Ci

Monitor CI Pipeline

Workflow

why-not-mechanizable: skill workflow guidance; each step requires understanding the surrounding context (repo, task shape, prior state).

1. Detect VCS platform: .gitlab-ci.yml -> glab, .github/ -> gh (review-time: see section note)
2. Start a Monitor with the matching script: (review-time: see section note)
   • GitHub: bash ~/.claude/skills/ci/scripts/gh-ci-monitor.sh (review-time: see section note)
   • GitLab: bash ~/.claude/skills/ci/scripts/glab-ci-monitor.sh (review-time: see section note)
   • Use persistent: false, timeout_ms: 3600000 (1 hour ceiling - CI pipelines can be long) (review-time: see section note)
   • Description: "CI pipeline on <branch-name>" (review-time: see section note)
3. React to Monitor notifications: (review-time: see section note)
   • no-runs|<branch>: no CI runs found for this branch - inform the user and stop (review-time: see section note)
   • error|persistent-failure: the monitor script hit 5 consecutive errors - report and stop (review-time: see section note)
   • Status change (e.g., in_progress|null → completed|success): acknowledge briefly (review-time: see section note)
   • Pipeline passes (completed|success): (review-time: see section note)
     • Run ~/.claude/scripts/notify.sh "CI passed - <branch-name>" (review-time: see section note)
     • Report success (review-time: see section note)
   • Pipeline awaiting manual action (completed|manual, GitLab only): (review-time: see section note)
     • All automatic jobs completed; the pipeline is paused on a manual gate and will not progress without user action (review-time: see section note)
     • Run ~/.claude/scripts/notify.sh "CI awaiting manual action - <branch-name>" (review-time: see section note)
     • Report status and stop watching - do NOT trigger the manual job automatically (review-time: see section note)
   • Pipeline fails (completed|failure or any non-success conclusion): (review-time: see section note) a. Fetch the job log:
     • GitLab: glab ci trace <job-id> (review-time: see section note)
     • GitHub: gh run view <run-id> --log-failed (review-time: see section note) b. Do NOT pipe output through head, tail, grep, or any other command - run the commands directly c. Analyze the root cause - identify the specific failure (test, lint, type check, build, coverage, etc.) d. Classify the failure as transient or real before proposing any change - transient = infra outage, rate limit, queued/timed-out runner, auth/network flake, registry or dependency propagation delay; real = a test/lint/type/build/coverage failure caused by the code under change. State the classification (review-time: see section note) e. If transient: re-run the failed job (gh run rerun <run-id> --failed / glab ci retry <job-id>) and keep monitoring - do NOT edit code for a transient failure. Escalate to the user only if it recurs after a re-run (review-time: see section note) f. Run ~/.claude/scripts/notify.sh "CI failed - <failure-summary>" g. If real, propose the fix to the user - explain what failed and what you'd change. Do NOT push automatically (review-time: see section note) h. Wait for user approval before implementing the fix (review-time: see section note) i. After approval: fix, commit with a descriptive message, push (review-time: see section note)

How it works

The monitor scripts poll CI status every 30 seconds but only emit a line when the status changes. This means:

• Zero token cost while the pipeline is running and status hasn't changed (review-time: see section note)
• Claude reacts within ~30s of a status change (vs up to 2 min with /loop) (review-time: see section note)
• No CronDelete cleanup needed - the script exits on terminal state, ending the Monitor (review-time: see section note)
• If gh/glab fails 5 times in a row (auth expired, network down), the script exits with an error notification (review-time: see section note)

Important: Non-interactive commands only

These commands require a TTY and will NOT work - never use them:

• glab ci view (interactive TUI) (review-time: see section note)
• gh run watch (interactive watcher) (review-time: see section note)
• Any command with --web flag (opens browser) (review-time: see section note)

Safe commands to use:

• glab ci status, glab ci list, glab ci trace <job-id> (review-time: see section note)
• gh pr checks, gh run list, gh run view <run-id> --log-failed (review-time: see section note)

Guardrails

• After 3 consecutive failures on the same issue, stop and escalate - something structural is wrong (review-time: see section note)
• Never weaken tests, skip linting, or lower coverage thresholds to make CI pass (review-time: see section note)
• Never use --no-verify or skip hooks (review-time: see section note)
• If a failure looks unrelated to your changes (flaky test, infra issue), flag it to the user rather than trying to fix it (review-time: see section note)