Repository navigation
feat(congestion)!: BBR ships through the selector, candidates compile behind moq-congestion's lab feature - #8850
Conversation
Quest-Claim: 8dc91da8afd60b0c
…moq-congestion-lab GCC and SCReAMv2 move to the workspace-only moq-congestion-lab crate, BBR3's delay bound gets a hidden setter, and moq-ffi's MoqCongestionControl becomes an object with constructors so a new algorithm is additive in every binding. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ort feature doc(hidden) is banned, so the pacer, a packet's GCC probe cluster, and BBR3's delay bound sit behind moq-congestion's documented lab feature, as test support does elsewhere. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…w the move Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
ℹ️ No blocking issues. One small doc gap is noted below.
Reviewed changes
I reviewed all four commits: the candidates move into a workspace-only crate, the lab test-support feature, and the conversion of the FFI congestion selector from an enum to an object in every binding.
moq-congestion-labcrate: GCC and SCReAMv2 move over unchanged except for imports and theprobe_clusteraccessors. The crate ispublish = false, and moq-quic and moq-tokio take it only as a path dev-dependency. It has its own large-MTU window test.labfeature onmoq-congestion: this feature gatespub mod pacer, thePacketprobe-cluster field and its accessors, and thebbr3::Config::delay_boundsetter. The license dropsBSD-3-Clause, and the unusedtracingdependency is removed.MoqCongestionControlas a uniffi object: it hasloss(),delay()andnone()constructors and exportsDebug/Eq/Hash.MoqQuicConfig.congestion_controlbecomesOption<Arc<_>>. Dart is regenerated. Go gets builder functions and a genericoptional[T]. Kotlin, Python, Swift, OBS C++ and the tests switch to the constructors.- Selector docs:
moq_tokio::quic::CongestionControlnow derivesHash. The docs in the relay config, the C API and the bindings saydelayis BBR v3 and the default. - Nightly and quests: the
rust-testgroup adds-p moq-congestion-lab, so the GCCbenchmark_per_packetstill runs and still matches its filter. The completed quest is removed, and the vendor-sync andquic-realtimequests point at the new paths.
I checked locally: cargo check --all-targets passes with no warnings for moq-congestion without lab, moq-congestion with --features lab, and moq-congestion-lab. A grep finds the old enum spellings (CongestionControl.DELAY, kLoss, CongestionControlLoss, and so on) only in historical MIGRATE.md entries.
ℹ️ doc/setup/repository.md doesn't list moq-congestion-lab
AGENTS.md sends contributors and agents to doc/setup/repository.md to find where a feature lives. Its "Congestion controllers and controller tests" row still lists only rs/moq-congestion. After this PR, someone looking for GCC or SCReAMv2 there will look in the wrong crate.
Technical details
# repository.md omits the new lab crate
## Affected sites
- doc/setup/repository.md:21 — the "Congestion controllers and controller tests" row links only `rs/moq-congestion`
## Required outcome
- The row also points to `rs/moq-congestion-lab` as the home of the candidate controllers under test (GCC, SCReAMv2)claude-opus-5-5 | 𝕏
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 25 minutes. View limit details
No actionable comments were generated in the recent review. WalkthroughThe changes add Merge Risk:
|
|
Maintainer decision (2026-11-15, /quest-complete): no new crate. GCC and SCReAMv2 stay in moq-congestion as modules compiled only behind a feature (✅ the existing documented (Written by Opus 5.5) |
…the lab feature The maintainer rejected a separate crate: the candidates are modules of moq-congestion again, compiled only with its `lab` test-support feature, which also gates BBR3's delay-bound setter. The pacer and a packet's probe cluster return to crate-private, the vendored WebRTC files keep their LICENSE and PATENTS beside them, and the license expression covers BSD-3-Clause again. moq-quic's and moq-tokio's comparisons enable `lab` as a dev-dependency, and the nightly GCC benchmark keeps its original series. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Applied the maintainer decision in 2f2095a: This also answers Pullfrog's note about (Written by Claude Opus 5.5 with medium) |
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
Since the last review, the branch merged main and dropped the workspace-only moq-congestion-lab crate. GCC and SCReAMv2 are modules of moq-congestion again, compiled only with its lab feature.
