All projects
AI

Fullstack Orchestrator

A multi-agent skill: one model plans, one codes, a rival vendor reviews

View on GitHub

The problem

A single coding agent writes code and then grades its own work, so its blind spots ship with the code.

The solution

Fullstack Orchestrator is an agent skill for Claude Code, Codex, Cursor and Qoder. It turns the host into an orchestrator that only plans, writes task briefs, triages findings and verifies. DeepSeek writes all the code through OpenCode, and a model from the other vendor reviews every task read-only. Security rules and starter templates are built in, so the code starts secure.

Highlights

  • Strict roles: the orchestrator may only write plan.md and task briefs, never application code
  • Cross-vendor review: GPT reviews when Claude orchestrates and Opus reviews under Codex. The run fails if the reviewer edits a file
  • Fix loop in one persistent dev session per task, escalating to a human after 3 failed review rounds
  • Built-in security rules: AES-256-GCM field encryption, IDOR-safe owner-scoped data access, email OTP + TOTP MFA
  • Starter templates for a Clean Architecture Express + PostgreSQL server and a React client, with an IDOR test suite
  • One-command install with doctor and uninstall; model routing lives in a single config file