AI Code Reviewer

An autonomous AI agent that continuously reviews pull requests on open source repositories.

What it does

1. Claims open source repositories from a pool
2. Clones each repo, understands the codebase structure
3. Monitors for open pull requests
4. Reviews each PR with full context — reads the diff, explores surrounding code, checks cross-file impact
5. Posts a detailed review comment on the PR
6. Moves on to the next repo, claims more when idle
7. Runs forever

How it works

The agent uses the Claude Agent SDK as its harness, driving the claude CLI (Claude Code) in a persistent loop. Each cycle:

• Call a claim API to get assigned repositories
• For each repo: git clone, check open PRs, fetch diffs
• For unreviewed PRs: read the diff + surrounding code, analyze for bugs, security, performance, architecture
• Post a review comment via GitHub API
• Report back to the claim API
• Sleep 30 seconds, repeat

The agent maintains state across cycles via session resume. It remembers which PRs it already reviewed and which repos it monitors.

Review format

Each review comment includes:

• Summary — what the PR does
• Architecture — how it fits the codebase
• Issues — file, severity (critical/warning/suggestion), explanation, fix
• Cross-file impact — anything in other files affected
• Assessment — approve / request-changes / comment

Tech stack

• Agent harness: Claude Agent SDK (Python) spawning the claude CLI
• LLM: Anthropic-compatible endpoint (Anthropic, z.ai GLM, etc.) — swap by pointing ANTHROPIC_BASE_URL at a different provider
• GitHub API: For reading PRs and posting comments
• Claim API: Central coordination so multiple agents don't review the same repos

Requirements

• Node.js (for @anthropic-ai/claude-code)
• Python 3.12+ (for claude_agent_sdk)
• GitHub PAT with public_repo scope
• An Anthropic-compatible LLM key (Anthropic OAuth, z.ai, etc.)

Configuration

Set these environment variables:

GITHUB_TOKEN=ghp_...                            # GitHub PAT for reading PRs and posting comments
ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic   # Or https://api.anthropic.com for Anthropic direct
ANTHROPIC_AUTH_TOKEN=...                        # z.ai API key, or use CLAUDE_CODE_OAUTH_TOKEN for Anthropic OAuth

Running

Install the harness:

npm install -g @anthropic-ai/claude-code
pip install claude_agent_sdk

Run the watchdog (which starts and supervises the agent loop):

./src/watchdog.sh

License

MIT