PR Babysitter

• IS: keeping one open PR moving: conflicts, CI across GitHub Actions/Buildkite/Vercel/Fly.io, inbound review comments, and merge readiness, then following the merge through the repo's post-merge watch, as a background monitor or as one-shot fixes.
• IS NOT: opening or editing the PR (pr-creator), reviewing or fixing the diff itself (tidy), or npm release PRs (autoship watches its own release CI; never babysit a release or Version Packages PR it drives).

Mode Selection

Standing rules, every mode:

• Monitoring or fixing code does not by itself authorize posting replies. Post, resolve threads, or request reviews only when the user authorized that communication; otherwise prepare replies and report them.

• Resolve scripts/fetch-comments.sh relative to this installed SKILL.md. ${CLAUDE_SKILL_DIR} below is a Claude Code adapter, not a portable environment variable.

• No setup questions. Auto-detect the PR, the CI platforms, and the defaults (poll every 2 minutes, auto-resolve noise, no auto-merge), then start. Overrides arrive inline: "poll every 5 minutes", "enable auto-merge".

• Skip closed PRs. Skip drafts (isDraft) unless asked. A merged PR goes on to the repo's post-merge watch: wait for the watch/<env> commit statuses on the merge SHA, stop at the first failure or error, and report the result (Post-merge Watch in references/monitoring-setup.md). If the base it merged onto (the merge commit's first parent, or for a rebase merge the commit before the PR's rebased commits) has no watch/* status, report "no post-merge watch". Either way, once the merge is live in production, smoke-test the change there read-only (Production smoke test in the same reference), report that too, and stop.

• Comment triage runs autonomously; the plan file is an audit trail, not an approval gate.

• Speak only on transitions. A quiet poll says nothing.

Reference Files

Monitor Loop

Phase 1 runs once in the foreground and starts the watch. Every event or tick then runs phases 2-5, diffs against the state file, and speaks only when something changed.

Copy this checklist to track progress:

PR babysit progress:
- [ ] Phase 1: Initialize (detect PR, pick watch mechanism, snapshot state)
- [ ] Phase 2: Conflict check
- [ ] Phase 3: CI check (diagnose, fix, gate, push)
- [ ] Phase 4: Comment check (triage new comments)
- [ ] Phase 5: Readiness check (report transitions, write state file)
- [ ] After merge: post-merge watch until watch/* passes, fails, or is absent, then a read-only production smoke test

Phase 1: Initialize

1. gh pr view [N] --json number,url,title,state,isDraft,headRefName,baseRefName,headRefOid,mergeable,mergeStateStatus,reviewDecision. No PR for the branch: say so and stop.
2. gh repo view --json owner,name for the calls that need owner/repo.
3. Detect CI platforms from gh pr checks --json name,link (pattern table in references/ci-platforms.md).
4. Pick the watch mechanism: the first rung of the watch ladder in references/monitoring-setup.md that applies (harness PR subscription, Monitor tool, cron, none). With no rung, do not claim monitor mode: run the matching one-shot mode or say this runtime cannot keep polling.
5. Snapshot state to .claude/pr-babysitter/babysit-pr-{N}.md: mechanism and ID, head SHA, mergeability, check states, open and awaiting-reply thread counts, review decision. This folder is never staged.
6. Confirm in five lines: PR, watch mechanism and ID, detected CI, current state, defaults in effect.

Phase 2: Conflict Check

gh pr view --json mergeable,mergeStateStatus. DIRTY resolves; BEHIND updates; UNKNOWN means GitHub is still computing, recheck next tick; anything else moves on.

git fetch origin {base} && git rebase origin/{base}
git push --force-with-lease --force-if-includes

• Clean rebase: push, notify.
• Conflicts only in lockfiles, generated files, or changelogs: regenerate per the reference, continue the rebase, push.
• Conflicts in source logic, migrations, or API contracts: git rebase --abort, then notify with the files and what each side changed. Human intent decides those.

Bare --force is never used. A refused lease means someone else pushed: abort and notify rather than overwrite their commits. More than one author on the branch means a rebase rewrites their commits: merge origin/{base} instead.

Phase 3: CI Check

1. gh pr checks --json name,state,bucket,link,workflow. bucket is pass, fail, pending, skipping, or cancel; link is the details URL.
2. Anything pending: wait. Diagnosing a half-finished run fixes the wrong thing.
3. Every fail: fetch logs with the platform's commands in references/ci-platforms.md (GitHub Actions, Buildkite auth chain, Vercel, Fly.io).
4. Classify per the reference: flaky (re-run once), stale dependency (reinstall and rebuild before touching source), code error (fix), knip (delete dead code or configure), infrastructure (notify; not fixable from code).
5. Fix, run the verification gate, commit, push. Flag regressions against the previous state (was passing, now failing).

One-shot loop ("fix CI"): after each push, gh pr checks --watch --fail-fast (exit 0 green, 1 a check failed, 8 still pending). Stop and summarize when checks are green, the failure is infrastructure, or the same check fails twice with the same error after a fix. Two identical failures is the signal to stop pushing, not to try a third variant.

Phase 4: Comment Check

1. Count open threads and threads awaiting my reply (newest comment not mine, in any resolution state, minus a reviewer who resolved their own last comment).
2. Compare both counts and the newest updated_at across review and issue comments against the state file. An edited-in-place bot comment and a reply on a resolved thread both have to register.
3. Any increase: run the Comment Triage Workflow, then notify "N new review comments on PR #{N}", counting only comments that are not noise. A poll whose new comments are all noise (such as the sticky <!-- pr-evidence --> comment, edited on every push) is a quiet poll.

