babysit-prs (worker tier): no deterministic signal that a bot re-review round has settled (codex signals 'clean' via a 👍 reaction on the PR issue, not a review record)

[kyle-sexton]

Observed while babysitting PR #528 (worker tier, 3 cascading codex re-review rounds).

Friction

After a worker pushes a fix and the fix commit goes green in CI, it must decide: merge now, or wait because another re-review may surface a new finding. There is no deterministic "re-review settled" signal, and the two review backends behave differently:

• When codex HAS a finding it posts a PR review record (pulls/<pr>/reviews, state COMMENTED) plus an inline comment — detectable by polling reviews for one whose commit_id == head.
• When codex is CLEAN it posts no review record; per its own About text it "will react with 👍" — i.e. a reaction on the PR issue (issues/<pr>/reactions, content=+1, user=chatgpt-codex-connector[bot]). A worker polling pulls/<pr>/reviews for a head-SHA review therefore waits indefinitely for a signal that never arrives, and cannot distinguish "still reviewing" from "reviewed, no findings".

There is also no per-commit correlation on the clean path: the 👍 is on the issue, not tied to a commit SHA — the worker must infer "clean for head" from the reaction created_at being after the head push time.

Impact

Worker-tier autopilot (auto-fix + gate-merge) needs to know when a re-review round is DONE to avoid either (a) merging into a pending finding, or (b) hanging on a review that will never post. Currently this is left to ad-hoc timing heuristics per worker.

Suggested direction

A shared helper (or an addition to the merge-gate / readiness surface) that reports a bot re-review as "settled" for the current head when EITHER a head-SHA review record exists OR a bot 👍 reaction on the PR issue post-dates the head-commit push time — normalizing the finding vs. clean signals across review backends (codex, claude[bot], cursor). Encode the backend-specific quirk (review record vs. issue reaction) once, so workers stop reimplementing timing heuristics.

Repro context

• Backend: chatgpt-codex-connector[bot] on melodic-software/claude-code-plugins
• PR fix(toolchain): resolve a present remote for default-branch detection instead of assuming origin #528 head d726a8b: clean round-4 review surfaced ONLY as issues/528/reactions +1 at 2026-07-19T18:35:59Z (push was 18:33:15Z); pulls/528/reviews count stayed at 3.

Category: improvement / operational.

[kyle-sexton]

This was generated by AI during triage. · Category: operational (improvement) · Priority: priority: high (source-control north-star; worker autopilot can merge into a pending finding OR hang on a review that never posts — precondition for trustworthy #476 autopilot). Routed status: ready.

Plan

Two review backends signal differently and a worker has no deterministic "re-review settled for head" signal:

• Finding present → codex posts a PR review record (pulls/<pr>/reviews, COMMENTED, commit_id==head) — detectable.
• Clean → codex posts no review record; it reacts 👍 on the PR issue (issues/<pr>/reactions, +1, chatgpt-codex-connector[bot]). A worker polling pulls/<pr>/reviews for a head-SHA review waits forever and can't tell "still reviewing" from "clean".

Fix: a shared helper (folded into the merge-gate / readiness surface) that reports a bot re-review settled for the current head when EITHER a head-SHA review record exists OR a bot 👍 reaction on the PR issue post-dates the head-commit push time. Encode the backend-specific quirk (review-record vs issue-reaction; the clean path's created_at > head-push-time correlation, since the 👍 carries no SHA) ONCE, normalized across codex / claude[bot] / cursor, so workers stop reimplementing timing heuristics.

Acceptance

• Given a green head with a clean codex round signaled only by a post-push issue 👍 (repro: PR fix(toolchain): resolve a present remote for default-branch detection instead of assuming origin #528 head d726a8b, reaction 18:35:59Z > push 18:33:15Z, reviews count stayed 3), the helper reports settled=clean for head — not an indefinite wait.
• Given a head-SHA review record, it reports settled=has-findings.
• No 👍 and no head-SHA review yet → reports not-settled (still reviewing), distinct from clean.
• The per-backend quirk lives in one helper; workers consume the verdict.

Composes with #534 (shared classifier + gate owns the merge-ready boolean — this settled-signal is part of that readiness surface) and #476 (autopilot must know a round is done before merging). Adjacent gate-signal bugs: #499 (approval parse), #465 (finding count).

Decision defaulted: one shared "re-review settled for head" helper on the readiness surface — settled when a head-SHA review record exists OR a bot 👍 issue-reaction post-dates the head push — normalizing finding-vs-clean across backends, rather than per-worker timing heuristics — veto before merge

[kyle-sexton]

Frontier disposition (drain session)

Ruling: Close as not planned.

Bot re-review settlement signal is tracked under the babysit-prs convention umbrella; re-open with a concrete acceptance brief if needed.