Babysit Pr

Babysit the current PR until it is merge-ready. Alternate between two signals and fix whatever is broken after each round:

1. CI checks — all required checks must pass (gh pr checks).
2. Bot review comments — no unanswered bot comments (npx agent-reviews --watch --bots-only).

Terminate only when both signals are clean at the same time.

Prerequisites

The current branch must have an associated PR. The agent-reviews CLI authenticates via gh or env vars; on cloud environments verify git config --global --get user.email is set to a real address so CI can attribute commits.

Substitute the user's package manager runner (pnpm dlx, yarn dlx, bunx) for npx where appropriate.

Phase 1: Initial sweep

Step 0: Ensure the PR is mergeable

Before touching CI or comments, confirm the PR is not conflicting with its base branch. A stale base can cause CI to pass on obsolete code or block merge even after every check is green.

Run gh pr view --json mergeable,mergeStateStatus,baseRefName.

• mergeable: MERGEABLE → go to Step 1.
• mergeable: UNKNOWN → GitHub has not computed mergeability yet. Wait ~5s and re-check (retry up to 3 times before escalating).
• mergeable: CONFLICTING or mergeStateStatus: DIRTY → rebase:

1. git fetch origin
2. git rebase origin/
3. If conflicts arise during the rebase:

• Trivial / mechanical conflicts (lockfile regeneration, import re-ordering, adjacent edits to separate symbols): resolve and git rebase --continue.
• Non-trivial conflicts (overlapping logic changes, schema/API changes, anything requiring product judgement): run git rebase --abort and ask the user.

1. After a clean rebase, push with git push --force-with-lease — never plain --force.
2. Re-run gh pr view --json mergeable to confirm the PR is now clean before proceeding.

Step 1: Resolve existing bot comments

If there are any unanswered bot comments, resolve them by invoking the resolve-agent-reviews skill (or following its workflow: fetch with --bots-only --unanswered --expanded, evaluate each finding, fix true positives, commit, push, reply with --resolve).

Skip to Step 2 if there are no unanswered comments.

Step 2: Check CI status

Run gh pr checks to read the current status.

• All passing → Phase 2.
• Any pending / in-progress → Step 3.
• Any failed → Step 4.

Step 3: Wait for in-flight checks

Run gh pr checks --watch to block until every check finishes. When it returns, re-run Step 2 to re-evaluate.

Step 4: Fix failing checks

For each failed check:

1. Identify the failing run: gh pr checks --json name,link,state,workflow (or gh pr view --json statusCheckRollup).
2. Fetch the failing logs: gh run view --log-failed. For very long logs, grep for the first error.
3. Analyze the root cause. Read the relevant code — do not guess from the log alone.
4. Fix the code minimally. Do not refactor unrelated code.
5. Reproduce the check locally where feasible (lint, type-check, unit tests).
6. Commit with a message describing the fix (e.g., fix(ci): resolve type error in FooBar) and push.
7. Return to Step 2.

If the failure is genuinely unclear, or the fix would require architectural changes, stop and ask the user.

Phase 2: Watch loop

All checks are green and all known bot comments are resolved. Now wait to see whether anything new appears.

Step 5: Watch for new bot comments

Run npx agent-reviews --watch --bots-only as a background task and wait for it to exit (default 10 min; override with --timeout).

• EXITING WITH NEW COMMENTS → handle them as in Step 1, then go back to Step 2 (new commits may have retriggered CI).
• WATCH COMPLETE → Step 6.

Step 6: Final CI sanity check

Run gh pr checks once more. Intermediate commits may have kicked off a new run. Also re-check gh pr view --json mergeable — the base may have moved during the watch window.

• All green and mergeable → Summary report. Done.
• Check failing or pending → back to Step 2.
• Conflicting with base → back to Step 0.

Summary report

## Babysit PR Summary
- Iterations: N
- Bot comments addressed: X (fixed Y, won't-fix Z, skipped W)
- CI failures fixed: M
- Final status: All checks green, no unanswered bot comments

Stopping rules

Bail out and ask the user when:

• The same check fails 3 times in a row after your fixes — your diagnosis is probably wrong.
• The same bot keeps re-raising a comment on the same lines after you replied — it is not accepting your fix or dismissal.
• A failure requires judgement outside the scope of the PR (architectural change, unrelated broken infra, flaky test needing a retry rather than a fix).
• The user has not authorized pushing to this branch, or the branch is protected.

Do not use destructive git operations to work around a failure. git push --force-with-lease is acceptable after a legitimate rebase in Step 0; plain git push --force, git reset --hard, and history rewrites intended to hide a failing check are not.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gzaripov
• Source: gzaripov/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.