Phase 5: Readiness Check

1. Ready means all of: mergeable == MERGEABLE, every required check pass, reviewDecision == APPROVED from a review whose commit_id is the head SHA, zero open blocking threads, zero threads awaiting my reply, every merge gate satisfied.
2. A merge-gate comment reading "Human review required" is a blocker to report with the criteria that forced it, not a finding to fix.
3. Ready: notify "PR #{N} is ready to merge." Merge is a one-way door: gh pr merge --auto with the repo's merge method, and only when the user opted in.
4. Not ready: name the blockers ("Waiting on: 2 checks pending", "Awaiting your answer: 2 questions from @reviewer", "Approval is stale: reviewed abc1234, head def5678").
5. Write the state file for the next tick.

Comment Triage Workflow

Inline from Phase 4 or one-shot. Autonomous: no approval gate, the plan file is the audit trail.

Fetch

Run ${CLAUDE_SKILL_DIR}/scripts/fetch-comments.sh {N}. It pages every thread and every thread's comments, recovers anchors, buckets threads, and computes owedReply against your own login. A non-zero exit prints one sentence on stderr saying why; fix that cause and re-run rather than hand-writing queries.

Check reviewers[] before classifying: every login that spoke must appear in the output with findings, a verdict, or an explicit "no content". A reviewer with reviews but zero comments is a fetch that lost something. anchor.source == "needs-translation" means finish the anchor ladder against the working tree before judging that finding.

Early exit only when open threads, awaiting-reply threads, actionable reviews, and actionable issue comments are all zero.

Classify

• Every inline comment from every author is read. An author absent from the bot table is unknown, not noise; noise needs a positive marker match.
• Classify per comment, not per thread: a human reply inside a bot's thread carries full human weight.
• Author type from content first, then login. github-actions[bot] is shared by reviewers and noise alike.
• Severity from the source's own markers; unknown sources default to Major. Severity orders the queue; it never decides whether a comment is read.
• Human intent: fix request, question, nitpick, or acknowledgement. A question gets an answer, not a code change. Human comments are never auto-ignored: fix unless the reviewer marked it optional.
• Merge-gate verdicts are Phase 5 inputs: record, never fix, never reply, never resolve.
• Deduplicate bots only (same path within 3 lines, keep the highest severity). A multi-location finding is one item.
• Every ignore carries one of the legal reasons from the plan template. "Author unrecognized" and "thread already resolved" are not among them.

Fix

1. Write the plan (references/fix-plan-template.md) to .claude/pr-babysitter/pr-{N}-review-plan.md, print the counts, proceed.
2. Ignored threads: one-line reply, then resolve.
3. Questions: post the answer, leave the thread open. The reviewer resolves it.
4. Resolved threads with an unanswered human reply: reply in place, do not unresolve, note it in the report.
5. Fixes, one commit per logical group. Run the verification gate before each commit and stage only that group's files.
6. Reply, then resolve, each fixed thread.
7. Re-run the script and report: open threads, threads still awaiting my reply, questions answered but not yet acknowledged, and current CI status. The re-run is the evidence; "addressed everything" is not.

Stopping

"Stop babysitting" or "cancel the monitor": stop the mechanism recorded in the state file per references/monitoring-setup.md (Stopping, Session Lifecycle), then report polls or events handled, fixes applied, conflicts resolved, comments triaged, and current state.

Gotchas

• Filtering threads on isResolved == false: GitHub collapses resolved threads, so a human reply posted after the resolve is the comment most likely to go unread.
• comments(first: 20): thread comments arrive oldest first, so a truncated page hides the newest comment, the one that decides whether you owe a reply. The script pages every thread to the end.
• viewerDidAuthor returned false on the viewer's own comments. Compare author.login to gh api user --jq .login.
• A null line means outdated or multi-line, not PR-level. Only a null path is PR-level. Recover the anchor before deciding anything.
• Triaging a bot's review body: the body is a count. Codex, Devin, Copilot, and Bugbot all put findings inline. Four empty-body human reviews are one review pass with its content in threads, not four reviewers with nothing to say.
• Bots that edit one comment in place (auto-approval assessments, DangerJS) keep the same id, so a state diff on ids sees nothing. Compare updated_at.
• Counting a review whose commit_id is not the head SHA as an approval: branch protection with "dismiss stale approvals" drops it on the next push, and the PR reads ready until then.
• git add -A after a fix commits hook artifacts (a root schema.gql) into the PR. Sweep git status --porcelain and stage paths.
• Resolving a thread without replying first: the reviewer sees a silent resolve and unresolves it.
• A subscription-only watch never sees conflicts: GitHub emits no webhook when the base branch advances into one.
• Cron when Monitor is available wakes the agent on every quiet tick and burns tokens for no signal. Polling under 2 minutes does the same to the GitHub rate limit.

Related Skills

• pr-creator: opens or edits the PR; babysitting starts after it exists
• planning: writes plans a fresh session executes. The fix plan this skill writes is an audit trail for one PR, not a planning deliverable
• tidy: reviews and fixes the diff itself; this skill applies GitHub review comments. Run it on monitor-authored fixes beyond a trivial patch
• autoship: npm release pipelines; it watches its own release CI, so never babysit a release PR it drives