.- The server now stores max at kind group, so the badge after a reload is 0 and there is no push. That is correct.
- The WebSocket
message_createstill lists max inmentions, because @all reaches max. The text names max, somessageMentionsMeDirectlyreturns true. - So the live badge shows 1, a desktop notification is raised, and the mention sound plays in the muted channel. After a reload the badge is 0.
Probe in the web store (a copy of
chatMentionPolicy.test.ts's setup, payloadmentions: ['me'], content "@all @max deploy", channel muted). Expected count 0 and no notification. Got:expected { count: 1, notes: 1 } to deeply equal { count: +0, notes: +0 }Before the fix, client and server agreed: both counted 1, and both were wrong. After it they disagree. That is the "live badge and badge after reload disagree" symptom H-C-01 described, now for H-C-02.
The new upgrade note and the webhooks.md sentence ("a named member is reached by that group only, which a mute silences") are not true for the web client's live state and desktop notification.
The same holds for the bot route (a bot whose
allowed_mentionsleaves outusersand does not list the member).The client cannot tell from the payload whether a name was allowed. A fix needs the server to say so, for example a list of the members addressed by name or role, or the kind per entry of
mentions. ThenmessageMentionsMeDirectlyreads that list instead of matching the text. The 3.0.9 comment on that function already says it wants exactly this ("a name it left as plain text addresses nobody").The fix also needs a web test mirroring
chatMentionPolicy.test.tsfor "@all @max" from a users-off sender.This is a client and payload change, so the desktop-client coupling rule may apply if the desktop client reads
mentions.
- SHOULD-FIX B-2: a bot named next to @all under NoUsers is still notified, now at a kind that cannot exist. This is the author's own suspicion, and it is confirmed.
- Probe: a users-off webhook (allow_everyone on, parse
["everyone","roles"]).- It posts "@all @
ping". The bot is stored at kind 1 (group) and is in the message's mentionslist. - The same webhook posts "@
ping" alone. The bot is not mentioned.
- It posts "@all @
- Cause:
mentionSkipped(mentions.go:120-128) keeps every id whentargets.allis set. For a bot, though, the SQL result (mentions.go:164-170) is there only because of the name. "@all and @here never include bots" (doc/bots.md:88). - Before the fix the bot got kind 0. Now it gets kind 1. Either way the bot sees its id in
MESSAGE_CREATE.mentions, which is how doc/bots.md:84-87 says a bot recognises a mention of itself. It may act on a mention the policy turned into plain text. - Harmless for badge, push and thread follow: bots do not read state or subscribe, and kind group does not follow.
- But it contradicts the new architecture.md sentence ("A name only counts when the sender may name members") and webhooks.md ("off turns every member mention into plain text").
- Fix: under
NoUsers, a non-allowed id that the group cannot reach must be skipped, not demoted. For example, letqueryMentionedreturn whether the row was reached by the group ($1 AND NOT is_bot, or$3 AND id = ANY($4)), and demote only those. Drop the rest. - Add the bot to the kinds table: "webhook users off @all" gives the bot "-".
- Probe: a users-off webhook (allow_everyone on, parse
- NOTE B-3: messages stored before the update keep their old kinds (upgrade note). I find this acceptable.
- The effect is limited to badges of unread messages and thread follows created from then on.
- Push has already been sent for old messages.
- A backfill would need to re-resolve old texts under policies that are no longer known: the webhook's settings at send time and the online set for @here. That cannot be done faithfully.
- NOTE B-4: "@here @Role". The reproducer reported that "@here @Ops" did not reproduce (count 1 before the fix). The code before the fix gives an online role member kind group through @here and then skips them in addRoleMentions. So "@here @Role" was affected for online members, as the commit message says ("or @here, when online").
- My inference, not re-run on the parent: the reproducer's probe member had no WebSocket connection. @here did not reach them, so the role alone added them at kind role.
- After the fix, the kinds table pins kind role for "@here @Dev" both online (
rolecolumn) and offline (awayrolecolumn).
Gates
- Unit
go test -count=1: ok forinternal/chat/mentions,internal/chat/core,internal/push,internal/chat/readstateandinternal/chat/messages. -tags=integration: ok forinternal/chat/mentionsandinternal/chat/core.- Full
internal/serverwith-tags=integration: ok (346.7 s, exit 0, no port exhaustion) - gofmt clean.
go vetwith and without-tags=integrationon mentions, core and server: clean. make complexity-check: clean.check-english-only: ok.check-doc-links: ok.internal/repo: okmention_kinds_integration_test.gohas 294 lines, under 800.- No allowlist, limit or test weakened. Commit type is
fix. The public API is unchanged:mentionsstays a list of ids, andmessage_mentions.kindis not exposed.
