Include available CI failures in review-round coder follow-ups

[wwind123]

Summary

Include already-failed CI checks in the coder follow-up without waiting for unfinished CI, and prevent the pending-CI filter from silently discarding real reviewer findings.

Why

In PR #757, the coder addressed reviewer findings and pushed another commit while an existing GitHub failure in test_completion_recovery.py remained unaddressed. Both CI runs failed on the same test. The pre-coder path already fetched a post-review check snapshot, but used it only for infrastructure-stall guidance; actionable CI findings were normally added at the approval gate.

Investigation also found that Sol's raw round-1 review contained one blocker and its round-3 review contained two blockers, but the tool published those reviews as approved with the findings removed. _is_pending_ci_only_review() accepted any substring matching a pending check name or generic CI keyword. With a check named test, genuine code findings requesting regression tests were misclassified as CI-wait-only objections. The checked-in regression payloads preserve those real summaries and findings; no provider calls are needed to replay them.

Behavior

• Reuse the post-review, reviewed-head CI snapshot at the coder handoff. Add observed failures, check links, and head identity to the unresolved-item ledger alongside reviewer findings.
• Tell the coder to inspect the failure logs, fix the failures, and run local regressions without waiting for queued or running CI.
• Do not turn missing checks, API unavailability, or recognized infrastructure stalls into code defects. Mixed failed/pending boards contribute only the failures.
• Preserve existing merge-conflict query suppression, managed intermediate-check filtering, and final approval/CI gates.
• Reuse an existing CI item instead of duplicating it; subsequent reviewers disposition the item through the normal ledger.
• Apply this behavior whether or not auto-merge is enabled. No new flag, CI dispatch, watcher, or LLM call is introduced.
• Replace pending-CI substring matching with conservative whole-statement matching. Check names may appear only in the statement's subject, not as arbitrary keyword matches across prose.
• Preserve all findings when any finding or substantive summary is mixed or ambiguous. Existing same-PR follow-up and unresolved-prior-item guards remain in place. An unrecognized phrasing stays blocking instead of being silently approved.
• Verify both early parallel publication and final reconciliation retain the real blockers and pass them to the coder's unresolved-item ledger.

Validation

• Regression matrix covers failures appearing during review, mixed failed/pending checks, pending/missing/unavailable checks, infrastructure stalls, and both auto-merge settings.
• End-to-end regression covers coder handoff followed by re-review and resolution of both the reviewer and CI items.
• Targeted CI-health, PR-orchestration, and parallel-review suite: 282 passed.
• python3 -m pytest -q --tb=short: 2,713 passed in 164.63 seconds, the complete local suite on the expanded fix, including completion recovery and parallel publication tests.
• git diff --check: passed.

-- OpenAI Codex

[wwind123]

Review verdict: Blocking

The CI-failure handoff is otherwise well covered, but the new pending-only classifier fails to recognize a canonical unavailable-status statement unless a check happens to be named status.

Blocking issues

• [src/coding_review_agent_loop/ci_health.py:104] Recognize generic GitHub (PR) check status is unavailable wording independently of check names. On an unavailable board, check_names is commonly empty, so the current subject grammar rejects this unambiguous CI-only statement and _is_pending_ci_only_review leaves the review blocking, dispatching a pointless coder round. The positive test at tests/test_ci_health.py:34 masks the defect by including status as a fictitious check name. Add check status to the generic GitHub subject grammar and cover the unavailable case with no pending or missing check names.

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

Review verdict: Approved

Reviewed 678464a against main. The new post-review CI snapshot correctly excludes infrastructure stalls, pending and missing-required checks, applies managed-CI filtering consistently with the other get_pr_checks call sites, skips entirely while conflicted, dedupes against a same-round GitHub PR checks item, and never waits on CI. The tightened pending-only filter now fails closed on ambiguous prose, which only ever preserves blocking findings. Snapshot field replacement matches the PullRequestChecks dataclass, item numbering stays consistent with the record-derived ledger, and source_round is display-only so refreshing a carried item is safe. README documents both behaviors. Parametrized tests cover failing, mixed, pending, missing, unavailable and stall boards under both auto-merge settings, plus a real-payload regression fixture. GitHub PR checks are passing.

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

PR review round 1 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented and pushed the requested generic GitHub check-status recognition fix.

Addressed items

• item-1: Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 1: [src/coding_review_agent_loop/ci_health.py:104] Recognize generic `GitHub (PR) check status is un...
  • Resolution: Added generic GitHub check-status subjects independent of dynamic check names and regression coverage with an unavailable board supplying no pending or missing check names; the focused CI-health test module covers the changed classifier and its conservative grammar cases.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_ci_health.py -q (agent-loop instrumented; whole-command timeout 120s)

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

Review verdict: Approved

Reviewed d380c52. The round-2 delta adds check status to the generic GitHub subject alternation in is_canonical_pending_only_text, so an unavailable board with no pending or missing-required check names now matches without relying on a fictitious check name; the positive test no longer passes status as a check name and a dedicated case covers check_names=() for both GitHub check status is unavailable. and the GitHub PR variant. Full-match anchoring is unchanged, so the existing negative cases (mixed CI-plus-defect prose, bare test, unescaped build linux) still fail closed and substantive findings stay blocking. Round-1 review of the CI-snapshot handoff logic still holds. GitHub PR checks are pending (test in progress), so no global test-pass claim is made here.

Prior unresolved item dispositions

• [item-1] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 1: [src/coding_review_agent_loop/ci_health.py:104] Recognize generic `GitHub (PR) check status is un... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Approved

The prior pending-CI classification defect is resolved, and the reviewed changes correctly add available non-infrastructure CI failures to coder follow-ups without waiting on pending checks. GitHub's test check remains in progress for the reviewed head.

Prior unresolved item dispositions

• [item-1] Blocking issue from OpenAI Codex: gpt-5.6-sol (medium), round 1: [src/coding_review_agent_loop/ci_health.py:104] Recognize generic `GitHub (PR) check status is un... -> resolved

-- OpenAI Codex: gpt-5.6-sol (medium)

[wwind123]

PR review round 2 reconciliation: settled reviewers: Claude, Codex. Finalization continues after reconciliation.

[wwind123]

Reviewers approved PR #758; watching GitHub checks in the foreground. No coder or reviewer agents will run while checks remain pending.

-- coding-review-agent-loop