patchwork-ai

CLI tool that analyzes GitHub issues, generates fixes, and opens pull requests automatically. Point it at an issue URL and it handles the rest — cloning, branching, fixing, testing, and PR creation.

Install

npx patchwork-ai --help

Or install globally:

npm install -g patchwork-ai

Setup

1. Configure your AI provider

patchwork-ai config set api-key <your-api-key>
patchwork-ai config set provider anthropic   # or openai
patchwork-ai config set model claude-sonnet-4-6

2. Set your GitHub token

export GITHUB_TOKEN=ghp_xxxxx

The token needs repo and workflow scopes for forking repos and creating PRs.

Usage

Fix an issue

Analyze the issue, generate a fix, run quality checks, and create a PR:

patchwork-ai fix https://github.com/owner/repo/issues/42

Override the AI model:

patchwork-ai fix https://github.com/owner/repo/issues/42 --model claude-sonnet-4-6

Analyze only (dry run)

Check whether an issue is suitable for auto-fixing without making changes:

patchwork-ai analyze https://github.com/owner/repo/issues/42

Discover issues

Find issues tagged "good first issue", "help wanted", or "bug" in a repo:

patchwork-ai discover https://github.com/owner/repo
patchwork-ai discover https://github.com/owner/repo --limit 20

Manage config

patchwork-ai config set api-key sk-xxxxx
patchwork-ai config get

Config is stored at ~/.patchwork-ai/config.yaml.

Quality Gate

Every fix goes through automated quality checks before a PR is created:

• AI artifact detection — strips any AI markers or co-authored-by lines
• Scope check — rejects changes touching more than 10 files
• Necessity review — AI verifies the fix is minimal and focused
• Readability review — checks code quality and style
• Side effect analysis — flags potential regressions
• Test execution — runs the repo's test suite

If any check fails, the pipeline stops and no PR is created.

Supported Providers

| Provider | Status | Default Model | |----------|--------|---------------| | Anthropic | Recommended | claude-sonnet-4-6 | | OpenAI | Supported | gpt-4o |

How It Works

1. Analyze — Fetches the issue, checks for existing PRs and claims, asks AI to assess difficulty
2. Repo setup — Forks the repo, clones it, creates a feature branch
3. Codebase analysis — Maps the file tree, identifies relevant files and conventions
4. Fix generation — AI produces minimal code changes following the repo's style
5. Quality gate — Automated checks ensure the fix is clean and focused
6. Test — Runs the project's test suite
7. PR creation — Pushes the branch and opens a pull request referencing the issue

Development

npm install
npm run build
npm test

License

MIT