AI Explore Review Loop Plugin

An open-source Claude Code plugin that lets a coding-agent session wait for a durable, exact-head pull-request review and continue the same review/fix/push loop without polling GitHub.

The plugin combines:

• PostToolUse, Stop, and SessionStart Hooks;
• an ordinary stdio MCP server;
• Supabase durable reads plus Realtime wake-ups;
• exact PR head-SHA validation;
• idempotent local and server-side review receipts;
• skills for status, pause, resume, stop, and review handling.

It does not require Anthropic MCP Channels and can run in Claude Code setups that use compatible third-party model providers.

What is included

This public repository contains only the installable plugin, its buildable source, tests, and public documentation. It intentionally excludes the private server-side control plane, Pod addresses, GitHub App configuration, Supabase migrations, Hermes deployment scripts, and operational baselines.

The full system requires a compatible backend exposing these authenticated Supabase RPCs:

• get_ai_explore_review_state(repo, pr_number, head_sha);
• ack_ai_explore_review_result(result_id, client_id, session_hash, status, new_head).

The authenticated user must also have RLS SELECT access to the corresponding github_pr_state row for Realtime wake-ups. Realtime is never treated as final delivery; the MCP tool always re-reads durable state.

Rush marketplace status

Rush marketplace distribution is not enabled yet. The current Rush Git importer and installer preserve recognized Skills and .mcp.json, but omit the Hook, bin/, and dist/ files required by this plugin. A trial entry was removed after this was verified with a real install.

Do not publish a reduced Skills+MCP package under the same name: it would appear installed while lacking the automatic review-loop and its MCP executable. Rush distribution can be enabled after the platform preserves arbitrary Claude plugin assets and Hooks end to end.

Install from GitHub without a checkout

claude plugin marketplace add kanyun-inc/ai-explore-review-loop-plugin
claude plugin install ai-explore-review-loop@ai-explore --scope user

Configure

The plugin needs a Supabase project URL, public anon/publishable key, and a developer Auth session. Run the login command through the public GitHub package in a real terminal, so no repository checkout is needed and the password does not enter the model transcript:

npm exec --yes \
  --package=github:kanyun-inc/ai-explore-review-loop-plugin \
  -- ai-explore-review-loop login \
  --url 'https://your-project.example.com' \
  --anon-key 'YOUR_PUBLIC_ANON_OR_PUBLISHABLE_KEY' \
  --email 'developer@example.com'

Passwords are read without echo and are not stored. Runtime state and refresh tokens live outside the repository in a mode-0600 user configuration file.

Develop

Requires Node.js 22 or newer:

npm ci
npm run build
npm run check
npm run validate:plugin
npm run security:scan

Generated plugin bundles are committed so marketplace installation does not run npm install on developer machines.

Security properties

• no database URL or Supabase service-role credential;
• no GitHub App private key or worker password;
• developer RLS instead of shared admin credentials;
• exact-head binding before returning or handling a review;
• duplicate Realtime events are harmless;
• Stop Hook fails open on infrastructure/authentication errors rather than trapping the developer;
• GitHub review text is treated as untrusted input and verified against the current diff.

See SECURITY.md for vulnerability reporting.

License

Apache License 2.0. See LICENSE, NOTICE, and THIRD_PARTY_NOTICES.md.