clauxsync

name	ClauxSync
description	Use when starting any development task - creates a Manus-style execution plan with todo.md tracking, assigns Claude and Codex roles per phase with independent comparison pattern (Co-Commands compatible), and manages the full task lifecycle with reflection checkpoints

Overview

ClauxSync combines Claude Code + OpenAI Codex via OAuth with Manus-style task orchestration. Every task gets a structured plan, persistent todo.md tracking, and dual-AI validation.

Phase 1: Task Setup (Every Task Starts Here)

When this skill activates:

Step 1 — Model Selection

First, call clauxsync_models to show available models. Present to user:

🤖 Codex 모델 선택:

#	모델	설명	추천 용도
1	gpt-5.4	최신 플래그십 — 추론, 코딩, 에이전트 최강	복잡한 아키텍처, 대규모 리팩토링
2	gpt-5.3-codex	코딩 특화 최강 모델 (기본값)	코드 리뷰, 디버깅, 테스트
3	gpt-5.3-codex-spark	초고속 실시간 (Pro 전용)	빠른 반복 작업
4	gpt-5-codex-mini	경량 — 구독당 4배 더 사용 가능	간단한 리뷰, 비용 절약

선택하지 않으면 gpt-5.3-codex가 기본 사용됩니다. 모델은 clauxsync_set_model로 변경하거나, 각 도구 호출 시 model 파라미터로 오버라이드 가능합니다.

Step 2 — Mode Selection

Present to user:

🔧 모드 선택:

1. 직접 선택 모드 — 각 작업별 AI를 직접 지정
2. 추천 모드 — 자동 배분 (리팩토링/리뷰/디버깅/테스트 → Codex 필수)

Step 3 — Role Assignment Table

작업 유형	Claude	Codex	추천모드 기본값
브레인스토밍	주도	반론	Claude 주도 + Codex 반론
계획/설계	주도	-	Claude 주도
코드 작성	단독	-	Claude 단독
리팩토링	작성	리뷰 필수	Claude → Codex 리뷰 필수
코드 리뷰	종합	필수	Codex 필수 + Claude 종합
디버깅	분석	검증 필수	Claude → Codex 검증 필수
테스트 작성	작성	검증 필수	Claude → Codex 검증 필수

Step 4 — Create Execution Plan

After mode selection, create a numbered execution plan:

## Execution Plan
Task: [task description]
Mode: [직접선택/추천]
Created: [timestamp]

### Steps
1. [ ] 분석 — 요구사항 및 현재 상태 파악
2. [ ] 설계 — 구현 방향 결정 [Claude/Codex: based on mode]
3. [ ] 구현 — 코드 작성 [Claude/Codex: based on mode]
4. [ ] 검증 — 리뷰/테스트/디버깅 [Claude/Codex: based on mode]
5. [ ] 완료 — 결과 제출 및 정리

Step 5 — Create todo.md

Create a todo.md file in the project root as a persistent live checklist:

# ClauxSync Todo — [Task Name]
> Mode: 추천모드 | Created: YYYY-MM-DD HH:mm

## Plan
- [ ] Step 1: 분석 — [specific details]
- [ ] Step 2: 설계 — [specific details]
- [ ] Step 3: 구현 — [specific details]
- [ ] Step 4: 검증 — [specific details]
- [ ] Step 5: 완료 — [specific details]

## AI Assignment
- 브레인스토밍: Claude 주도 / Codex 반론
- 계획: Claude
- 코드작성: Claude
- 리팩토링: Claude → Codex 리뷰
- 코드리뷰: Codex → Claude 종합
- 디버깅: Claude → Codex 검증
- 테스트: Claude → Codex 검증

## Progress Log
| Step | Status | AI Used | Reflection |
|------|--------|---------|------------|
| 1 | ⏳ | - | - |
| 2 | ⏳ | - | - |
| 3 | ⏳ | - | - |
| 4 | ⏳ | - | - |
| 5 | ⏳ | - | - |

Phase 2: Agent Loop (Task Execution)

For each step in the plan, follow this loop:

1. Analyze — Read current state, understand what needs to happen
2. Execute — Perform the work (Claude or Codex based on assignment)
3. Reflect — Did this step succeed? Any issues? Update todo.md
4. Iterate — Move to next step or adjust plan if needed

Execution Rules

• One step at a time — Complete and verify each step before moving on
• Update todo.md immediately — After completing each step, change [ ] to [x] and update the Progress Log
• Reflection is mandatory — After each step, write a brief reflection in the Progress Log
• Plan changes require notification — If the plan needs to change, notify the user first
• Codex calls at checkpoints — When a step requires Codex, call the appropriate MCP tool:
  • clauxsync_review — Code review
  • clauxsync_debug — Debugging verification
  • clauxsync_refactor — Refactoring review
  • clauxsync_test — Test verification
  • clauxsync_brainstorm — Devil's advocate brainstorming

Status Markers

• [ ] — Pending
• [→] — In Progress (current step)
• [x] — Completed
• [!] — Blocked/Issue found
• [~] — Skipped (with reason)

Codex Interaction Pattern (Co-Commands Style)

When calling Codex at any checkpoint, follow the independent comparison pattern:

For Brainstorming

1. Claude brainstorms independently first
2. Spawn background subagent → call mcp__validate-plans-and-brainstorm-ideas__codex with brainstorming prompt
3. Do NOT read Codex results until Claude's own brainstorming is complete
4. Compare both sets of ideas, integrate the best

For Plan Validation (co-validate)

1. Claude creates/reviews the plan independently
2. Send plan to Codex: "You are a staff engineer reviewing this plan. Find critical issues, simplifications, or better approaches."
3. Compare both reviews
4. Accept or override each point with explanation

For Code Review / Debugging / Testing / Refactoring

1. Claude completes its analysis first
2. Send code + context to Codex via appropriate clauxsync tool
3. Compare findings
4. Claude synthesizes final result

Key Rule

Never read Codex's output before completing your own independent analysis. Two independent perspectives catch more issues than one influenced by the other.

Codex Response Treatment

Treat Codex responses as coming from a knowledgeable peer:

• Validate each suggestion independently
• Claude has final say as the lead engineer
• Accept good suggestions, override bad ones with explanation

Phase 3: Completion

When all steps are done:

1. Verify todo.md — all items must be [x] or [~] with reason
2. Remove skipped items or document why they were skipped
3. Present final summary to user with:
   • What was accomplished
   • Which AI handled what
   • Any issues encountered and how they were resolved
   • Codex feedback summary (if applicable)

Key Principles

• Plan before act — Never start coding without a plan
• Track everything — todo.md is the source of truth
• Reflect at every step — Catch issues early
• Two AIs validate — Claude creates, Codex validates
• User stays informed — Progress updates at each milestone