What is pr-agent?

pr-agent is a GitHub App webhook service that automatically reviews pull requests using an AI agent. When someone opens, updates, or reopens a pull request, the service receives a webhook notification from GitHub, runs an agent loop that inspects the PR, and posts a review comment summarizing its findings. It also responds to slash commands typed in PR comments: /review triggers a manual review on demand. The agent loop works by giving an AI model access to GitHub's REST API through a tool-based interface. The model calls these tools to examine code changes, read file contents, and post a review. The loop runs until the model finishes or hits a configurable round limit (MAX_TOOL_ROUNDS). A finalization step ensures the loop always ends with a complete text response rather than an unfinished tool call. Webhook deliveries are deduplicated using the X-GitHub-Delivery header (or a SHA-256 hash of the body as fallback), so retries from GitHub do not trigger duplicate reviews. Concurrent review runs are bounded by a REVIEW_CONCURRENCY setting (default 2) to prevent large bursts of pull request activity from spawning unbounded parallel AI calls. Incoming webhook payloads are validated with Zod shapes before processing. The service is built with TypeScript and uses Effect TS as its runtime. It can be deployed using Docker Compose with a multi-stage Dockerfile, exposes a /health endpoint, and requires a GitHub App configured with repository permissions for issues, pull requests, and contents. A provider API key (for the underlying LLM) must also be set in the environment.