- Candidates behind
lab:gccandscreamare#[cfg(feature = "lab")]modules again.pacerstays crate-private because BBR3 and Cubic also use it.Packet'sprobe_clusterfield ispub(crate)and compiled only withlab, so the accessor pair from the previous round is no longer needed. - Optional
tracing:lab = ["dep:tracing"], and only GCC logs. Hakari addedtracingtoworkspace-hackto match. - License and vendored notices: these are unchanged from
main. The license is(MIT OR Apache-2.0) AND BSD-3-Clause, andsrc/gcc/{LICENSE,PATENTS,AUTHORS}ship with the crate. - Dev-dependency wiring: moq-quic and moq-tokio enable
labthrough theirmoq-congestiondev-dependency. Their test code matchesmainagain. - Large-MTU test: the large-MTU window test uses a
Vecof factories, and the GCC and SCReAM entries are gated bycfg. - Nightly:
rust-testaddsmoq-congestion/labto--features, sobenchmark_per_packetkeeps its original series name. - README and quests: the moq-congestion README describes the feature. The
quic-realtimequest records the maintainer's decision, and the vendor-sync quests no longer change.
I ran cargo check --all-targets -p moq-congestion locally, with and without --features lab. Both pass with no warnings. No references to moq-congestion-lab or moq_congestion_lab remain. The previous review's note about doc/setup/repository.md no longer applies now that the lab crate is gone.
claude-opus-5-5 | 𝕏

Problem
moq-congestion 0.1.0 has never been published, yet its public API carried three candidates no session can select: GCC (
gcc, about 16k vendored WebRTC lines under BSD), SCReAMv2 (scream), and BBR3's delay bound (bbr3::Config::delay_bound). The maintainer's pick on moq-dev/moq#8764: ship BBR through one selector, keep the candidates off the published API until tested, and make new algorithms additive. moq-ffi'sMoqCongestionControlwas a closed enum, so a new algorithm would break every Swift, Kotlin, and Dart caller that switches over it.Completes
quest/m2/congestion-candidates.md.Approach
src/gcc/) and SCReAMv2 (src/scream.rs) stay modules of moq-congestion, compiled only with its documentedlabtest-support feature, the waysimandfuzzwork elsewhere.labalso gates BBR3's delay-bound setter. Without it, neither module, nor a packet's GCC probe-cluster field, nortracing(now optional, only GCC logs) is compiled. The pacer stays crate-private.labthrough their moq-congestion dev-dependency, whichcargo publishdrops, so their code is unchanged from main. The large-MTU window test covers GCC and SCReAM again whenlabis on, and the nightlyrust-testgroup namesmoq-congestion/labexplicitly, so the GCC per-packet benchmark keeps its original series.LICENSE,PATENTS, andAUTHORSbeside them and still ship in the published package, so moq-congestion's license stays(MIT OR Apache-2.0) AND BSD-3-Clause, as on main.moq_tokio::quic::CongestionControl(#[non_exhaustive], relay/CLIcongestion_control). Its docs now say plainly thatdelayis BBR v3 and the default, and that a new algorithm is added as a variant named after it.MoqCongestionControlbecomes an immutable uniffi object withloss(),delay(), andnone()constructors, following theMoqAudioCodecprecedent, with Eq/Hash/Debug exported. A new algorithm is a new constructor, which breaks no binding. Wrappers, tests, and docs are updated in Swift, Kotlin, Dart (bindings regenerated), Python, Go, C++ (OBS plugin), and moq-ffi.quic-realtime's stale "addRealtime" naming item is aligned with the newer decision to name each winner after its algorithm, and it records the feature decision below.Impact
moq-congestion(unpublished):gcc,scream, andbbr3::Config::delay_boundnow require the newlabfeature (test support, meant for dev-dependencies).tracingbecomes an optional dependency thatlabenables. License unchanged.moq-tokio:quic::CongestionControlalso derivesHash(additive). Docs only otherwise.moq-ffi(breaking):MoqCongestionControlis an object instead of an enum.MoqQuicConfig.congestion_controlisOption<Arc<MoqCongestionControl>>. Migration:.delay/.loss/.none->CongestionControl.delay()/.loss()/.none()CongestionControl.DELAY->CongestionControl.delay()CongestionControl.delay->CongestionControl.delay()CongestionControl.DELAY->CongestionControl.delay()CongestionControlLoss/Delay/Noneconstants are gone; usemoq.LossCongestionControl(),moq.DelayCongestionControl(),moq.NoneCongestionControl(), assigned directly (noptr(...))moq::CongestionControl::kDelay->moq::CongestionControl::delay(); compare witheq()loss/delay/nonestring)delay,loss,none).Decisions
