Repository navigation
fix(net): the ingest fuzzer counts a refused GROUP_END as a failed stream - #9216
Conversation
Quest-Claim: 8d548765a77cf078
…ream The harness dropped the result of settling the group, so a GROUP_END before a frame it held read as a clean end over an aborted, empty group, tripping the stream-size bound. The session already fails that stream with MalformedTrack; the harness now does too. The crash input joins the regressions. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
✅ No new issues found.
Approach: sound. The harness now treats a group.end(fin) the group refuses as the run's error, as the session does, instead of fixing the ingest path that was already correct.
Reviewed changes
Reviewed the full diff: the harness fix, the regression input and the quest file deletion.
rs/moq-net/src/fuzz/ingest.rs:runnow keeps the result ofgroup.end(fin), so a refused end is reported asErrand no longer recorded as a clean end. TheErrarm still aborts the group and returns the original error. The types line up with the laterend.map_err(|err| err.to_string()).rs/moq-net/fuzz/regressions/ingest/lite07-group-end-before-frame: adds the crash input, whichjust checkreplays viafuzz::tests::regressions. The diff shows this file as binary, so I did not inspect its bytes.quest/m0/nightly/fuzz-ingest-bound.md: removes the finished quest plan.
I did not run the tests. The PR reports fuzz:: passing 6/6 and a 300 s just rs fuzz ingest run with no crash.
claude-sonnet-5-5 | 𝕏
|
No actionable comments were generated in the recent review. WalkthroughThe ingest fuzzer now uses the group settlement result as its result when ingest succeeds. If ingest fails, it still aborts the group and returns the ingest error. The document describing the crash investigation and planned regression-corpus update was deleted. Merge Risk:
|
|
Merge pick (maintainer away, granted the coordinator merges it confidently recommends): ✅ /quest-merge at d20a736. (Written by Claude Opus 5.5 with high) |

Problem
Fuzz (ingest)failed in Nightly run 44562804083:The input is a lite-07 timed stream: one 93-byte frame stamped about 15e9 ms, then
GROUP_ENDwith End 0. Ingest reads both and returnsOk(Some(0)). Thenrecv::Group::end(Some(0))refuses an end before a frame the group holds and aborts the group withMalformedTrack, which is what the session does with that stream (lite/subscriber.rs,Ok(end) => group.end(end)). Ingest is correct. The bug is in the harness:rundiscarded the result ofgroup.end(fin), recorded the end as clean, and then read zero frames from the aborted group. The size bound then fired on the mismatch.Changes
rs/moq-net/src/fuzz/ingest.rs: a refused end now becomes the run's error, so the harness sees the stream fail as the session does.fuzz/regressions/ingest/aslite07-group-end-before-frame.quest/m0/nightly/fuzz-ingest-bound.md.Impact
Public API: none. Wire: none. Test harness only.
Decisions
Choices I made without the maintainer (all undoable in review):
GROUP_ENDbefore a held frame.recv::Groupalready enforces that check for the session and for this harness, so a second copy in ingest would duplicate it.crash-<hash>name, followingdraft22-untimed-extension-block.Validation
fuzz::tests::regressionsbefore the fix (ingest.rs:191, the same message as Nightly) and passes after it.cargo test -p moq-net --features fuzz --lib fuzz::passes 6/6.just rs fuzz ingest -- -max_total_time=300 -timeout=10on nightly 1.102 with cargo-fuzz 0.13.2, corpus seeded with the committed regressions including this input. Done 477,595 runs in 301 s, no crash.just checkpasses.Caused by
group.finish()with the falliblegroup.end(fin)but kept discarding its result (found by check: the Nightly fuzzer)Follow-ups
None.
Models
Size: S
Implementer: Claude, claude-opus-5-5, medium effort
(Written by Claude Opus 5.5 with medium)
🤖 Generated with Claude Code
Review
Rewritten by
quest reviewafter each round.Head: d20a736
Base: origin/main at 1a34abe
Risk: low,
Only the ingest fuzz harness, a regression input, and a quest plan change, so no consumer or peer behavior can break, and a revert fully undoes the change.Classified: low by Codex, gpt-6.1-sol, low effort, session
01a1f6f8-f5c7-7821-ab95-c591a7e2f616Reviewer: Codex, gpt-6.1-sol, medium effort
Session:
01a1f6f9-0be5-7b32-8533-41835251cb0f, round 1Score: 10/10,
The focused change fixes false clean-end classification by propagating group-finalization errors and adds the crash input to regression replay; source review found no issues, though tests were not run.Rounds: 1; first round 0; findings 0: fixed 0, refuted 0, open 0
Findings: none