ci_skill

Overview

This skill monitors CI and workflow status, diagnoses failures by fetching only the failing logs, and applies targeted fixes plus systematic prevention. It focuses on pragmatic checkpoints around pushes, reviews, and merges to avoid noisy continuous checks. The skill automates reruns for flaky failures and prescribes changes that prevent recurrence.

How this skill works

It checks PR CI status using gh pr checks and lists recent runs for the current branch to identify failures. For each failed run it fetches failed-step logs (gh run view --log-failed), extracts the failing step, error message, root cause, and affected files, then recommends and applies an immediate fix. After the acute fix, it proposes and commits prevention measures (linter rule, pre-commit hook, CI check, stricter types, test, or docs) and pushes the change so CI re-runs.

When to use it

• After push settles (avoid burst commits)
• Before requesting review or declaring work done
• Before merging a PR into a mainline branch
• When CI shows failing checks or red status
• When tests intermittently fail and need diagnosis

Best practices

• Check CI at clear checkpoints rather than continuously to avoid noise
• Differentiate flaky tests from real failures; rerun flaky ones first
• Fetch only failed logs to speed diagnosis and reduce noise
• Always pair the acute fix with a prevention change and document both in the commit message
• Use PAL or an expert chat for non-trivial failures before broad refactors

Example use cases

• A unit test fails in CI due to missing null check — add the check, add a unit test, enable strictNullChecks, commit and push
• CI shows network timeout tests intermittently — rerun the failed run, if flaky add retry logic or stabilize test environment
• Build fails with a TypeScript error after refactor — fix types, update tsconfig or add a type assertion, and add a linter rule if needed
• A lint step is newly failing — add or adjust the ESLint rule and a pre-commit hook to catch it locally
• Release blocked by a flaky integration test — rerun failed runs and add a targeted integration test and CI gating to prevent regressions