Babysit CI

Mode

• watch: inspect and report. No edits, pushes, or reruns.
• fix ("fix CI", "get checks green"): scoped fixes and pushes. Never merge.
• Ambiguous request: watch.

Loop

1. Find the PR: gh pr view <pr> --json number,url,headRefName,headRefOid. No PR given: pass the branch from git branch --show-current.
2. Read checks with gh pr checks <pr> --json name,bucket,state,workflow,link. Judge the PR by this output alone: Actions runs are only part of it, since apps and status contexts report here too.
3. Diagnose failures before waiting on anything. Each failing check gets one root error: the earliest log line that names a cause; errors after it often follow from it.
   • Actions check: take the run id from its link; gh run view <run-id> --log-failed.
   • External check: open its link and find the step or service that failed.
4. Pending checks: run gh pr checks <pr> --watch --fail-fast as a long-running command if this runtime supports one and keep diagnosing meanwhile; otherwise run it in the foreground.
5. Watch mode: once the watcher exits, reread the checks and the head SHA, report, and stop. Fix mode continues below.

Fix

Sort each failure before touching code:

• Not from this PR's diff: see whether the default branch already passes that check. If it does, git fetch origin and git merge origin/<default> bring the fix in; this PR carries no patch for it.
• Flake: one rerun (gh run rerun <run-id> --failed), and cite the evidence: it passed on rerun with no change, or it is a known flake (say where it is recorded).
• Caused by this PR: if the failing command is cheap to run locally, reproduce the failure first and rerun it after the change. One cause per push. Stage explicit paths. Never pass --no-verify.

After any push, restart at step 2. A push can trigger workflows the old list lacked or retire ones it showed, so check state read before the push is stale.

Stop, report, and ask before continuing

• 3 fix pushes done.
• The same failure survives a fix: switch to $fable-mode, rethink the cause, and name the fix tier.
• The fix needs secrets, infra, permissions, or a user decision.

Before calling it green

Re-read headRefOid; if it moved, run the loop on the new head. A missing, skipped, or unavailable expected check does not pass; list it.

Each fix push leaves a head without the pre-merge review. State that in the report. Rerunning the cross-vendor $claude-review uses the user's Claude subscription and needs their OK.

A CI quirk that cost a retry goes to pitfalls via $recall.

Report

• Each iteration: check → first root error line → link → fix applied.
• Final: status and PR URL, or blocked plus the next action.