babysit-pr（保姆公关）

Babysit PR

Autonomously monitor a PR (or an entire Graphite stack), fix CI failures, respond to review comments, and push verified fixes. Runs as a long-running session — loops internally with intelligent sleep intervals until all PRs are ready or the user stops the session.

Invocation

/babysit-pr

Session Loop

The skill runs in a continuous loop:

1. Detect if the current branch is part of a Graphite stack
2. Check PR status (CI + reviews) across all PRs in the stack
3. Fix issues starting from the lowest affected branch
4. If all PRs are ready → print success and exit
5. Otherwise → sleep for an intelligent interval, then repeat

digraph babysit {
    node [shape=box];

    "Start session" [shape=doublecircle];
    "Detect Graphite stack\n(gt ls)";
    "Collect PRs for\nall stack branches";
    "Any PRs found?" [shape=diamond];
    "Exit: no PRs found" [shape=doubleoctagon];
    "For each PR\n(bottom of stack first):\ncheck CI + reviews";
    "Any issues found\nacross stack?" [shape=diamond];
    "For each branch (bottom-up):\nfix, verify, commit,\nrestack";
    "gt submit --stack\n(single push)";
    "All PRs: checks green +\ncomments addressed?" [shape=diamond];
    "Print: Stack is ready" [shape=doubleoctagon];
    "Print remaining issues\n+ sleep interval";
    "Sleep with\nintelligent interval";

    "Start session" -> "Detect Graphite stack\n(gt ls)";
    "Detect Graphite stack\n(gt ls)" -> "Collect PRs for\nall stack branches";
    "Collect PRs for\nall stack branches" -> "Any PRs found?";
    "Any PRs found?" -> "Exit: no PRs found" [label="no"];
    "Any PRs found?" -> "For each PR\n(bottom of stack first):\ncheck CI + reviews" [label="yes"];
    "For each PR\n(bottom of stack first):\ncheck CI + reviews" -> "Any issues found\nacross stack?";
    "Any issues found\nacross stack?" -> "All PRs: checks green +\ncomments addressed?" [label="no"];
    "Any issues found\nacross stack?" -> "For each branch (bottom-up):\nfix, verify, commit,\nrestack" [label="yes"];
    "For each branch (bottom-up):\nfix, verify, commit,\nrestack" -> "gt submit --stack\n(single push)";
    "gt submit --stack\n(single push)" -> "All PRs: checks green +\ncomments addressed?";
    "All PRs: checks green +\ncomments addressed?" -> "Print: Stack is ready" [label="yes"];
    "All PRs: checks green +\ncomments addressed?" -> "Print remaining issues\n+ sleep interval" [label="no"];
    "Print remaining issues\n+ sleep interval" -> "Sleep with\nintelligent interval";
    "Sleep with\nintelligent interval" -> "Detect Graphite stack\n(gt ls)" [label="loop"];
}

1. Detect Graphite Stack

Determine if the current branch is part of a stack:

gt ls

This shows the full stack structure. Parse the output to get the ordered list of branches from bottom (closest to trunk) to top.

If the branch is standalone (not part of a stack), treat it as a single-branch stack — the rest of the flow works the same way.

2. Collect PRs Across the Stack

For each branch in the stack (bottom to top), find its PR:

gh pr view <branch> --json number,url,headRefOid

Build a list of (branch, pr_number, headRefOid) tuples. Skip branches that don't have PRs yet.

If no PRs exist for any branch, print an error and exit.

3. Check CI and Reviews Across All PRs

Use sub-agents for parallel status collection. Dispatch one sub-agent per PR to fetch CI status, review threads, and general comments concurrently. Each sub-agent runs the checks from steps 3a–3c for its assigned PR and returns a structured summary of issues. This keeps the main context clean and avoids sequential API calls across the stack.

Similarly, when investigating failing checks, dispatch a sub-agent per failing run to fetch logs (gh run view <run-id> --log-failed) and analyze the root cause. CI logs are verbose — isolating them in sub-agents prevents context pollution.

Collect all issues into a list tagged by branch, so you know which branch to fix first. Then fix bottom-up per step 4.

3a. Check Required CI

gh pr checks <pr_number> --json name,state,bucket,link,required

• Only act on checks for the HEAD commit — verify via headRefOid.
• Fix failures immediately — don't wait for all checks to finish. If some checks have already failed while others are still pending, fix the failed ones right away. The push will trigger a fresh CI run for everything anyway.
• For each failing check (required and optional):

1. Get logs: gh run view <run-id> --log-failed 2. Investigate the root cause 3. Fix the code

• If checks are pending or in_progress, note how long they've been running (use gh run view <run-id> --json createdAt) to inform sleep interval.
• Optional check failures are best-effort — fix them if possible, but they do not block the exit condition.

