reviewer: posted reviews can still fail when completion marker parsing or marker verification misses them

[mrcfps]

Summary

Some reviewer runs successfully post a GitHub review with a Looper marker, but the run is still marked failed because the agent did not emit a parsed __LOOPER_RESULT__ marker or because publish-time marker verification does not find the already-posted review.

This creates false failed reviewer loops and can trigger duplicate follow-up attempts.

Evidence

Posted review but run failed as missing completion marker

Loop 133 (powerformer/looper#125) shows the agent posted a GitHub review successfully:

{"id":4202572321,...,"state":"COMMENTED","commit_id":"fdd4dd124989cae611642ba392bfe462694f01b6"}

The review body includes a Looper marker:

<!-- looper:review id=reviewer:333e9f1e-75cb-466d-9224-13432ab972fe:fdd4dd124989cae611642ba392bfe462694f01b6 head=fdd4dd124989cae611642ba392bfe462694f01b6 outcome=actionable -->

But the loop ended as:

last_error_kind = non_retryable
last_error = Reviewer agent did not report a valid completion marker after publishing review

Publish step failed despite existing reviews

Loops 325 and 333 failed in step=publish with:

Reviewer agent completed but no matching GitHub review marker was found; retrying marker verification before rerunning review

Then their queue items were cancelled with loop_terminated, even though the relevant PR already had reviews/approvals in the review list.

Root cause

The runner has two separate success signals:

1. Agent stdout marker (__LOOPER_RESULT__) parsed by the executor.
2. GitHub review marker verification (<!-- looper:review ... -->).

When (1) is missing, the runner tries to recover by verifying (2), but that verification can miss valid posted reviews due to marker shape, idempotency key mismatch, event type, author, or timing. The fallback then treats the run as failed even though the external side effect already happened.

Relevant code:

• internal/agent/executor.go:499-502 — failed/missing parse marker handling
• internal/reviewer/runner.go:2083-2118 — missing parse marker recovery and final FailureNonRetryable
• internal/reviewer/runner.go:2144+ — publish step verification of pending review markers

Expected behavior

If a Looper-marked GitHub review exists for the expected head, the reviewer run should converge to success or skipped/idempotent completion even when the agent fails to emit __LOOPER_RESULT__.

It should not leave a failed loop after an external review was already posted.

Proposed fix

• Make marker verification tolerant of legacy/current marker id shapes where safe.
• Search recent reviews/comments by expected head and marker prefix before failing missing completion marker cases.
• Treat found existing marker as PendingReview and advance to publish/idempotent completion.
• If verification is inconclusive, prefer retryable_after_resume over non_retryable so the runner can re-check GitHub state before rerunning the full review.
• Add tests where the agent posts a valid marked review but omits __LOOPER_RESULT__.

[mrcfps]

Additional current production evidence from ~/.looper/looper.sqlite on 2026-05-06:

Recent reviewer runs are still failing after the agent appears to have reached the post-review/publish path:

ended_at UTC	local time	PR	run	last_completed_step	current_step	summary
2026-05-06T08:45:42Z	16:45:42	nexu-io/open-design#637	run_802d25a467b536d19e6f8d041197c88e	thread_resolution	review	Reviewer agent did not report a valid completion marker after publishing review
2026-05-06T08:40:57Z	16:40:57	nexu-io/open-design#568	run_7408ab2a0f3e68f5831b2fd78bff059e	thread_resolution	review	same marker failure
2026-05-06T08:40:24Z	16:40:24	nexu-io/open-design#644	run_ff78bfa3b2244cf78a9d12080c4059a8	thread_resolution	review	same marker failure
2026-05-06T08:35:57Z	16:35:57	nexu-io/open-design#642	run_1d2f9b32f424f27e654d416ecb32e562	thread_resolution	review	same marker failure

There are also adjacent publish-step failures on the same day where the agent completed but the runner could not find a matching GitHub review marker:

run_ef9a3f8fd62063c631436f6c73249bf8: Reviewer agent completed but no matching GitHub review marker was found
run_8088d6d7c4461e6ae1192de2895a9659: Reviewer agent completed but no matching GitHub review marker was found; retrying marker verification before rerunning review
run_ca4903a932633e8b6c9d2df64990d96b: same marker lookup failure
run_b8661b66225f6c5e0ea3dccf4ff9d6c8: same marker lookup failure with retry wording

For contrast, successful reviewer publish records on the same host do produce both runs.status=success / last_completed_step=publish and pr.review.posted events:

local time	PR	run	evidence
2026-05-06 14:27:46	nexu-io/open-design#627	run_fec5f09d31af83ec904ed6932805d4d6	looper review submit ... returned { "submitted": true }; DB emitted pr.review.posted
2026-05-06 14:27:10	nexu-io/open-design#617	run_1821430daf9f3e560437f9ffb1b291ca	looper review submit ... returned { "submitted": true }; DB emitted pr.review.posted

A query over today's reviewer runs showed 55 successful publish completions since 2026-05-06T00:00:00Z, so the submit path works generally. The failure mode appears to be convergence/verification after a side effect: if the agent or wrapper has already posted a Looper-marked review, missing __LOOPER_RESULT__ or a missed marker lookup should not leave the loop failed.

Suggested acceptance criteria to add:

• When pr.review.posted exists for the run/head, a missing agent completion marker should converge to success/idempotent completion.
• When GitHub contains a Looper-marked review for the expected head, marker verification should accept it even if the agent parse status is missing.
• If verification is temporarily inconclusive, classify as retryable_after_resume and re-check GitHub state before rerunning a full review.
• Regression coverage should include a completed agent execution with parse_status=missing after a successful review submit, plus a publish-resume path that finds an existing review marker.

[mrcfps]

Looper finished work on this issue.

Pull request: #231

_{Generated by Looper 0.6.0 · runner=worker · agent=codex}