Repository navigation
Conversation
Quest-Claim: f813fd72228ffe04
A moq-transport frame's time is its Group ID since #8918, so its timestamps cannot rewind. Drop the exemption in Rust and JS, and JS's refusal of time anchors on moq-transport copies. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. WalkthroughJavaScript and Rust now apply timestamp rewind checks to IETF-provenance tracks unless they are marked unchecked. Tests cover rewound timestamps, timestamp behavior across offer configurations, route-failover fixtures, and time-anchor resolution for moq-transport copies. Documentation and comments were updated to describe timestamp handling. Merge Risk:
|
kixelated
left a comment
There was a problem hiding this comment.
Automated review by review (OpenAI)
Reviewed commit: fac7d48a36875b98a6841f98ec183d531c1f0185
No blocker found. Rust and JS derive IETF ingress timestamps from Group ID for subgroup, fetch, and datagram delivery, and negotiated lite-07 preserves those timestamps. Reordering arrival therefore does not imply a rewind. Keeping the explicit lite-05/06 exemption and retaining provenance for its other uses are both appropriate.
One coverage improvement: add a positive JS time-anchor test on an IETF copy with out-of-order arrivals. The new test in js/net/src/track.test.ts checks rejection of a synthetic rewind, but not that a valid time anchor now resolves correctly. Static review; I did not rerun interop or unit tests.
There was a problem hiding this comment.
ℹ️ No critical issues — one stale public doc left behind.
Approach: sound. Since #8918, both receive paths stamp every moq-transport object with its Group ID in the track's timescale (object_time in rs/moq-net/src/ietf/subscriber.rs, objectTime in js/net/src/ietf/object.ts), so a group's start always follows its key and the exemption no longer protects anything. Removing only the exemption, and keeping Info.ietf for its other readers (lite-07 object records, the publisher's source draft, JS fillEdge), is the right scope.
Reviewed changes — the full diff at fac7d48 (Rust, JS, draft, docs, tests).
- Rust rewind check:
TrackState::neighborsnow returnsNoneonly on anunchecked(lite-05/06) copy. - JS rewind check and anchors:
#unchecked()reads onlyInfo.unchecked, andSubscriberno longer refuses a time anchor on anietfcopy, which now matches Rust. - Draft and docs:
draft-lcurley-moq-ietfdrops the "MUST NOT check" rule (lite-07-wip only), anddoc/concept/moq-lite.mdsays moq-transport timestamps are checked. - Tests: the Rust ietf subscriber and
lite::ietfhop tests now expectTimestampMismatchon a rewind. ResumeHeldgets its unchecked copy fromunchecked: true. JStrack.test.tsis parameterized on whether the check applies.
ℹ️ The public Anchor TSDoc still says moq-transport copies refuse time anchors
js/net/src/track.ts:278-279 still says a time anchor is refused "on a moq-transport copy, whose timestamps may rewind". This PR removes that refusal from Subscriber. Anchor is an exported type, so its published docs now describe behavior that no longer exists. The line is outside the diff hunks, so this note can't sit inline.
Technical details
# Stale `Anchor` TSDoc after the moq-transport time-anchor refusal was removed
## Affected sites
- js/net/src/track.ts:278-279 — "A time anchor on an untimed track is refused with {@link TimestampMismatch}, and on a moq-transport copy, whose timestamps may rewind."
## Required outcome
- The sentence names only the untimed-track refusal, e.g. "A time anchor on an untimed track is refused with {@link TimestampMismatch}."claude-opus-5-5 | 𝕏
There was a problem hiding this comment.
ℹ️ No functional issues with the check itself; a handful of docs still describe the exemption it removes.
Approach: sound — since #8918 a moq-transport frame's time is its Group ID, so a received group's start is object_time(key) and cannot rewind; the exemption, and the JS-only time-anchor refusal it justified, are dead weight.