C. H-C-03 part A: the event comment delete dialog shows the plain preview
Verdict: MERGE.
The excerpt is now clampChars(previewText(c.content), 160) (EventComments.vue:67), exactly as the message delete dialog builds it (useMessageActions.ts:32, useMessageMenu.ts:140). previewText is stripMarkdown plus whitespace folding (chat/lib/replies.ts:25-27). clampChars cuts by code points and does not split a combining mark or a ZWJ sequence (shared/lib/textLength.ts:43). The blocked-member branch is unchanged.
Own probes, mounted as in the recorded test:
| Input | Excerpt |
|---|---|
||only a spoiler|| | "Spoiler" |
> quoted **bold**, blank line, line two `code` | "quoted bold line two code" |
| a fenced code block "secret code" | "secret code" (as in the sibling dialogs) |
[link text](https://example.com) ~~gone~~ | "link text gone" |
| "e" + U+0301, 170 times | 80 whole "é" (160 code points), no stray mark |
| the ZWJ family emoji, 170 times | 32 whole families (160 code points), no split sequence |
a moderator deleting a blocked member's comment the **butler** ||did it|| | "Blocked message" (blocks.collapsed), so the 3.0.9 branch still folds |
Revert check: with the parent's EventComments.vue, both tests in EventCommentsExcerpt.test.ts FAIL ("does not show the text of a spoiler or the Markdown marks", "does not cut a character in half"). The it.fails is now it.
Findings: none. One note: an empty preview (a comment that is only whitespace after stripping) gives an empty excerpt. The message dialog falls back to "Attachment", but comments have no attachments, so this is fine.
Gates: ESLint clean on both files. vitest run src/events and the blocked-member test: 27 files, 211 tests passed. vue-tsc: no errors (vue-tsc --noEmit -p tsconfig.app.json). check-english-only: ok. Size: 166 and 53 lines. Commit type is fix. No documentation names this dialog. No allowlist or test weakened.
What I ran
- Fetch. Three temporary worktrees.
npm ciin the C worktree'sweb/. The web probe for B ran there too, because fix B touches no web file (git diff e4d472b15 1c69fb153 -- web/is empty). - Revert checks and all suites and probes as listed in each section.
What I did not check
- A: a live pion server with real forged UDP datagrams (the quota handler was called directly, as in the author's test). The desktop client.
- B:
- The desktop client's use of
mentions. - The gateway delivery of
MESSAGE_CREATEto a bot in B-2: I checked the stored row and doc/bots.md, not a gateway connection. - Why the reproducer saw "@here @Ops" give 1 before the fix (B-4).
- The desktop client's use of
- C: the dialog's visual rendering.
Round 2: msch128/fix-h-c-01-02 (21c17b742, 453e3e965 on 1c69fb153)
Verdict: MERGE. B-1 and B-2 are resolved. No new defect found. Two notes on the wording of the documentation.
What changed
- 21c17b742 (B-1).
- Every message payload now carries
direct_mentions: the members ofmentionswhose stored kind is not group. It is filled inmessages.loadMentions, the one place that fillsmentions(kind <> groupin the same query).ScanMessagesand the fallback inpollstart it as an empty list. messageMentionsMeDirectlychecks in this order:- a reply to me;
- when the field is present, whether my id is in it;
- otherwise the old rule (the text or a role of mine).
- Every message payload now carries
- 453e3e965 (B-2).
queryMentionedreturnsgroupedfor each row:NOT is_bot AND (@all OR (@here AND online)).mentionSkippedkeeps a user who is not allowed byNoUsersonly whengroupedis true. A bot that only its name reaches is dropped.
Probes
- Web store (a copy of the setup in
chatMentionPolicy.test.ts; the channel is muted):- A users-off sender posts "@all @max", with
mentions: [me]anddirect_mentions: []. Result: live badge 0, no desktop notification. AfterfetchReadState(server value 0) the badge is still 0, so the live badge equals the badge after a reload. - A role message whose text does not name me, with my id in
direct_mentions: badge 1. - A reply to me with an empty
direct_mentions: badge 1. A reply always breaks the mute. - A frame without the field (an old server): the old text rule applies, badge 1.
