agent-reviews

Agent skills automate the entire triage-fix-reply loop for PR reviews. The workflow follows these steps:

1. Fetch: Retrieves unanswered comments (filtered by all, bot-only, or human-only).
2. Evaluate: Analyzes each finding (e.g., true positive, false positive, actionable).
3. Fix: Applies fixes for real issues and runs lint/type-checks.
4. Dismiss: Dismisses false positives with an explanation.
5. Reply: Responds to every comment with the outcome (e.g., Fixed in {commit} for true positives or Won't fix: {reason} for false positives).
6. Watch: Enters a watch mode to poll for new comments and repeats the cycle until no new comments appear for 10 minutes.
7. Report: Provides a summary of all actions taken.

All fixes are batched into a single commit before the polling phase begins.

The agent-reviews package provides three distinct skills tailored for different workflows. Instead of a single tool, you choose a skill based on whether you want to address human feedback, bot feedback, or both:

• resolve-reviews: Resolves all comments (both human and bot).
• resolve-agent-reviews: Resolves bot comments only.
• resolve-human-reviews: Resolves human comments only.

These skills are compatible with any agent that supports Agent Skills, such as Claude Code, Cursor, or Codex.

The resolve-agent-reviews skill follows a two-phase lifecycle to automate PR maintenance:

Phase 1: FETCH & FIX (Synchronous)

1. Fetch: Retrieve all unanswered bot comments using --expanded.
2. Evaluate: Categorize findings as TRUE POSITIVE (fix it), FALSE POSITIVE (dismiss it), or UNCERTAIN (ask the user).
3. Fix & Commit: Apply code changes, run lint/type-checks, and push a single commit containing all fixes.
4. Reply: Respond to every comment. Use --resolve only for false positives or skipped items to ensure reviewers can still verify actual fixes.

Phase 2: POLL FOR NEW COMMENTS (Looping)

After Phase 1, start a watcher loop using --watch. This phase runs until the watcher exits with WATCH COMPLETE, indicating no new bot findings were posted within the timeout period.

npx agent-reviews --bots-only --unanswered --expanded

To use the resolve-reviews skill, ensure the following are installed and configured:

• Tools: git, gh (GitHub CLI), and Node.js.
• Package Manager: You can use npx, pnpm dlx, yarn dlx, or bunx to run agent-reviews commands.
• Cloud Environments: If running in a remote environment (like Codespaces), verify your git identity by running git config --global --get user.email. If it is empty or a placeholder, set it manually so commits map to your user.

Note: In local environments, you can skip the git identity check.

To use the resolve-human-reviews skill, ensure the following are installed and configured:

• Tools: git, gh (GitHub CLI), and Node.js.
• Package Manager: You can use npx, pnpm dlx, yarn dlx, or bunx to run agent-reviews commands.
• Cloud Environments (Codespaces/Remote Agents): You must verify your git author identity so CI can map commits to you. Check your email with git config --global --get user.email. If it is empty or a placeholder, set it manually before proceeding.

Note: In local environments, the git identity check can be skipped.

To use the resolve-agent-reviews workflow, ensure the following are installed and configured:

• Tools: git, gh (GitHub CLI), and Node.js.
• Package Manager: You can use npx, pnpm dlx, yarn dlx, or bunx to run agent-reviews commands.
• Git Identity (Cloud/Remote environments only): In environments like Codespaces or remote agents, verify your git author identity so commits map to your user. If git config --global --get user.email is empty, set it manually.

Note: Skip the git identity check if working in a local environment.

git config --global user.email "your-email@example.com"

agent-reviews automatically detects authentication using the following resolution order (first match wins):

1. GITHUB_TOKEN environment variable
2. GH_TOKEN environment variable
3. .env.local in the repository root
4. gh auth token (via the GitHub CLI)

If you are already logged in with gh auth login, no additional configuration is required. For cloud/remote environments or HTTPS proxy setups, providing GITHUB_TOKEN or GH_TOKEN is recommended. The tool also supports HTTPS_PROXY via undici ProxyAgent support.

You can add agent-reviews as a slash command to any agent that supports Agent Skills (such as Claude Code, Cursor, or Codex). This allows the agent to automate the triage-fix-reply loop.

Skills use npx agent-reviews at runtime, so the CLI is fetched automatically.

Use the following command to add a skill:

npx skills add pbakaus/agent-reviews@<skill-name>

Available skills:

• resolve-reviews: Resolves all comments (human + bot).
• resolve-agent-reviews: Resolves bot comments only (Copilot, Cursor, etc.).
• resolve-human-reviews: Resolves human comments only.

npx skills add pbakaus/agent-reviews@resolve-agent-reviews

You can install the agent-reviews CLI globally using npm to manage GitHub PR review comments directly from your terminal.

npm install -g agent-reviews

You can redirect the tool to use GitHub Enterprise Server (GHES) or a local API-compatible server by setting the following environment variables:

• GITHUB_API_URL: Sets the base URL for REST and GraphQL endpoints. For GHES, providing a URL ending in /api/v3 will automatically have the GraphQL endpoint rewritten to /api/graphql.
• GITHUB_GRAPHQL_URL: An optional variable for specifying a custom GraphQL endpoint in unusual setups.
• GH_TOKEN: An alternative to GITHUB_TOKEN for authentication.
• GH_REPO: Used to target specific repositories when running in detached environments.