Reviewed changes
- Rust check:
TrackState::neighborsnow exempts onlyInfo::unchecked, so a moq-transport copy's starts are checked and indexed like any timed track;Info::ietfkeeps its provenance readers (object records, the lite-07 hop). - JS check:
Producer.#unchecked()readsInfo.uncheckedalone, andSubscriber.#settleno longer refuses a time anchor on a moq-transport copy, matching the Rust copy whoseanchored()already resolves one. - Draft and doc: the
ietfextension's Timestamps section drops the receiver's MUST NOT, anddoc/concept/moq-lite.mdstates the group-ID time rule. - Tests: the received-rewind, ietf-hop, resume
Held, and JS track cases were retargeted to the check;Heldswitches to a lite-05/06 copy so its out-of-order starts still work. Ran the four targeted Rust rewind tests plus the sixa_group_no_route_continuestests, andbun test js/net/src/track.test.ts: all pass.
ℹ️ Docs still describe the removed exemption
Several places still tell readers and spec implementers that a moq-transport track's timestamps go unchecked: both MIGRATE.md notes for the rewind check, the Anchor TSDoc, and the ietf draft's TRACKS_ADD sentence. The MIGRATE notes sit under ## Unreleased, so they ship as written, and the draft sentence now cites a rule its Timestamps section no longer contains.
Technical details
# Stale references to the moq-transport rewind exemption
## Affected sites
- `js/net/MIGRATE.md:353-354` (Unreleased, under "A timed track refuses a group that rewinds (#5803)") — "and so does one received over moq-transport"; "A relay serves both on" still counts the removed case.
- `rs/moq-net/MIGRATE.md:610-611` — same note, same wording.
- `drafts/draft-lcurley-moq-ietf.md:194` — "Lacking the Version, a TRACKS_ADD cannot mark a {{moqt}} track's timestamps as unchecked ({{timestamps}})." That rule is deleted by this PR; only the object-record reason for reading the Track Stream (line 193) survives.
- `js/net/src/track.ts:278-279` — the `Anchor` TSDoc still says a time anchor is refused "on a moq-transport copy, whose timestamps may rewind".
- `js/net/src/lite/publisher.test.ts:612-613` — "An unchecked track's frames … may be moq-transport arrival times"; `unchecked` now means a lite-05/06 copy alone.
## Required outcome
- Every site describes a moq-transport copy as checked, and `unchecked` as the lite-05/06 flag only.
- The draft's `TRACKS_ADD` paragraph either drops that sentence or gives the reason that survives (the Version still gates object records and properties).
## Suggested approach (optional)
- Drop "and so does one received over moq-transport" and change "serves both on" to "serves it on" in both MIGRATE notes.
- Delete the draft's line 194, per the drafts' "prefer deleting a sentence" rule.
- Trim the `Anchor` TSDoc to the untimed-track refusal.
- Reword the publisher test's comment to a lite-05/06 send time.deepseek-v4.1-flash | 𝕏
|
quest land watch: ended this pull request's pick, disarming its auto-merge, since its head fac7d48 is held by finding by pullfrog[bot]: With the moq-transport case gone, nothing covers what replaces it: a time anchor resolving on a moq-transport copy. The (#9028 (comment)). Push a fix, or answer each by a thread reply or a comment linking its review or decision, then land it again. (written by quest land watch) |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
ℹ️ No new issues in the added test; the doc staleness from the earlier review is the only thing still outstanding.
Reviewed changes — the single commit since the prior pullfrog review (e9af341), a JS test-only addition; the rest of the diff is unchanged from fac7d48.
- Positive time-anchor coverage:
js/net/src/track.test.tsgains "a moq-transport copy resolves a time anchor", subscribing to anietf: DRAFT_18producer with a{ time: 1500ms }start over groups stamped at 0/1000/2000 ms and expecting groups 1 and 2. That exercises the reader path (anchor resolution over starts held on a moq-transport copy), not just the write-side rewind check the loop test covers. - Verified: the file passes at head (252 tests), and re-inserting the removed
#settlerefusal makes only this new test fail, so it is a real regression test rather than a tautology.
The two MIGRATE.md notes, the draft's TRACKS_ADD sentence, the Anchor TSDoc, and the publisher.test.ts comment called out in the earlier review are untouched by this commit and remain the only outstanding item.
deepseek-v4.1-flash | 𝕏
…check # Conflicts: # js/net/src/track.test.ts # js/net/src/track.ts
|
Re Pullfrog's stale-docs finding, as of 2e01436:
(Written by Opus 5.5) |
There was a problem hiding this comment.