labfeature, one feature for all three candidates (rejected: the workspace-onlymoq-congestion-labcrate). This supersedes the quest's recorded rejection of "ship them behind an unstable feature", which paraphrased the quest: first pass of the pre-cut feature review, and its follow-ups #8764 pick. The quest file is deleted by this PR, so the record lives inquest/m4/quic-realtime.md, the bake-off that owns the candidates.lab: it already names "candidates under test" in this PR, and a second feature (say,candidatesbesidelab) would split one test-support switch in two.labfeature, not#[doc(hidden)](banned bysh/text.sh), the same rule moq-dev/moq#8392 set forsim/fuzz.MoqCongestionControlas an object (theMoqAudioCodecshape) rather than wrapper-only types in Swift/Kotlin/Dart, so every binding grows the same way and the API names match across languages.Validation
just check(changed scope): every module passes exceptcheck quest, which fails only because this host's dev shell carries a stale quest (62d6b5e) older than the flake's pin; the pinned v0.15.0 (nix run github:moq-dev/quest/3122d76 -- check --base origin/main --branch quest/m2/congestion-candidates) reports485 documents ok. Itstest rsran 13818 tests, all passing, including 251gcc::and 9scream::tests in moq-congestion (enabled through moq-quic's dev-dependency).lab: clippy--all-targets -D warningsandcargo doc -D warningspass with no dead-code warnings; nextest runs 125 tests (no candidates compiled).--features lab: clippy and doc pass; nextest runs 385 tests, including the large-MTU window check over GCC and SCReAM.just rs test --release -p moq-congestion --features lab -E 'test(benchmark_per_packet)' --run-ignored onlyrunsmoq-congestion gcc::bench::benchmark_per_packet, the series name main already records.just rs test -p moq-tokio --test integration -E 'test(scream::near_zero)' --run-ignored onlypasses.cargo package --list -p moq-congestion: shipssrc/gcc/withLICENSE,PATENTS,AUTHORS, andsrc/scream.rs, under(MIT OR Apache-2.0) AND BSD-3-Clause.just py test(128 passed),just go check,just kt check,just dart check, and the OBS settings test on the new constructors; moq-ffi'sclient_connects_with_every_knob_setand rung tests. Swift isn't run on Linux; CI's macOS lane covers it.Alternatives
moq-congestion-labcrate (this PR's first approach): rejected by the maintainer; it split one crate's tests, benchmark history, and license notice across two for no published-surface gain over a feature.Follow-ups
None.
Models
Size: M
Implementer: Claude Code, claude-opus-5-5, unknown effort (arm A)
Implementer: Claude Code, claude-opus-5-5, medium effort (iterate, M, arm A)
🤖 Generated with Claude Code
(Written by Claude Opus 5.5 with medium)
Review
Rewritten by
quest reviewafter each round.Head: 2f2095a
Base: origin/main at f3f4cc5
Risk: high,
The branch changes public controller and binding APIs that downstream applications may compile against, so consumers may need migration even though reverting restores the previous API.Classified: medium by default
Reviewer: Codex, gpt-6.1-sol, medium effort
Session:
01a1dfcc-cfd7-75e1-8793-d48ad864e19e, round 2Score: 9/10,
The incremental refactor preserves controller behavior and updates feature wiring and callers consistently, with no actionable defects found in source review; builds and tests were not run as instructed.Raised the risk to high:
The branch changes public controller and binding APIs that downstream applications may compile against, so consumers may need migration even though reverting restores the previous API.Rounds: 2; first round 0; findings 0: fixed 0, refuted 0, open 0
Findings: none