Multi-Model Collaborative Execution

Advanced skill — Uses external model backends to generate code prototypes, then refactors into production-grade implementations.

Usage

/multi-execute .codex/plan/<feature-name>.md
/multi-execute <direct task description>

Workflow

Phase 1: Read Plan

• Parse plan file or task description
• Extract: task type, steps, key files, session IDs
• Route: Frontend → Gemini, Backend → Codex, Fullstack → parallel

Phase 2: Context Retrieval

• Retrieve project context for target files
• Ensure complete definitions and signatures available

Phase 3: Prototype Acquisition

• Call external model backends for code prototypes
• Receive unified diff patches (models have zero write access)

Phase 4: Code Implementation

• Parse received diffs
• Simulate applying to target files (mental sandbox)
• Refactor "dirty prototype" to production-grade code
• Apply changes with minimal scope
• Run self-verification (lint, typecheck, tests)

Phase 5: Audit and Delivery

• Call external models for code review
• Integrate review feedback
• Execute necessary fixes
• Report changes, audit results, and recommendations

Key Rules

1. Code Sovereignty — All file modifications by the orchestrator
2. Dirty Prototype Refactoring — External model output is a draft, must refactor
3. Minimal Changes — Only modify necessary code
4. Mandatory Audit — Must perform review after changes