- Sound and thread follow:
- Both are decided by the same function.
messageSound.ts:36andthreadAllows(notifications.ts:75) callmessageMentionsMeDirectly, so they follow the same answer (checked by reading the code, not by playing a sound). - The server's thread follow and push read
kind <> group. That is the same set asdirect_mentions.
- Both are decided by the same function.
- A users-off sender posts "@all @max", with
- Every other web reader of
mentions. Checked by grep.messageMentionsMe(the highlight of a message, and the unmuted badge) still readsmentions. That is right: there @all counts.- With
suppress_everyoneit goes throughmessageMentionsMeDirectly, so it now matches the readstate rule (kind <> group). - No other reader of message mentions exists in
web/src.
- Push.
namedMentionsandthreadAudienceread the stored kinds and followers, not the payload. No change was needed there. - Bot route for B-2 (integration probe, which the author had not tested). A second bot exists. The first bot posts "@all @otherbot @zed" with parse
["everyone"]:- otherbot is not in
mentions; - zed is in
mentionsand not indirect_mentions. - With
allowed_mentions.users: [zed], zed is indirect_mentions.
- otherbot is not in
- message_update.
- A member edits "@all hello" to "@all hello @zed": zed is in
direct_mentions. - Editing it to "@all hello again":
direct_mentionsis[]and zed stays inmentions.
- A member edits "@all hello" to "@all hello @zed": zed is in
- REST history. None of the messages lacked the list. Payloads with
direct_mentions: nullin the golden files: 0. - Human members (the kinds tables).
TestMentionKindsOfEveryRouteAndGroup,TestEditedMentionsGetTheSameKindsandTestPushFollowsTheStoredMentionKindsstill pass. The only change for humans is the bot cell.
Compatibility
- Old client against the new server. The generated decoders check only the keys they know (
hasOwn), so the extra field is ignored. The old live behaviour remains until a reload, as upgrade.md and webhooks.md say. - New client against an old server. The field is missing, so the client uses the old rule. The new decoder accepts a missing field or an array of ids. It would reject
null, and the server never sends null. - doc/api-stability.md ("Not breaking: adding ... a response field"). The field is additive and
mentionskeeps its meaning.make openapi-diff: "no breaking changes since v3.0.9".make openapi-check: exit 0. Theapi/events.jsonmessage events refer to the shared schema, and the registry tests pass. - Bots' MESSAGE_CREATE (the gateway sends
core.Messageunchanged). The field is additive there too. - Desktop repo. The claim holds.
linux/andexperimental/app/handle unread and mention counts that the web app reports, plus notification settings. Neither parses message payloads: grep finds nomessage_createand no.mentionsof a message.
Revert checks
- 453e3e965, code reverted:
TestMessageCreateSaysWhoIsAddressedDirectlyFAILS ("a users-off webhook mentions the bot it names next to @all").- The unit test no longer compiles against the old signature (
mentionHit). The integration test carries the behaviour check.
- 21c17b742, server half (
loadMentionsno longer fills the list):TestMessageCreateSaysWhoIsAddressedDirectlyFAILS ("want the role holder and not the plain-text name"). - 21c17b742, web half (
notifications.tsreverted): "a name next to @all from a sender that may not name members does not break the mute" FAILS.
Notes
- NOTE R2-1: "before 3.0.10" in webhooks.md.
- This is safe only if the fix ships in the next release. Release-please makes that 3.0.10 for a
fix:on top of v3.0.9. - If a 3.0.10 is tagged before this branch merges, the sentence is wrong.
- The other docs that name a version name one already released ("since 3.0.0", "in 1.2.0").
- Suggestion: "a web app that does not know
direct_mentions". Otherwise, check the version of the release PR when merging.
- This is safe only if the fix ships in the next release. Release-please makes that 3.0.10 for a
- NOTE R2-2: doc/bots.md (Mentions, lines 84-88) says nothing of
direct_mentionsinMESSAGE_CREATE, and the field is in the bots' payload too. This is optional: bots may ignore unknown fields.
Suites and gates
- Unit:
./internal/chat/...,./internal/poll/and./internal/events/...all ok. - Integration: mentions, messages and poll ok. The full
internal/serverintegration run is ok (235 s, exit 0, no port exhaustion). - Web: full vitest gives 666 files passed (2 skipped), 12,058 tests passed, 1 expected fail, 58 skipped. vue-tsc exit 0. ESLint clean on the touched files.
- gofmt clean. go vet with and without the integration tag clean.
make complexity-check,internal/repo,check-english-onlyandcheck-doc-linksall ok. - No allowlist, limit or test weakened. Both commits are
fix.