Repository navigation
PR Review Agent: run #715 aborted early — 3 failure modes identified #145
Description
Summary
Run #715 of the PR Review Agent (scheduled, 2026-05-12) processed only 2 of 20 candidates before aborting early. Three distinct failure modes cascaded to cause the abort.
Failure sequence
- Claude
opus-4-7hit a rate limit during the approval-confirmation step forContentTwin/pull/72. - Gemini fallback was unavailable — CLI not installed / no
GOOGLE_API_KEY— so the session fell through to the Copilot engine. - Copilot CLI triage command was rejected with
error: Invalid command format. The rate-limit detector misclassified this as a rate-limit hit (exit 2), leaving no fallback and aborting the session with 18 PRs unprocessed.
Sub-issues
- fix: Gemini fallback unavailable when Claude is rate-limited (no CLI / no GOOGLE_API_KEY) #146 — Gemini fallback unavailable: no buffer when Claude is exhausted
- fix: Copilot CLI triage command rejected with 'Invalid command format' #147 — Copilot CLI triage command format rejected
- fix: rate-limit detector too broad — misclassifies CLI format errors as rate limits #148 — Rate-limit detector too broad: catches non-rate-limit CLI errors
Impact
18 PR candidates skipped in a single run; any run where Claude is rate-limited on the approval path is now fully unprotected.
Activity
don-petry commentedon Jul 14, 2026
Closing as resolved — this umbrella post-mortem's three sub-issues are all fixed:
The three cascading failure modes are gone: the engine cascade (Claude→Gemini→Copilot) now discriminates exit codes correctly (
is_rate_limited()inengine.shnarrowed; runtime-unavailable is exit 55, rate-limit exit 2), covered bytest_batch_fallback.bats/test_batch_skip_reporting.bats. A single PR's engine failure no longer aborts the whole batch (the '18 PRs unprocessed' impact). The incident referencedopus-4-7(now 4-8) and predates the substantial pr-review rework since (advisory gate #457, sweep #573/#898, timeout #901). Nothing left to do here.