AI-powered tool that autonomously fixes CI failures, resolves merge conflicts, and merges GitHub pull requests using Claude AI.
- Fix any PR - Use the CLI to fix CI failures on any GitHub pull request
- Auto-fix - Automatically fixes test failures, linting issues, security vulnerabilities, merge conflicts, and build errors
- Smart classification - Categorizes failures and applies appropriate fix strategies
- Organization-wide automation - Includes workflows to scan VectorInstitute repos daily for bot PRs (Dependabot, pre-commit-ci)
- Observable - Full execution tracing with dashboard analytics
- Transparent - Comments on PRs with status updates
┌─────────────────────────────────────────────────────────────┐
│ aieng-bot │
│ │
│ CLI Usage (any PR): │
│ aieng-bot fix --repo owner/repo --pr 123 │
│ │
│ Automated Workflows (bot PRs only): │
│ Daily scan (00:00 UTC) for Dependabot/pre-commit-ci PRs │
│ with failing checks across VectorInstitute org │
└─────────────────────────────────────────────────────────────┘
1. Create Anthropic API Key
- Get from Anthropic Console
- Add as repository secret:
ANTHROPIC_API_KEY
2. Create GitHub Personal Access Token
- Go to Settings → Developer settings → Personal access tokens → Fine-grained tokens
- Configure: Resource owner:
VectorInstitute, Repository access:All repositories - Permissions:
contents: write,pull_requests: write,issues: write - Add as repository secret:
ORG_ACCESS_TOKEN
3. Enable GitHub Actions
- Go to Actions tab → Enable workflows
The bot now monitors all VectorInstitute repositories automatically.
1. Classification
- Analyzes PR and failure logs using Claude Haiku 4.5
- Categorizes failure type: test, lint, security, build, or merge conflict
- Routes to appropriate fix strategy
2. Fix Loop
- Clones target repository and PR branch
- Loads appropriate AI prompt template for the failure type
- Uses Claude Agent SDK to automatically apply fixes
- Commits and pushes fixes to PR
- Polls CI status, retries on failure (up to max_retries)
3. Merge
- Auto-merges when all checks pass
- Comments on PR with status updates
Automated Bot PR Monitoring (optional, for VectorInstitute org)
- Daily scan at 00:00 UTC for Dependabot/pre-commit-ci PRs with failures
- Automatically dispatches fix jobs for discovered PRs
Required Secrets
ANTHROPIC_API_KEY- Anthropic API access for ClaudeORG_ACCESS_TOKEN- GitHub PAT with org-wide permissions
Model Configuration
- Classification: Claude Haiku 4.5 (
claude-haiku-4-5-20251001) - cost-efficient - Fixing: Claude Sonnet 4.5 (
claude-sonnet-4-5-20250929) - agentic capability - Override with
CLAUDE_MODELenvironment variable
Workflows
discover-and-dispatch.yml- Daily scan (00:00 UTC) for Dependabot/pre-commit-ci PRs with failuresfix-pr-agent.yml- Per-PR agent workflow (6-hour timeout), can be triggered manually for any PRcode_checks.yml- Ruff + mypy checksunit_tests.yml- pytest suite
AI Prompt Templates (customize for your needs)
fix-merge-conflicts.md- Resolve merge conflicts with best practicesfix-test-failures.md- Test failure resolution strategiesfix-lint-failures.md- Linting/formatting fixesfix-security-audit.md- Security vulnerability handlingfix-build-failures.md- Build/compilation error fixes
Can fix:
- Merge conflicts (dependency files, lock files, code)
- Linting and formatting issues
- Security vulnerabilities (dependency updates)
- Simple test failures from API changes
- Build configuration issues
Cannot fix:
- Complex logic errors
- Breaking changes requiring refactoring
- Issues requiring architectural decisions
# Install dependencies
uv sync
# Fix and merge a PR
aieng-bot fix --repo owner/repo --pr 123
# Fix with dashboard logging
aieng-bot fix --repo owner/repo --pr 123 --log
# Custom retries and timeout
aieng-bot fix --repo owner/repo --pr 123 --max-retries 5 --timeout-minutes 180Trigger via CLI:
# Fix a PR using the CLI (recommended)
aieng-bot fix --repo owner/repo --pr 123
# Or trigger via GitHub workflow
gh workflow run fix-pr-agent.yml \
--field target_repo="owner/repo" \
--field pr_number="123"
# Run the bot PR discovery workflow (VectorInstitute org only)
gh workflow run discover-and-dispatch.ymlTrigger via GitHub UI: Actions → Select workflow → Run workflow → Enter parameters
View comprehensive analytics and agent execution traces:
- 📊 Dashboard - Interactive dashboard with:
- Overview table of all PR fixes
- Success rates and performance metrics
- Detailed agent execution traces (like LangSmith/Langfuse)
- Code diffs with syntax highlighting
- Failure analysis and reasoning timeline
Features:
- Real-time PR status tracking
- Agent observability (tool calls, reasoning, actions)
- Historical metrics and trends
- Per-repo and per-failure-type analytics
- Sortable/filterable PR table
Authentication:
- Restricted to @vectorinstitute.ai email addresses
- Google OAuth 2.0 sign-in
View activity:
- Dashboard - Comprehensive analytics and traces
- Actions tab - All workflow runs and success/failure rates
- PR comments - Detailed status updates on each PR
- Run summary - PR count and actions taken per run
Debug commands:
# View recent workflow runs
gh run list --workflow=discover-and-dispatch.yml --limit 5
# View logs for specific run
gh run view RUN_ID --log- Setup Guide - Configuration and permissions
- Deployment Guide - Rollout and monitoring
| Issue | Solution |
|---|---|
| Workflow doesn't run | Check Actions enabled and secrets are set |
| Can't find PRs | Verify ORG_ACCESS_TOKEN has correct permissions |
| Can't merge PRs | Ensure token has contents: write permission |
| Can't push fixes | Check token has write access to target repos |
| Claude API errors | Verify ANTHROPIC_API_KEY is valid |
| Rate limits | Reduce monitoring frequency in workflow cron schedule |
See Setup Guide for detailed troubleshooting.
🤖 aieng-bot - AI-powered PR maintenance by Vector Institute AI Engineering