rstacruz/babysit-pr

babysit-pr

Monitor a PR's CI status in a loop. Delegate log analysis to a subagent to keep the main context lean. Fix branch-related failures and push. Stop when CI passes or a blocker requires user input.

Inputs

Accept any of:

• No argument — auto-detect PR from current branch
• PR number
• PR URL

Setup

Find the directory containing this SKILL.md at runtime — that is SKILL_DIR. The polling script lives at $SKILL_DIR/scripts/poll-pr.sh.

Run it as:

bash "$SKILL_DIR/scripts/poll-pr.sh" [<pr-number-or-url>]

Requires gh (≥ 2.32) and jq.

Main loop

LOOP:
  snapshot = poll-pr.sh [<pr>]

  status=merged  → "PR #<n> merged." — STOP
  status=closed  → "PR #<n> closed without merging." — STOP
  status=passing → "CI green. PR #<n> is ready to merge." — STOP

  status=pending →
    unchanged_count += 1
    interval = min(60 × 2^(floor(unchanged_count / 3)), 300) seconds
    wait interval, loop

  status=failing →
    unchanged_count = 0
    Spawn ONE @general-alpha subagent with ALL failing_run_ids:
      """
      Analyse CI failures for PR #{{PR_NUM}}.
      For each run ID in {{FAILING_RUN_IDS}}, run:
        gh run view <id> --log-failed
      Classify the overall failure set as branch-related OR flaky/infra.
      - Branch-related: compile errors, test failures, lint violations,
        type errors in files touched by this branch.
      - Flaky/infra: network timeouts, runner provisioning failures,
        GitHub Actions outages, rate limits.
      If branch-related: describe exactly what to fix and in which file(s).
      If flaky: confirm no changed code is implicated.
      """

    Read classification:
      branch-related →
        Check worktree: if unrelated uncommitted changes exist → STOP, ask user
        Fix the code
        git add, commit, and push
        If push rejected (non-fast-forward, protected branch, etc.) → STOP, report
        unchanged_count = 0, loop

      flaky/infra →
        For each failing run_id:
          retry_count[run_id] += 1
          if retry_count[run_id] > 3 → STOP, "Flaky retry budget exhausted: <workflow_name>"
        gh run rerun <run-id> --failed  (for each failing run)
        loop

Git safety

• Verify no unrelated uncommitted changes before editing (git status --short)
• Work only on the PR head branch
• Never force-push; if git push is rejected, stop and report the reason
• Flaky retry budget: 3 per run ID, tracked independently

Output cadence

• Emit a progress update only on status changes (pending → failing → pushing, etc.)
• Emit a heartbeat every 5 unchanged pending polls ("still pending, next check in Xs…")
• On stop: output a final summary — PR number, final status, commits pushed, retries used

Stop conditions

1. merged or closed
2. passing — all checks green
3. Branch-related failure where fix attempt fails or push is rejected
4. Flaky retry budget (3) exhausted for any run
5. Unrelated uncommitted changes in worktree
6. Auth or permission failure (gh auth errors, push access denied)