Independent review reports

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

Reports of the independent, read-only reviews that a change under the paths in .github/CODEOWNERS gets before it is merged (rule 9 in AGENTS.md). A reviewer gets the code and the task, not the author's conclusion, changes nothing, and writes down what it checked, what it ran and what it did not check. The reports are kept as written; a finding that was fixed later stays in the report.

The reviewers so far were agents, not people. A report here is no external audit.

2026-11: hardening run on 3.0.0 (releases 3.0.1 to 3.0.7)

ReportChangeSubject
828#828shared links use the server's public address (public_url)
771-third#771, merged with #780metadata removal for JPEG, PNG and WebP, third round
775#775, merged with #780repair of the media group protocol
777-keys#777key rotation (JWT_SECRET_PREVIOUS, DATA_KEY, rotate-keys)
777-ws#777panic isolation in the hub and the SFU
838#838the vanity link's Link active switch and the inactive page
779#779group merge, first round (rejected)
779c#779group merge, third round
779d#779group merge, fourth round
779-final#779the three lines that changed after the fourth round, reviewed after the merge
831#831members creating invites, and inviting into a channel
780#780the merge train as a whole
780-events#780typed events
780-outbox#780webhook queue written in the transaction, health under a storage outage
780-tt#780Trusted Types and the CI changes

2026-11: follow-up to the hardening run (3.0.8)

ReportChangeSubject
789#789timer panics, fail-closed permission enforcement, live storage health; two rounds: the should-fix findings S1 to S4 of the first were fixed in the same pull request and checked in the second
790-sweepthe nightly simulator sweepthe settle table of the 43 failing seeds, the sweep of seeds 1..400 with a failure budget, a test for the key package renewal rule; its two should-fix findings were fixed in the same pull request
798#798the data export after a desktop download: one more hour after the first complete download, the length sent, a single download claim that holds under parallel requests; two rounds: the should-fix findings of the first were fixed in the same pull request and checked in the second
bot-platformthe bot platformmanifest install, message components and modals, interactions, both SDKs and the release packaging; seven rounds, every finding fixed in the same branch and checked by reverting its fix
bot-platform-webthe bot platformthe web UI: the add-bot flow, components, selects, forms and their states, at six viewports in two browsers; two rounds
webhook-call-chat#803an outgoing webhook of a conversation's call Talk heard its call chat (releases up to 3.0.13; fixed in 3.1.0): the routes, the queue insert and the delivery each refuse a private channel
816#816the flaky ws, sfu and macOS stage tests fixed at their cause: each stated cause checked against the code, two reproduced on main, and each test shown to fail again when the product behaviour it guards is broken
819#819a Talk join sent before the realtime socket was up was dropped and the call never connected: the join waits for the socket, backs out with its microphone released, and a reconnect sends no resume beside it; two rounds: both findings of the first fixed in the same pull request and checked in the second
824#824each lower camera layer encodes its own copy of the capture, so a bandwidth-cut best layer no longer shrinks the thumbnail; the clones' lifecycle (camera light), the frame encryption of their lines, and the e2e noise filter turned off from the fixture
815-delta#815messages with a lifetime, the last delta: new, duplicated and conversation Talk channels answer the auto-delete they follow (every channel INSERT ... RETURNING checked), the narrow header chip, the admin Messages tab heading and the layout spec; MERGE, two non-blocking suggestions (the GoWaited comment is fixed in the same pull request)
829#829a bot can be given access to every channel in one step: the install and settings endpoints under the bot invariant, the snapshot rule that keeps private channels and categories hidden from such a bot, replace semantics when switching back; three rounds, the findings of the first two fixed in the same pull request
822#822the lobby's join and a join during the microphone prompt could be dropped or doubled around a socket drop: both wait for the socket and a reconnect sends no resume beside a held join; two replay ideas (a leave, a stop) were rejected in review as ineffective or unsafe, and the lost leave stays a ledger follow-up
836#836each lower screen layer encodes its own copy of the capture, as #824 did for the camera: the clones' lifecycle (sharing indicator), a quality change, reconnects, the frame encryption of their lines; two rounds: the required finding of the first (a stale hint after a quality change) was fixed in the same pull request and checked in the second
clip-sweep#820the full clipping sweep moves out of ci.yml into a nightly workflow that the release gate judges; the gate reads every run of the last 49 hours, so pull request runs cannot push the deciding night off a page; one round
850#850a leave lost while the socket was down no longer leaves a ghost seat for the server's grace time: a seatless connection's voice_leave naming its Talk and tab ends the lost seat (same user, same tab only), including a stale connection the server has not noticed yet; the client sends it after the reconnect
851#851bots can download attachments of messages they can see: unified authorization and streaming path with members, forced attachment disposition, existence privacy (404), Range and 304 leak prevention, mutation checks

2026-11: security audit of the features after 3.0.10 (3.3.x)

ReportChangeSubject
914#914replies of a timed thread root end with it for every read and copy, display names without direction and invisible characters (also stored ones, migration 0111), a command's channel option needs the bot's view, a command restriction that cannot be stored is refused, a stored message keeps its file when the durability wait fails; two rounds
mnemo-39talk-mnemo-bot #39Mnemo's attachment reading: PDF extraction bounded in time, memory and length, a pixel limit, names in the consent prompt as plain text, refusals under the rate limits, no file text in thread memory; two rounds
bot-sdks-go-9-js-18bot-sdk-go #9, bot-sdk-js #18the bot token stays with the configured server (resume address, redirects), interaction tokens out of errors and logs, bounded answers; two rounds each
931#931the test shard script excludes packages by Go pattern resolved with go list instead of a regex from the command line (CodeQL js/regex-injection); the CI shards stay the same

2026-11: defect hunt on 3.0.8

Each finding of the hunt went to two agents that had not written it (one reproduced it from a clean checkout, one tried to defeat it), and each fix to a reviewer that had written neither. These are the fix reviews; several took two or three rounds, and a round that ended in "do not merge" is part of the report.

