AI-powered PR babysitter — automatically fixes merge conflicts, failing CI, and unresolved review comments using Claude Code

---

AI-powered PR babysitter that runs locally on your Mac. Automatically fixes merge conflicts, failing CI, and unresolved review comments using Claude Code.

Runs every 5 minutes via macOS launchd. No GitHub Actions minutes burned — uses your own Claude tokens.

How it works

1. Scans your open PRs on GitHub
2. For each PR with issues (merge conflicts, failing checks, unresolved reviews):
   • Clones the repo into a temp directory
   • Invokes Claude Code CLI to diagnose and fix the issues
   • Runs your project's verify/test/lint commands before pushing
   • Pushes fixes and resolves review threads
3. Tracks attempts via GitHub labels (babysitter-attempt-N)
4. Gives up after 10 attempts (babysitter-gave-up label)

Prerequisites

• macOS (uses launchd for scheduling)
• Claude Code CLI installed and authenticated
• GitHub CLI (gh) installed and authenticated
• jq installed (brew install jq)

Review SECURITY.md before starting unattended automation. It covers the trust model, required GitHub token permissions, log sensitivity, and revocation steps.

Install

git clone https://github.com/mstuart/pr-babysitter.git
cd pr-babysitter
./install.sh

This copies babysitter and pr-babysitter.sh to ~/.local/bin/.

Update later with git pull --ff-only && ./install.sh. Uninstall with ./install.sh --uninstall after stopping the agent with babysitter stop.

Setup

cd /path/to/your/repo
babysitter init          # creates .babysitterrc
# edit .babysitterrc with your repo and username
babysitter doctor        # read-only preflight check
babysitter start         # starts the launchd agent

Configuration

Place a .babysitterrc in your repo root. See .babysitterrc.example for all options.

Required settings:

BABYSITTER_REPO="owner/repo"        # GitHub repo
BABYSITTER_AUTHOR="your-username"   # Only watch your PRs

Optional settings:

BABYSITTER_MAX_ATTEMPTS=10          # Give up after N attempts
BABYSITTER_MODEL=sonnet             # Claude model
BABYSITTER_INTERVAL=300             # Seconds between runs

Command examples. Keep these non-interactive and avoid destructive operations such as deploys, seeds, database resets, package publishing, git push, git reset, or starting long-running services.

# npm
BABYSITTER_INSTALL_CMD="npm ci --prefer-offline --no-audit --no-fund"
BABYSITTER_VERIFY_CMD="npm run verify"
BABYSITTER_TEST_CMD="npm test -- --runInBand"
BABYSITTER_LINT_CMD="npm run lint:fix"

# pnpm
BABYSITTER_INSTALL_CMD="pnpm install --frozen-lockfile --ignore-scripts"
BABYSITTER_VERIFY_CMD="pnpm verify"
BABYSITTER_TEST_CMD="pnpm test -- --runInBand"
BABYSITTER_LINT_CMD="pnpm lint:fix"

# Turborepo
BABYSITTER_INSTALL_CMD="pnpm install --frozen-lockfile --ignore-scripts"
BABYSITTER_VERIFY_CMD="pnpm turbo run typecheck build --filter=...[HEAD^]"
BABYSITTER_TEST_CMD="pnpm turbo run test --filter=...[HEAD^]"
BABYSITTER_LINT_CMD="pnpm turbo run lint --filter=...[HEAD^]"

# Python
BABYSITTER_INSTALL_CMD="python -m pip install -r requirements.txt"
BABYSITTER_VERIFY_CMD="python -m compileall ."
BABYSITTER_TEST_CMD="python -m pytest"
BABYSITTER_LINT_CMD="python -m ruff check --fix ."

Custom prompts

For full control over what Claude does, point to a prompt template file:

BABYSITTER_PROMPT_FILE=".babysitter-prompt.txt"

The file supports variable substitution via envsubst: $REPO, $NUMBER, $TITLE, $ATTEMPT, $MAX, $MERGEABLE, $FAILING, $UNRESOLVED, $VERIFY_CMD, $TEST_CMD, $LINT_CMD.

Or append extra rules to the default prompt:

BABYSITTER_EXTRA_RULES="- This is a monorepo with Turborepo
- NEVER run 'npm run seed' (destroys all data)"

Config resolution

Config is loaded from the first match:

1. $BABYSITTER_CONFIG env var (explicit path)
2. .babysitterrc (walking up from cwd to /)
3. ~/.config/pr-babysitter/config

When babysitter start creates the launchd plist, it records the resolved config path in BABYSITTER_CONFIG so unattended runs use the same file checked by the CLI.

CLI

babysitter <command>

COMMANDS
  init         Create a .babysitterrc config in the current directory
  doctor       Check config, auth, dependencies, and command safety without starting automation
  start        Load and start the launchd agent
  stop         Unload and stop the launchd agent
  status       Show agent status and current PR overview
  dashboard    Live-updating dashboard (refreshes every 30s)
  run          Run one babysitter cycle immediately (foreground)
  logs         Tail the babysitter log
  logs clear   Clear the log file
  reset <pr>   Remove all babysitter labels from a PR
  reset-all    Remove all babysitter labels from all open PRs
  version      Print version

Dashboard

babysitter dashboard       # refreshes every 30s
babysitter dashboard 10    # refreshes every 10s

Shows: agent status, each PR's health (conflicts, failing checks, unresolved threads), attempt counts, and recent log activity.

What it fixes

Issue	How
Merge conflicts	git merge origin/main + AI conflict resolution
TypeScript errors	Runs verify command, fixes reported errors
Test failures	Runs test command, fixes failing tests
Lint errors	Runs lint command with auto-fix
Review comments	Reads unresolved threads, makes requested code changes, resolves threads

The babysitter will not:

• Force-push
• Guess at answers to discussion questions in reviews
• Push code that fails verification
• Touch PRs by other authors

Labels

Label	Meaning
babysitter-attempt-N	The babysitter has made N fix attempts on this PR
babysitter-gave-up	Exceeded max attempts — babysitter will no longer touch this PR

Use babysitter reset <pr-number> to clear labels and let it try again.

License

MIT