3b. Inline Review Threads

gh api graphql -f query='
  query($owner: String!, $repo: String!, $pr: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $pr) {
        reviewThreads(first: 100) {
          nodes {
            id
            isResolved
            comments(first: 10) {
              nodes { body author { login } path line }
            }
          }
        }
      }
    }
  }
'

3c. General Review Comments

These are top-level review comments — the body text of a review submission, not tied to any specific file or line. They often contain important high-level feedback, architectural concerns, or summary requests.

gh api repos/{owner}/{repo}/pulls/{pr}/reviews --jq '.[] | select(.body != "" and .body != null) | {id, body, state, user: .user.login}'

Also fetch standalone PR comments (conversation tab, not part of a review):

gh api repos/{owner}/{repo}/issues/{pr}/comments --jq '.[] | {id, body, user: .user.login}'

Processing Rules

Skip any thread/comment that:

• Is resolved (for threads)
• Has already been addressed by the babysitter (last reply is from the bot / starts with 🤖)

Classify and act on each remaining thread or comment.

All replies MUST start with 🤖 to identify as AI-assisted.

For general review comments, reply using the appropriate API:

• Review comments: gh api repos/{owner}/{repo}/pulls/{pr}/reviews/{review_id}/comments or reply to the review
• Issue comments: gh api repos/{owner}/{repo}/issues/{pr}/comments -f body="..."

4. Fix Issues Bottom-Up

When issues are found across multiple PRs in the stack, always fix the lowest branch first. A fix in a lower branch may resolve issues in higher branches after restacking.

Important: only push once per loop iteration. Apply all fixes locally, then push the entire stack at the end.

For each branch with issues (bottom to top):

1. Checkout the branch: gt checkout <branch>
2. Fix all CI failures and review comments for that branch (per step 3 rules)
3. Run targeted local verification — lint and typecheck the changed files, and run only the tests that exercise the changed code (not the full suite). Use the project's CLAUDE.md/AGENTS.md to determine how to run scoped tests (e.g., passing specific test files, directories, or filter patterns). If verification fails, fix the issue and re-verify until it passes.
4. Commit locally: gt modify --commit
5. Restack so higher branches pick up the changes: gt restack
6. Continue to the next branch with issues

After all branches have been fixed:

1. Push the entire stack once: gt submit --stack

5. Sleep Interval Logic

After each iteration, choose a sleep duration based on current state:

When babysitting a stack, base the interval on the most urgent situation across all PRs (i.e., use the shortest applicable sleep).

Print the chosen interval and reason before sleeping: "Sleeping {N} minutes — {reason}"

Use shell sleep for the wait (e.g., sleep 300 for 5 minutes).

6. Exit Condition

All PRs in the stack are ready when:

• All required CI checks are passing on HEAD for every PR
• All review threads across all PRs are either resolved OR have a babysitter reply as the last comment
• All general review comments across all PRs have been addressed (babysitter reply exists)

Optional check failures do not block this exit condition. If optional checks are still failing when the exit condition is met, note them in the output but still declare the stack ready.

Print: "All required checks passing and all review comments addressed across the stack. PRs are ready."

Common Mistakes

• Acting on stale CI results — Always verify checks are for the HEAD commit before investigating failures. If checks are pending, wait.
• Replying to the same comment twice — Check if the last reply in a thread is already from the babysitter before responding.
• Pushing without local verification — Always run the project's verification commands before gt modify --commit.
• Resolving threads you shouldn't — Only resolve threads where you fixed a real issue. Leave scope-change and non-issue threads open for the reviewer.
• Missing general review comments — Don't only check inline threads. Always also fetch top-level review bodies and issue comments — reviewers often put their most important feedback there.
• Sleeping too long after pushing — After pushing a fix, use a short interval so you catch CI results quickly.
• Fixing the wrong branch in a stack — Always fix the lowest branch that has the issue. Fixing higher up can cause merge conflicts or get overwritten by a restack.
• Pushing multiple times per loop — Fix all branches locally first (gt modify --commit + gt restack for each), then push once with gt submit --stack. Multiple pushes waste CI cycles and create race conditions.
• Forgetting to restack after fixing — After modifying a branch in the middle of a stack, always gt restack before moving to the next branch so higher branches pick up the changes.
• Assuming higher branches are unaffected — After fixing and restacking, re-check all PRs from scratch. The restack may introduce new CI failures in higher branches.
• Fetching PR status and logs sequentially in the main context — Use sub-agents to check each PR's CI/reviews in parallel and to fetch verbose CI logs. This speeds up each loop iteration and keeps the main context window clean for decision-making and code fixes.