Review of PR #914: timed thread replies, display name cleaning, bot channel options, command permission storage, durable-commit file cleanup

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

Two rounds. The first reviewed the five changes as first pushed; its should-fix findings were fixed in the same pull request and checked in the second round (at the end).

Round 1

Verdict: MERGE AFTER FIXES. No blocker. All five changes are correct, the tests fail without the fixes, and the reviewer found no regressions. Several read paths still used the old lifetime check instead of core.LiveSQL, two of them (new report copies, the bulk delete's audit entry) writing stored copies that outlive the sweep, so doc/timed-messages.md promised more than the code did.

What was checked

1. A reply ends with its timed root.

  • LiveSQL(alias) is ((a.expires_at IS NULL OR a.expires_at > now()) AND NOT EXISTS (... ended_root.id = a.parent_id AND ended_root.expires_at <= now())). Neither part can be NULL, so NOT LiveSQL(x) (saved list, media serving) means "this row ended, or its root did". A root has parent_id IS NULL and never matches the subquery.
  • All 13 callers embed the expression after AND, and it is fully parenthesised. pins.go uses LiveSQL("messages") on an unaliased FROM messages; the inner table is ended_root, so messages.parent_id is the outer row. No caller uses live or ended_root as an outer alias. MessageSelect embeds LiveMessages twice, each in its own scope. TestLiveMessagesSpellsOutLiveSQL keeps the constant and the function together.
  • Through MessageSelect / LiveMessages: history, threads, search, inbox, bot API reads, forum posts, announcements, polls, interaction responses and reply previews leave out replies of an ended root; saved, reminders, forward, reactions, media serving and the reported message itself too.
  • Still on the raw check: report context, the bulk delete, the conversation list, the account export, the edit and reaction locks (findings below). The other raw uses only see roots (forum activity, FirstUnread, /thread, reply counts).
  • Gateway resume replays no stored messages.
  • Performance, EXPLAIN ANALYZE on a scratch PostgreSQL 18 with 1M rows and the production indexes: channel history of 50 roots 0.23 ms before, 0.08 ms after (anti join on the primary key, no probes for roots); trigram search with 751 hits 8.5 ms before and after; thread replies 0.08 ms.

2. Display names. cleanDisplayName (httpx.CleanName) is used for registration, UpdateProfile and UpdateDisplayName; /nick goes through UpdateDisplayName. No other path writes a member's display name from input (the admin seed uses the username, bots and webhooks already use CleanName, the erasure writes a fixed string). C0 controls are now refused; the test covers it. On the web, mentionName strips the direction and zero-width characters, keeps ZWJ and ZWNJ, and falls back to @username; unicode-bidi: isolate on .md-mention is right for an inline span.

3. Bot command channel option. The botUser passed in interactions.go is the command's bot. Built-ins pass uuid.Nil and keep the member-only check (none has a channel option today). Categories never get View and fail the type filter; conversations and Talks of conversations fail the bot view and dm_id IS NULL; threads are messages. The SQL filter matches resolveChannels.

4. storedDefaultPerms. A value above MaxInt64, unknown bits or an unparsable value is refused instead of stored as NULL. Clean already refuses them, so this is defence in depth.

5. ErrNotDurable. Both error returns of AwaitDurable wrap it with %w: %w; errors.Is follows it through "record upload: %w". Of every caller of the insert functions only the media upload and the webhook delete objects on error, and both are fixed. Scheduled sends delete their row in the insert transaction; the nonce replay has nothing to clean up.

Tests, limits, docs. On a copy with the origin/main version of each fixed file, TestAReplyEndsWithItsTimedRootBeforeTheSweep, TestAnUploadCommittedButNotConfirmedKeepsItsFile, TestAWebhookFileOfACommittedMessageStaysWhenTheWaitFails and TestSlashCommandRoundTrip fail; the two unit tests call functions that do not exist on origin/main and assert the opposite of its behaviour. Largest changed file 386 lines, largest test 578; gocognit at most 13 on changed functions. doc/bots.md matches; no documented response in api/openapi.json changed.

What was run

  • The four integration tests above, alone and on the copy without the fixes (fail as expected).
  • go test -count=1 on chat/core, auth, bot, db, media, webhook, chat/...: ok. go vet on the changed packages: clean.
  • npx vitest run src/shared/lib/markdown.test.ts: 31 passed.