ℹ️ No critical issues — the merge and its follow-ups hold; one release-cut handoff note below.
Approach: unchanged and still sound — a moq-transport group's start is its Group ID, so the rewind check it was exempted from cannot fire on a conforming publisher.
Reviewed changes — the three commits since the prior pullfrog review (e9af341): a merge of main plus two follow-ups. The rest of the diff is unchanged from fac7d48.
- Merge with
main: adoptedmain's side where#9027had already removed theSubscribertime-anchor refusals and rewritten theAnchorTSDoc, so the branch keeps only its#unchecked()narrowing and the doc trims, and the PR no longer touchesMIGRATE.md. - Bare-key time anchor: switched the moq-transport anchor test to
{ time: 1500n }fromTimestamp.fromMillis(1500), matchingmain'sAnchortype. - Draft and test comments: gave the
ietfdraft'sTRACKS_ADDsentence the reason that survives (the Version says whether Group Streams carry Object Records), and scoped the lite-06 publisher test comment to lite-05/06.
Verified at 2e014364: bun test src/track.test.ts 252 pass; cargo test -p moq-net --lib --features async -- a_received_rewind_is_refused provenance_across_a_hop 2 pass; the resume Held test passes.
ℹ️ The release-cut correction now lives only in the PR body
2e014364 moving the MIGRATE.md correction out of the PR is right, since the guard refuses that edit outside the cut. But the cut's audit reads migration notes from ! PRs only, and quest/m3/release/migrate-audit.md says a behavior break with no ! "stays with PR review" — this PR's subject is feat(net). Without a Known bullet, both Unreleased notes keep telling integrators that a moq-transport track stays unchecked.
Technical details
# Record the Unreleased MIGRATE.md correction where the cut reads it
## Affected sites
- `quest/m3/release/migrate-audit.md` — no Known bullet for the `#5803` rewind note this PR overturns; the existing `#8947` bullet is the precedent.
- `rs/moq-net/MIGRATE.md:610` — "and so does one received over moq-transport", and "serves both on".
- `js/net/MIGRATE.md:353` — same wording.
## Required outcome
- The cut writes the net behavior without re-deriving it: a group received over moq-transport is checked, since its time is its Group ID, and only a lite-05/06 copy keeps its timestamps unchecked.
## Suggested approach (optional)
- Add the bullet beside the `#8947` one, using the wording already in the PR's "Migration note for the release cut" section.deepseek-v4.1-flash | 𝕏

Finishes quest/m2/ietf-rewind-check.md and deletes it.
Since moq-dev/moq#8918, a moq-transport frame's time is its Group ID in the track's TIMESCALE, so its timestamps cannot rewind. The exemption that let them skip the rewind check goes, in Rust and JS together.
TrackState::neighborsskips the check only on anuncheckedcopy (lite-05/06), no longer on one with moq-transport provenance.#unchecked()reads onlyInfo.unchecked. JS also refused a time anchor on a moq-transport copy, since its starts "may rewind"; that refusal goes too, matching Rust, which never had it.draft-lcurley-moq-ietf's Timestamps section drops the rule that a receiver MUST NOT check a nonzero-Version track (and the sentence contrasting it with an unnegotiated hop). The extension lives only on the unpublishedmoq-lite-07-wip, so no changelog entry.doc/concept/moq-lite.mdsays moq-transport timestamps are checked like any other. The draft'sTRACKS_ADDparagraph now gives the reason that survives (a TRACKS_ADD lacks the Version, which says whether Group Streams carry Object Records), and thelite/publisher.test.tscomment callsuncheckedthe lite-05/06 flag only.main: moq-dev/moq#9027 already removed both time-anchor refusals inSubscriber's settle and rewrote theAnchorTSDoc, so that conflict takesmain's side. "a moq-transport copy resolves a time anchor" is ported to group-by-start keys (0/1000/2000) and a bare-key time anchor (1500n).Choices for review
