--- name: review-ci-logs description: Diagnose a CI failure or noisy CI output from the logs, find the first cause the repository owns, and fix it without hiding diagnostics. Use when a CI run fails or flakes, when asked to make CI green or quieter, or when asked to add a retry or silence output in a workflow. --- # Review CI logs Read every applicable `AGENTS.md` first, then the CI docs. Repository-local policy overrides these defaults. 1. Pick the runs. For a run you were given, inspect only that run. Otherwise choose a few failed and a few successful runs inside a bounded window. 2. Download logs to a temporary directory and read the failed steps there. Do not stream whole archives into the conversation. Remove the files afterwards. 3. Classify each finding: a real error, misleading output, a cascade from an earlier error, a diagnostic worth keeping, or volume only. Find the first cause the repository owns. Later errors are usually the cascade. 4. Compare the failing job with what was run locally before the push: the same commit, project, target, fixtures, and configuration. Say which of these holds: the local check was skipped, it passed in a different scope or environment, the failure can only happen in CI, or it is unknown. The same command name does not prove the same scope. Do not call an outside infrastructure failure a gap in local checks. 5. Make one bounded fix at the cause. Keep non-zero exits, the primary error, artifacts, summaries, and diagnostic output. Do not hide stderr, quiet output globally, or add a retry that masks the cause. A retry is right only for the same idempotent operation failing for a transient outside reason. 6. Add a focused regression check, run the workflow checker, and compare the output before and after where that means something. Report what was left alone. Do not file issues unless asked. 7. When a local check was skipped or did not match CI, say so in the pull request description. The `pr-description` skill owns what that note holds.