Findings

Should-fix:

  1. internal/chat/reports/reports.go:94,96,100: the neighbours of a reported message use the raw check, so a report filed after a root ended and before the sweep stores the thread's replies in its snapshot.
  2. internal/chat/messages/moderation.go:222,224: the bulk delete counts a reply of an ended thread and writes its text into the audit entry.
  3. internal/chat/dm/dm.go:135,151: the conversation list can show such a reply as the last message.
  4. internal/chat/account/account_export_messages.go:36, account_export.go:24: the export includes the member's own replies of an ended thread.
  5. doc/timed-messages.md:121 promised exports and new report copies leave such replies out.
  6. Names stored before the change are only cleaned inside web mention pills; author lines and member lists show them as stored.

Nits: the edit and reaction locks (message_mutations.go:77,222) still use the raw check (the routes are refused earlier); two different refusal texts for one rule in commands_validate.go; the conversation check is defence in depth and untested on its own; the web strip list is narrower than the server's (U+2060 to U+2064, U+00AD); the option type comment in commands.go could say what the docs say. Not changed by this pull request: after ErrNotDurable an upload still answers 500 for a stored message; bot/components.go loadComponentMessage has no lifetime check.

Not checked

Admin-only reads (media listings, audit log viewer), unread counts in readstate.go (counts, no content), the desktop and mobile clients beyond the shared web bundle, the full integration suite, production query statistics.

Round 2

Verdict: MERGE. All round 1 should-fix items are fixed; three nits remain.

What was checked

  • S1 to S4: report context (the anchor and both neighbour selects), the bulk delete (m and the nested r), the conversation list (last activity and the preview) and both account export queries use core.LiveM or core.LiveSQL with the alias they embed; editTx and lockMessage use LiveSQL("messages") against an unaliased FROM messages. FOR UPDATE OF m and FOR NO KEY UPDATE still work with the added subquery. TestLiveMSpellsOutLiveSQL keeps the constant and the function together.
  • S5: doc/timed-messages.md matches the code for every read it lists.
  • S6: the character classes of migration 0111, decoded, are exactly the web NAME_FORMAT set (U+00AD, U+061C, U+180E, U+200B, U+200E, U+200F, U+202A to U+202E, U+2060 to U+2064, U+2066 to U+206F, U+FEFF); ZWJ and ZWNJ stay. A name with nothing left becomes LEFT(username, 24), the same as defaultDisplayName for ASCII usernames. The migration does not strip the wider set that CleanName removes on input (all Cf, private use, C1 controls, blank-glyph letters); the direction and zero-width characters are covered. The doc/upgrade.md row is accurate.
  • Push: the rewritten mentions reach the notification body only through clean(plainLine(...)) (internal/push/payload.go:182), which turns controls and every Cf except ZWJ and ZWNJ into spaces; sender and channel names go through the same clean. The round 1 concern about pushes does not hold.
  • Nits of round 1: one refusal text, "a channel you and the bot can see", for every bot-command case (built-ins keep "a channel you can see"); the option type comment and api/openapi.json say what doc/bots.md says.
  • Size limits: largest changed file dm.go at 390 lines.

What was run

  • TestAReplyOfAnEndedThreadIsNotCopiedOrEdited, TestAReplyEndsWithItsTimedRootBeforeTheSweep, TestTheMigrationCleansStoredDisplayNames, TestSlashCommandRoundTrip: ok.
  • go test ./internal/chat/... ./internal/bot/ ./internal/auth/ ./internal/push/: ok. scripts/check-english-only.mjs: ok. Vitest markdown.test.ts: 31 passed.
  • TestAReplyOfAnEndedThreadIsNotCopiedOrEdited on a copy with the previous versions of reports.go, moderation.go, dm.go and message_mutations.go: fails on the report copy, the bulk-delete audit entry and the conversation list.

Remaining findings (nits)

  1. The edit assertion does not prove the editTx change: the handler's LoadMessageRef already refuses the edit, so the lock change is defence in depth; the account export change has no test of its own.
  2. Migration 0111 held its format characters literally, invisible to readers. Fixed after this round: the classes are written as \uXXXX escapes (ASCII only), and TestTheMigrationCleansStoredDisplayNames passes on that file.
  3. Clients see the cleaned names after a reload: the migration sends no user_update, and it runs at startup before clients connect.

Not checked in round 2

The full suite was not run again by the reviewer (it passed for the author on this state); no new EXPLAIN (the plan shape is unchanged from round 1).

Something wrong on this page? Tell us.