detect-flaky-ci

Run the detect-flaky-ci skill, which scans recent failed workflow runs, fetches failed job logs, filters out infrastructure noise, and matches test failure identities against existing tracking issues. Use --window-hours to control how far back the scan looks.

It uses mechanical signals: diff/PR mismatch, sandwich patterns across runs, matrix divergence on the same commit, and measured cross-file slowdowns for setup-hook timeouts. Playwright in-run retries count as immediate confirmation, while vitest failures accumulate observations across runs.

No. It is detection-only and never touches source code, branches, or PRs. Its only outputs are new GitHub issues, comments on existing issues, and label changes on issues it created; fixing is delegated to a separate investigation skill.

The skill uses the gh CLI with the api and run view subcommands, or the mcp__github__get_job_logs MCP tool in cloud environments where the egress proxy blocks blob storage redirects. The fetch method is probed once per run, not per log.

A fixed run count silently stops covering older failures on high-volume days, dropping them from the scan with no warning. A time window sized to at least twice the cron interval guarantees coverage even if one scheduled run is skipped entirely.

The mechanical suspicion signals are not exhaustive, so subtle flakiness still requires passive observation accumulation or human review. Playwright log attribution can also be ambiguous, in which case the skill falls back to a coarser job-level identity rather than fabricating a precise one.