fix-ci
Diagnose a failing CI run from the log — reproduce locally, minimal fix, verify.
SKILL.mdskills/fix-ci/SKILL.md
GitHubWhen to use
The user shows a failing CI log or says the pipeline/build broke.
How
- Read the log bottom-up — the first error matters; everything after is usually cascade. Quote the exact failing line back.
- Classify: test failure / lint-format / type error / dependency or install / flaky-infra (timeout, network, runner died).
- Reproduce locally with the same command CI ran (from the workflow file —
read_file .github/workflows/*.yml); same Python/node version if pinned. - Fix the cause, not the symptom: a failing assertion means the code or the test is wrong — decide which from the requirement, then change that one.
- Re-run the exact failing command locally until green; mention any check you could NOT run locally.
- Flaky-infra failures: say so explicitly, suggest retry + a guard (timeout bump, retry-on-503), don't "fix" unrelated code.
Avoid
- Deleting/skipping a test to make CI green — that's a last resort the user must explicitly approve.
- Pinning random versions until it passes, with no explanation.
- Fixing three unrelated things in one go.
Done well
The failing line is named, the cause is explained in one sentence, the fix is minimal, and the previously-failing command passes locally.