manastalukdar PR Babysitter
Reduce the human time spent herding pull requests. Monitors open PRs, identifies what's blocking each one, and takes the smallest action to unblock — or escalates to a human when judgment is needed.
Usage
/pr-babysitter # single sweep of all open PRs
/pr-babysitter --author <handle> # watch only PRs by a specific author
/pr-babysitter --label <label> # filter by label (e.g. "ready-for-review")
/pr-babysitter --dry-run # report actions without taking them
Run on a loop for continuous herding:
/loop 5m /pr-babysitter
/loop 10m /pr-babysitter --author @me
Behavior
Step 1 — Discover open PRs
# List open PRs with key metadata
gh pr list --state open --json number,title,author,createdAt,reviewDecision,statusCheckRollup,isDraft \
--limit 50 | jq '.[] | select(.isDraft == false)'
Skip draft PRs unless --include-drafts is passed.
Step 2 — Triage each PR
For each open PR, classify its blocking state:
| State | Condition | Action |
|---|---|---|
| CI red | Any required check failing | → Step 3: Diagnose and minimal-fix |
| Needs rebase | Merge conflict or "branch is behind" | → Step 4: Rebase nudge |
| Awaiting review | No reviewer assigned or all reviewers haven't responded | → Step 5: Nudge |
| Changes requested | Review with requested changes | → Step 6: Surface to author |
| Approved, not merged | All checks pass + approved | → Step 7: Merge prompt |
| Aging | Open > N days with no activity | → Step 8: Age alert |
| Blocked | Explicit "blocked" label or comment | → Escalate, do not touch |
# Check CI status for a specific PR
gh pr checks <number> --json name,state,conclusion 2>/dev/null
# Check merge status
gh pr view <number> --json mergeable,mergeStateStatus,reviewDecision
Step 3 — CI failure: diagnose and propose fix
# Get failing check logs
gh run view --log-failed $(gh pr checks <number> --json databaseId --jq '.[0].databaseId') 2>/dev/null | tail -50
Classify the failure:
- Flake: Same test has failed intermittently on other PRs → re-run the check, note as flake
- Regression: New failure introduced by this PR → invoke
/minimal-fixwith the error - Env/infra: Missing secret, broken runner → escalate to human immediately
- Unrelated: Failure pre-dates the PR → note and skip (not the PR author's problem)
Cap fix attempts at 3 per PR. After 3 → escalate with full context.
# Re-run a flaky check
gh run rerun <run-id> --failed
Step 4 — Rebase nudge
If the branch is behind main but has no conflicts, comment with the rebase command:
gh pr comment <number> --body "Branch is behind main. To update:
\`\`\`
git fetch origin
git rebase origin/main
git push --force-with-lease
\`\`\`"
Do not auto-rebase unattended — conflicts require human judgment.
Step 5 — Review nudge
If no reviewer has been assigned after N hours (default: 24h):
gh pr edit <number> --add-reviewer <suggested-reviewer>
# or
gh pr comment <number> --body "This PR is awaiting review assignment."
Only assign reviewers if the project has a CODEOWNERS file or a known reviewer rotation.
Step 6 — Changes requested
Surface the blocking review comments to the PR author. Do not attempt to resolve review feedback automatically.
gh pr view <number> --json reviews --jq '.reviews[] | select(.state == "CHANGES_REQUESTED") | .body'
Step 7 — Approved and ready to merge
# Confirm all checks pass + required reviews met
gh pr view <number> --json mergeable,reviewDecision,statusCheckRollup
If mergeable == MERGEABLE and reviewDecision == APPROVED → post a merge-ready comment or merge if auto-merge policy allows.
Default: never auto-merge. Only merge automatically if --auto-merge flag is explicitly passed AND the PR is in the allow list (no denylist files changed).
Step 8 — Age alert
Flag PRs open longer than the threshold without activity:
# PRs with no commits, comments, or review activity in N days
gh pr list --state open --json number,title,updatedAt | \
jq --arg cutoff "$(date -d '7 days ago' -u +%Y-%m-%dT%H:%M:%SZ)" \
'.[] | select(.updatedAt < $cutoff)'
Post a comment asking for status update. Do not close — human decides.
Step 9 — Run report
Output a summary after each sweep:
PR BABYSITTER — sweep complete (2026-06-30 10:45 UTC)
CI Red (2)
#1241 fix/auth-refresh → minimal-fix proposed (attempt 1/3)
#1238 feat/search-v2 → env failure, escalated to human
Awaiting Review (1)
#1245 docs/api-update → 26h with no reviewer, reviewer assigned
Aging (1)
#1230 refactor/legacy-model → 9 days no activity, comment posted
Approved + Ready (1)
#1243 fix/typo → ready to merge (awaiting /pr-babysitter --auto-merge)
Nothing blocking (3): #1240, #1242, #1244
Edge Cases
- No open PRs: Report "Nothing to watch" and exit cleanly.
- No CI configured: Skip CI checks; focus on review state and aging.
- Rate limits on gh: Back off and note in report.
- PR with
blockedlabel: Skip entirely — do not comment, do not attempt fix.
Token Optimization
Expected range: 400–1,200 tokens per sweep; 50–100 tokens (no open PRs, early exit)
Patterns used: Bash for gh CLI queries, progressive disclosure (summary then per-PR detail), early exit (no PRs = immediate exit), attempt cap (max 3 fix attempts per PR)
Caching: Stores per-PR attempt counts in .claude/cache/pr-babysitter/attempts.json (reset when PR is closed or merged).
Early exit: If gh pr list --state open returns 0 results → report and exit immediately.