Multi AI

Multi-AI Pipeline Orchestrator

You are starting the multi-AI pipeline. Follow this process exactly.

Reference Documents

First, read the standards that guide all reviews:

• skill/multi-ai/reference/standards.md - Coding standards and review criteria

Step 1: Prepare task state

Inspect .task/ if it exists. Resume an active run instead of deleting its state. For an explicitly requested new run, archive the previous directory to a unique sibling path and verify the archive, then create a fresh .task/:

mkdir -p .task

Do not overwrite an unfinished task or discard its review evidence.

Step 2: Capture User Request

Write the user's request to .task/user-request.txt.

Step 3: Create Initial Plan

Write .task/plan.json:

{
  "id": "plan-YYYYMMDD-HHMMSS",
  "title": "Short descriptive title",
  "description": "What the user wants to achieve",
  "requirements": ["req1", "req2"],
  "created_at": "ISO8601",
  "created_by": "claude"
}

Step 4: Refine Plan

Research the codebase and create .task/plan-refined.json:

{
  "id": "plan-001",
  "title": "Feature title",
  "description": "What the user wants",
  "requirements": ["req1", "req2"],
  "technical_approach": "Detailed how-to",
  "files_to_modify": ["path/to/file.ts"],
  "files_to_create": ["path/to/new.ts"],
  "dependencies": [],
  "estimated_complexity": "low|medium|high",
  "potential_challenges": ["Challenge and mitigation"],
  "refined_by": "claude",
  "refined_at": "ISO8601"
}

Step 5: Sequential Plan Reviews

Run reviews in sequence. Fix issues after each before continuing:

1. Invoke /review-sonnet

   • Read .task/review-sonnet.json result
   • If needs_changes: fix issues in plan, update .task/plan-refined.json

2. Invoke /review-codex

   • Read .task/review-codex.json result
   • If needs_changes: fix issues and restart from step 5.1
   • If approved: continue to implementation

Step 6: Implement

Invoke /implement-sonnet

This skill will:

• Read the approved plan from .task/plan-refined.json
• Implement the code
• Add tests
• Output to .task/impl-result.json

Step 7: Sequential Code Reviews

Run reviews in sequence. Fix issues after each before continuing:

1. Invoke /review-sonnet

   • Read .task/review-sonnet.json result
   • If needs_changes: fix code issues

2. Invoke /review-codex

   • Read .task/review-codex.json result
   • If needs_changes: fix issues and restart from step 7.1
   • If approved: continue to completion

Step 8: Complete

Write .task/state.json:

{
  "state": "complete",
  "plan_id": "plan-001",
  "completed_at": "ISO8601"
}

Report success to the user with:

• Summary of what was implemented
• Files changed
• Tests added

Important Rules

• Follow this process exactly - no shortcuts
• Resolve review findings before continuing. If reviewers disagree or a cycle repeats without new evidence or a relevant change, report the unresolved blocker; do not loop indefinitely or mark the task complete.
• If codex rejects, restart the review cycle from sonnet
• Keep the user informed of progress at each major step

State Files Reference

Reference Directory