ReportFix for
h-l1-01-04four gates and tests that did not check what their stated reason said (complexity gate, a false skip, the compatibility baseline, the release checks)
h-l4-01leaving a group conversation while seated in its Talk
h-l3-s1-01edits and poll votes after a block in a conversation of two
h-l3-12the chat of the conversation's Talk after a block
h-l3-01-04link rules: the server judges every host the client could open
h-l3-05-06a muted channel and resolved mentions; snippets and pushes hide what the renderer hides
h-l3-09wiki titles compared by one Unicode-aware key
h-l3-10a blocked member's words folded on every surface
removal-fixthe removal of silent members in the media group protocol (the committer waits for newcomers; a renewal that won a conflict counts toward the limit), three rounds, with a design note: a mitigation, not a cure
h-l3-07-h-l2-01-h-l3-11document sentences that differed from the code; the composer's message text

One change in this set was made by the coordinator and reviewed by nobody else: after "Show", focus goes to the message's row in the place where it was pressed (web/src/shared/lib/revealFocus.ts).

2026-11: defect hunt, second part (lanes 5 to 11, on 3.0.9)

The same procedure. Several findings of this part came from the fix reviews themselves, and a review that ended in "merge after fixes" stays in the report with its later rounds appended.

ReportFix for
h-l8-08the SFU vets a client's answer against its own offer (media sections, one stream per section) before the connection reads it; two rounds
h-l8-07caps on RTCP streams nothing reads on a media connection; four rounds
h-l8-07-test-changea test of the stream caps that the coordinator corrected just before the merge (a rare stray buffered stream), reviewed after the release: right as it is
h-l8-09memory kept per subscription switch: the connection is renewed at a sender cap, a resume budget per publisher and source; three rounds
h-l8-03, h-l8-04-06the bandwidth estimator forgets a stopped sender; a hub map entry left at zero; a counter for dropped presence audit entries
h-l8-02-h-c-01-03the TURN relay's allocation budget per member; mention kinds next to @all (the field direct_mentions); an event comment's delete dialog; two rounds
h-l9-01-03the SFU's stored times on a steady clock; sequence numbers and picture ids across a layer switch; two rounds
h-l2-04the media group protocol: an add the member never saw relayed; two rounds
h-l2-05-h-l9-02the handshake pacer and the back-offs on the monotonic clock; with a measured judgement of the availability under a hostile relay and the trace of five simulator seeds (one new finding, not fixed: an equal-epoch fork)
lane2-harnessthe protocol sweeps themselves (replays, roster lies, clocks, crash points): three rounds; the first found that two perturbations never ran
h-l5-02-03-05, h-l5-01-04the web client's lists after a reconnect and a delete; the follow before the first reply; unread counts corrected after a delete or edit; read-state answers ordered against live events (migration 0104)
h-l6-01, h-l6-01-round2-h-l6-03audit entries of a private forum and of a removal from a private Talk reach only auditors who can see them
h-l6-02a hidden message, category or channel answers like a missing one; three rounds
h-l6-04, h-l6-05-06-07a conversation's call Talk: membership decides, and it is no community channel anywhere (events, the admin channel routes, the media list, reports, auto-delete)
h-l7-01-03, h-l7-06-07, h-l7-09limits and rules under concurrent requests: reset links, the first Owner, role templates, the group conversation cap, the Owner's reply and a pinned reply against deletes and the auto-delete sweep
h-l7-04-05, h-l7-08audit entries inside the change's transaction; answers read inside it; no background job after the shutdown drain began
h-l10-01-04, h-l10-05account erasure and export: what was left out; nothing new may name an erased account (migration 0103), with a dump and restore comparison
h-l11-01-03the security floor of the desktop bundle: the trailer is read the same way by the pull request check and the release; Go reads signed documents as the desktop clients do; three rounds
h-l10-06an account deletion aborted by the database as a deadlock with the member's own write is retried (found by the integration review, fixed after 3.0.10)
follow-upsevent comments and event texts go through automod; reactions compared in one emoji form
release-gatea red, stale or missing night of the deep sweeps blocks the release, with one deliberate override; more paths in CODEOWNERS; the known open defects named in SECURITY.md; two rounds
bundle-integrationthe collection branch as a whole: every merge with a hand resolution, the coordinator's own commits, interactions between fixes

Not kept from the hardening run (lost with a temporary directory before they were copied): the review of #768 (permission checks refuse without a service), the first two rounds for #771 and the second round for #779.

Something wrong on this page? Tell us.