Claude Codex Loop — Productivity skill for Claude Code
Claude plans, Codex executes: a two-mode local workflow with verification, repair loops, and shared project memory.
How to install Claude Codex Loop
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open Yuwangzhi/claude-codex-loop and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Claude Codex Loop does
Claude plans, Codex executes: a two-mode local workflow with verification, repair loops, and shared project memory.
Alternatives in Productivity
- Reviews Plans — Review all active architecture plans relevant to the current task 15.6k ★
- Agentacct — See what your coding agents did and what it cost 665 ★
- ComfyUI Skills OpenClaw — Agent-friendly ComfyUI workflow skills for OpenClaw, Hermes Agent, Codex, and Claude Code; complementary to Co 395 ★
README
Claude–Codex Loop
[](LICENSE)
**只和 Claude 对话,让 Claude 定方案、Codex 干活,最后由 Claude 验收。也可以选择全部由 Claude 完成。**

概览图展示主要分工;具体分支见下方[模式选择](#两种模式)与[执行闭环](#闭环是怎样实现的)。[图片说明与生成提示词](docs/visual-assets.md)。
A local, two-mode workflow for Claude Code and Codex CLI: plan → execute → verify → repair → review → shared memory. Python standard library only; uses your existing CLI authentication.
适合 Claude 额度较少、Codex 额度较多,又希望保留 Claude 统筹体验的个人工作流。将大规模执行交给 Codex,可以减少 Claude 承担的执行工作;实际额度消耗取决于任务规模、上下文和返修次数,不承诺固定节省比例。
两种模式
| 模式 | Claude 的职责 | Codex 的职责 |
|---|---|---|
| 1:Claude 统筹 + Codex 执行 | 理解需求、形成计划、定义验收标准、检查结果 | 修改代码/文档、分析、运行测试、按反馈返修 |
| 2:仅 Claude | 独立完成规划、执行、验证和交付 | 不调用 |
每项新任务都说明当前模式,可回复 `1`、`2`、`一`、`二`、`模式1`、`模式二`,也接受“二吧”“就用1”等表达。未指定时沿用当前 session 的上次模式;新 session 则采用用户最近一次明确选择的偏好。从无历史选择时才询问并等待。同一任务内的补充、检查和返修继续当前模式,其他窗口的选择不会切换正在执行的任务。
flowchart TD
U["提出新任务"] --> S{"本次明确选择
否则沿用历史模式"}
S -->|无选择且无历史| Q["询问模式,等待回答"]
Q --> S
S -->|模式一| M1["Claude 规划,Codex 执行
独立检查后由 Claude 验收"]
S -->|模式二| M2["Claude 规划、执行与验证
不调用 Codex"]
M1 -->|验收通过| D["按需记录已验证知识
交付结果,保留模式"]
M2 -->|验证完成| D
D --> N["等待下一项任务
可选 1 或 2,不选则沿用"]
classDef neutral fill:#f5f7fa,stroke:#8293a6,color:#102a43;
classDef claude fill:#fff1e6,stroke:#d68a51,color:#102a43;
classDef loop fill:#e5f5f2,stroke:#438d87,color:#102a43;
classDef done fill:#eaf4e8,stroke:#6a9463,color:#102a43;
class U,S,Q,N neutral;
class M1 loop;
class M2 claude;
class D done;
模式一的返修、阻塞与停止分支见[执行闭环](#闭环是怎样实现的)。模式二不经过 CCL 的 Codex 任务执行器;两种模式都不会仅因时间流逝而自动开始下一项任务。
快速开始
1. 准备环境
- Python 3.9+、Git;macOS/Linux 使用 POSIX shell,Windows 的 Claude Bash 工具使用 Git Bash。
- 能独立工作的 Claude Code CLI 和 Codex CLI,并完成各自登录。
- Claude 客户端需支
Related Skills
Xdev Hive
Self-hosted control plane for teams running coding agents (Claude Code, Codex, Gemini CLI): task board, runs o
Relay Task Kernel
One handoff contract that Claude Code, Codex, Gemini CLI and Copilot all read. Zero-dependency CLI for shared
Workflow Orchestration
A Claude Code skill for disciplined task execution with planning, verification, and self-improvement loops
Eval
Run eval-driven development workflow. Execute verification loops and evaluate session outcomes against defined
Plan Execute Verify Repair
Run the autonomous-engineering operating loop on a task (plan→implement→verify→repair→report)
Subagent Driven Development
Claude Code skill: execute implementation plans with fresh subagents per task, two-stage review gates, and par
Related Agents
Bench Runner
Executes a11y skill benchmarks across hosted and local model families. Runs cloud/Codex/Ollama benchmark scrip
Dual Reviewer
Second-phase Codex code-review loop with Claude adjudication, after pr-reviewer in the branch review pass or m
Fullstack Agent
Team lead agent — researches, designs, specs, and plans. Creates an agent team, spawns teammates, coordinates