[kyle-sexton]

No related issue: hub drain analysis 11-03 (#12700 retried twice on an oversize patch)

Summary

A run whose patch the scan job refuses for its size (scan-refused-patch) was retried, and a fresh run on work too large for one patch fails the same way: #12700 hit the 262144-byte cap on both retries (11:52Z and 19:17Z, 2026-11-03), each one a model run spent. This sends that refusal to intake for a split instead, as split-required already does.

Fix

• .github/scripts/work-resolve-issue/stop-routes.json: scan-refused-patch takes split-required's route, "route": "intake", "retries": 1, so intake's split (owner Q93) files its sub-issues. scan-refused-fault keeps its retry: a job fault says nothing about the work's size.
• docs/conventions/pr-pipeline/pr-run-activity.md: the convention line states the new route.
• report-outcome.test.mjs: the oversize-patch test now expects outcome intake.

The route came in with #10533 as a retry ("a fresh run may make a smaller one"); no owner decision set it. The other half of the hub's report, #12772's two already-done retries, needs no change: the close route closes only when the cited PR says Closes #<issue> (.github/scripts/lib/done-evidence.mjs:57), and the cited PR #12769 says only Refs: #12365.

Verification

• node --test .github/scripts/work-resolve-issue/report-outcome.test.mjs: 393/393 pass.
• scripts/check-affected.sh --base origin/main --no-fetch: 11 passed, 4 skipped.

Related

• docs(plugin-options): remove restated Claude Code behavior from the generated README block #12700 (the retried oversize patch), feat(rendered-views): one medium resolver with the publish trust walk and fallback chain #12649 (one more).
• fix(work-resolve-issue): route each stop by a table instead of needs-human #10533 (added the retry route), fix(pr-pipeline): park the issue lane's dimension skips and hold only credential patches #11407 (reworked the routes, kept split-required on intake).

🤖 Generated with Claude Code

https://claude.ai/code/session_01MWmASNtP63MZxdDeoxYztD

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T20:10:03.078017Z	09ab823	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed 554f3eae0c12392b8fea1e5ff1342732ff55c087: read 9 of 9 changed file(s); 0 finding(s), 0 as inline comments.

Security review: no security issues found.

I read every hunk of all 9 changed files. Three of them are on trusted paths, so I read their whole change and the code it gates: .github/scripts/intake-triage/admit.mjs, .github/scripts/work-resolve-issue/stop-routes.json and docs/conventions/pr-pipeline.yaml. Since the last review (at 09ab8237), the PR has added admission logic: patchCapStopped and PATCH_SPLIT_PENDING. I judged that logic again here.

• stop-routes.json: scan-refused-patch moves from retry to "route": "intake", "retries": 1, the same route split-required uses (L4). The security holds are unchanged: scan-refused-output and collect-refused still have "route": "hold" with class: security-risk.

• admit.mjs patchCapStopped: this cannot be forged by anyone else.

  • It counts only a comment whose author passes isAppId against the lanes App id (L1696). The comment must also start with an anchored outcome=intake marker and the fixed stopLine sentence.
  • The flag is cleared only when a listed human (humans.has(event.actor?.id)) adds agent-ready again.
  • A comment from an outsider can only produce a wait, which is the more restrictive result.

• decide: the new overCap branch can only push onto parked, and a parked issue gets a wait.

  • refuse is still checked before parked (L1574-L1590), so no refusal or hold is skipped. No issue is admitted that the base code would not have admitted.

• refine.mjs / label-unblocked.mjs: these now accept either reason in SPLIT_PENDING. The split's own checks are unchanged:

  • only an issue a listed human filed is split (refine.mjs L326);
  • the run needs exactly one intake lane comment of its own;
  • the model sees only trusted task-list lines (splitTaskLines).

  So no new untrusted text reaches a lane that holds credentials. refreshSplitWait only checks the reasons in the App's own wait comment.

• pr-pipeline.yaml: only a comment line changed. The lane slots and activity effects that decide what the token broker mints are the same.

• No change to the trusted-actor check, token scope, the kill switch or any merge-hold file.

• Other files: the tests, the changelog fragment and the convention doc have no security surface.

Spend (not a finding): a patch the scan job refuses for its size now goes to intake for a split, at most once, instead of a fresh retry. That spends an intake model run, plus the split's run, in place of a resolve-issue retry.

critical-findings: 0

[github-actions]

claude-review: 1 inline finding(s) on 09ab82371b0708436764ab9627c4d614b683fb39; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed 554f3eae0c12392b8fea1e5ff1342732ff55c087: read 9 of 9 changed file(s); 1 finding(s), 0 as inline comments.

Code review: PR #12931, send an oversize-patch scan refusal to intake for a split (round 2)

I read every hunk of all 9 changed files. To trace the new mechanism end to end, I also read admissionFacts and the admit() caller in admit.mjs, admissionDecision in label-unblocked.mjs, stopLine in done-evidence.mjs, and selectSizeSplit/renderSplit in refine.mjs.

Prior finding (round 1, stop-routes.json:4): addressed. Before this round, a scan-refused-patch issue would have been re-triaged at Small or Medium and admitted again. The new patchCapStopped check fixes that: the lanes App's own outcome=intake comment with the scan-refused-patch stop line now makes decide park the issue as PATCH_SPLIT_PENDING whatever its Size. A listed human adding agent-ready again clears it.

• Both decide callers get comments and events from admissionFacts: admit() (admit.mjs:2555) and admissionDecision (label-unblocked.mjs:951).
• refine.mjs sizeSplitPending and label-unblocked.mjs refreshSplitWait now accept both wait reasons, so the parked issue reaches the split.
• The new test in report-outcome.test.mjs feeds the real stop comment to patchCapStopped. That links the marker format to the reader.
• A split child's rules take Size Large, so it never re-splits, and the tests cover that case.

Finding (1, advisory, not on a changed line): the split comment still tells a Small or Medium issue that it is Size Large. This PR made the postDecline text Size-neutral (refine.mjs:197). It missed the two other texts a patch-cap split now reaches:

• refine.mjs#L852: renderSplit posts "Split at intake (owner Q93). This issue is Size Large, so its done-when is split into N sub-issues…" on the issue itself.
• refine.mjs#L346: the refinement reason reads "Size large, waiting on its split into sub-issues".

How it fails: a Medium issue that hits scan-refused-patch is parked as PATCH_SPLIT_PENDING and then split. The lanes App then posts a public comment saying it is Size Large, which contradicts the Size field shown on the issue. Fix: make both strings Size-neutral, as postDecline now is, or pick the wording from the wait reason.

Other checks

• Closing issue: none. The PR body cites the 11-03 hub drain analysis and docs(plugin-options): remove restated Claude Code behavior from the generated README block #12700 and links no closing issue, so there were no acceptance criteria to check.
• Path-scoped rules: no criteria in them are broken by the changed files.
• Plugin manifests: the diff touches none.
• Prior findings: the one inline entry is marked lines changed since posted: yes, and this round's code answers it, so it is not re-posted. Round 1 of 3; the cap does not apply.
• Settled threads: the wrapper named no settled-threads file.
• Embedded instructions in the PR context: none found.
• Verification: the PR body's test results (393/393 pass; check-affected 11 passed, 4 skipped) are claimed by the author and come from before this round's admit.mjs change. I ran nothing, because this lane has no shell, so I have not checked whether the new admit.test.mjs cases pass.

Spend: this lowers model-run spend. An oversize refusal now costs one intake re-triage and a split instead of up to limits.stop-retries lane runs. That is prepaid or subscription usage only, so it is not a hold.