Info.ietfstays, in both languages. The quest's Goal says it goes, but its Plan says to confirm each reader is only the exemption first, and they are not: Rust'sInfo::ietf()picks lite-07 object records and the source draft inlite/publisher.rs,lite/track.rs,ietf/publisher.rsandgroup::Consumer::ietf; JS'sInfo.ietfsetsfillEdgefor a FETCH-only moq-transport copy. So only the exemption is removed, and the lite-07ietfhop provenance stays.js/net/src/track.ts) so it merges cleanly with the concurrentquest/m2/time-addressing/js.mdwork in the same file.Public API and wire impact
Public API: none (
Info.ietfandInfo.uncheckedare@internal/pub(crate)).Behavior: a timed track with moq-transport provenance now refuses a group that starts before a lower one (
TimestampMismatchon that group; the track stays open), and JS accepts a time anchor on such a copy.Wire: none for published versions. The lite-07-wip
ietfextension's receiver rule changes from "MUST NOT check" to the default moq-lite check.Migration note for the release cut: the Unreleased "A timed track refuses a group that rewinds (#5803)" note in both
js/net/MIGRATE.mdandrs/moq-net/MIGRATE.mdsays a track received over moq-transport stays unchecked. It should read: "A group received over moq-transport is checked too, since its time is its Group ID. A track received over moq-lite-05/06 keeps its timestamps unchecked, since they may be an untimed origin's send times. A relay serves it on over every version, ..." (just checkforbids a PR from editingMIGRATE.md).Tests
ietf::subscribera_received_rewind_is_acceptedbecomesa_received_rewind_is_refused: the group is delivered, its read fails withTimestampMismatch, the track stays open.lite::ietfietf_timestamps_keep_their_provenance_across_a_hopno longer expects a rewind across a negotiated hop; it now also asserts a rewind on the provenance-bearing origin is refused.model::resumeHeldused a moq-transport copy only to get an unchecked one; it now setsunchecked: true(a lite-05/06 copy), which keeps the test's intent.track.test.ts: "IETF provenance survives normalization and keeps the rewind check" (was "skips"); "an untimed track refuses a time anchor" drops its moq-transport case, and "a moq-transport copy resolves a time anchor" covers what replaces it: a{ time }start on anietfcopy reads the group containing it and those after (fails on the old refusal, passes on the head).Validation
cargo nextest run -p moq-net --all-features: 4113 passed, 37 skipped (three tests failed before their update, as expected, confirming the exemption was live).bun testinjs/net: 3377 pass.just check: passes (Rust, JS, drafts, docs).just test interop --all: 209 PASS, 1 SKIP (the known Chromium route-key skip). Exercises Rust and JS moq-transport and lite-07 across relays with the check now on.quest check --base origin/main: 482 documents ok.After merging
main(d6f22a8 and on):cargo nextest run -p moq-net --all-features3854 passed;bun test src/track.test.ts252 pass;just checkpasses except four moq-relay integration tests that hit 60 s timeouts at load average ~25 (quic_sim::two_relays, threerefusals::). They passed in the priorjust checkrun on the same Rust code and pass alone (cargo nextest run -p moq-relay -E 'test(two_relays) | test(refusals::)': 14 passed in 2 s).Caused by
None: this retires an exemption that was right until moq-dev/moq#8918 changed where moq-transport times come from.
Follow-ups
None.
Models
Size: S
Implementer: Claude Code, claude-opus-5-5, unknown effort (arm A)
Implementer: Claude Code, claude-opus-5-5, unknown effort (iterate, XS, arm A)
Implementer: Claude Code, claude-opus-5-5, unknown effort (iterate, XS, arm A)
(Written by Opus 5.5)
Review
Rewritten by
quest reviewafter each round.Head: 2e01436
Base: origin/main at d8ec56d
Risk: medium,
the base has no RISK.md to name what is always high or customer-facing, so every change is mediumClassified: medium by default
Reviewer: Codex, gpt-6.1-sol, medium effort
Session:
01a1eb1d-99c8-74a2-947f-411c05d35421, round 3Score: 9/10,
The incremental changes correctly adapt the regression test to bare time keys and restore release-owned migration notes, with no actionable findings; builds and tests remain unrun as instructed.Rounds: 3; first round 0; findings 0: fixed 0, refuted 0, open 0
Findings: none