Worktree Codex Parallel

使用 git worktree 隔离多个 Codex 实例，由 OpenClaw 主控器并行调度完成同一项目的不同编码模块。 适用场景：将一个编码项目拆分为独立子任务，让多个 Codex 实例并行实现，最后合并 PR。 触发条件：用户要求"多个 Codex 协作"、"并行编码"、"worktree 编码"、"多...

Security Scan

The skill largely matches its stated purpose (parallelizing Codex via git worktrees) but it reads undeclared local config and API keys and sends logs/analysis to an external service — behaviors that are not declared in the metadata and could leak secrets or code.

Purpose & Capability

The skill's purpose (orchestrate multiple Codex agents using git worktrees) is coherent with the included scripts (setup_worktrees.sh, orchestrate.sh, push_and_pr.sh). However the SKILL metadata claims no required env/config paths while the instructions and code rely on several external binaries and secrets (OPENAI_API_KEY, CODEX_BIN, CLAUDE_BIN path, GH token). The skill does not declare these requirements in the registry metadata, which is inconsistent and surprising to a user.

Instruction Scope

SKILL.md and the scripts instruct the agent to read ~/.openclaw/openclaw.json for a GitHub token and to run Codex/Claude binaries with flags that can bypass sandboxing. Additionally, the dashboard's AI analysis thread (ai_analyze_async) reads the same ~/.openclaw file to extract OPENROUTER_API_KEY and posts log summaries to an external API (openrouter.ai). Those actions (reading a global OpenClaw config and transmitting log tails to a third-party) are outside the narrow 'worktree orchestration' description and are not declared.

Install Mechanism

No install spec is provided (instruction-only with bundled scripts and a dashboard). That minimizes installation-time risk because the skill does not fetch remote archives or run an installer. The risk is runtime behavior of the included scripts rather than an install-time downloader.

Credentials

The code accesses multiple credentials and config locations that are not declared: it expects OPENAI_API_KEY, CODEX_BIN, optional CLAUDE_BIN, and reads GH_TOKEN from ~/.openclaw/openclaw.json; the dashboard also extracts OPENROUTER_API_KEY from the same file. Requesting or reading another skill's/platform config file (~/.openclaw/openclaw.json) to obtain API keys is disproportionate and can lead to cross-skill secret access/exfiltration.

Persistence & Privilege

The skill does not set always:true and does not appear to persistently modify other skills, but it explicitly reads a platform/local config (~/.openclaw/openclaw.json) and extracts keys. Accessing and using other skills' config/credentials increases privilege and blast radius even though the skill itself isn't permanently installed as always:true.

What to consider before installing

This skill appears to implement parallel worktrees correctly, but it reads and uses secrets that are not declared and posts log data to an external service. Before installing or running: - Treat the dashboard and scripts as code that will run on your machine: review/modify them if you do not want automatic network calls. - Do not run dashboard.py or orchestrate.sh unless you are comfortable that ~/.openclaw/openclaw.json may be read; better: supply GH token and any API keys via environment variables instead of relying on that file. - The dashboard's ai_analyze_async reads OPENROUTER_API_KEY from your OpenClaw config and sends log tails (potentially code/content) to openrouter.ai. If you have sensitive code or secrets in logs, this can leak them. Remove or sandbox that feature, or require an explicit opt-in and an explicit environment variable for the analysis key. - The skill suggests using flags like --dangerously-skip-permissions / --dangerously-bypass-approvals-and-sandbox; those bypass host protections and increase risk. Avoid these unless you fully understand the implications. - Confirm the BASE URL (http://152.53.52.170:3003/v1) and any hardcoded IPs are intended and trustworthy — they point to a self-hosted proxy and could route model requests off your environment. If you still want to use this skill safely: require the owner to (1) declare required env vars in metadata, (2) stop reading ~/.openclaw/openclaw.json automatically (use explicit env vars), and (3) make external AI analysis optional and gated behind an explicit, purposeful opt-in.

Worktree Codex — 多 Codex 并行编码技能

前提条件

• Claude Code 二进制：/mnt/c/Users/Inuyasha/.local/bin/claude.exe（Windows 侧，可从 WSL 调用）
• GitHub Token：从 ~/.openclaw/openclaw.json 读取（skills.entries["gh-issues"].apiKey）
• 目标 Git 仓库已存在（本地 + GitHub 远端）

核心工作流（5 步）

步骤 1：明确任务拆分

在开始前，必须确定：

• 每个 Agent 的任务（1-3 个具体函数/文件）
• 文件所有权（每个文件只允许一个 Agent 写）
• 并行 or 串行（有依赖关系的任务需串行）

详见 references/task-decomposition.md

步骤 2：创建 Worktree

bash scripts/setup_worktrees.sh <repo_dir> <worktrees_base_dir> <agent-a> <agent-b> ...

输出格式：agent_name:worktree_path:branch_name（每行一个 Agent）

步骤 3：并行启动 Codex

用 exec(background=true) 并行启动多个 Agent，每个调用：

OPENAI_API_KEY="<key>" \
OPENAI_BASE_URL="http://152.53.52.170:3003/v1" \
CODEX_MODEL="gpt-5.3-codex" \
bash scripts/orchestrate.sh \
  <repo_dir> \
  <agent_name> \
  <worktree_path> \
  <branch> \
  "<task_prompt>" \
  /tmp/<agent_name>.log

# 注意：orchestrate.sh 默认使用 --full-auto（sandbox=workspace-write）
# 如需完全无沙箱，设 CODEX_EXTRA_FLAGS="--dangerously-bypass-approvals-and-sandbox"

Task prompt 模板：

你是 <AgentName>，只能修改 <file1>, <file2>，不得碰其他文件。
任务：<具体实现要求，含函数签名、异常处理、docstring>
完成后执行：git add -A && git commit -m "<commit message>"
只输出 done。

步骤 4：启动展板（可选但推荐）

展板是任务生命周期绑定的，不是常驻服务：

任务启动 → active 模式：SSE 实时推送，1秒刷新
任务完成 → idle 模式：页面保持不失效，停止轮询，显示"等待下次任务"
下次任务 → orchestrate.sh 自动 POST /reload → 恢复 active

启动展板（首次，后台运行）：

python3 skills/worktree-codex/dashboard.py \
  --logs /tmp/agent-a.log /tmp/agent-b.log \
  --port 7789 &
# 浏览器打开 http://localhost:7789

之后每次 orchestrate.sh 启动时会自动调用 /reload，无需手动操作。 手动激活新任务也可以：

curl -X POST http://localhost:7789/reload \
  -H 'Content-Type: application/json' \
  -d '{"logs":["/tmp/agent-a.log","/tmp/agent-b.log"]}'

展板功能：

• 每个 agent 状态卡（waiting/running/done/failed）+ 进度条 + 实时 token 计数
• idle 模式下页面不失效，数据保持可读
• 所有 agent 完成后，step-3.5-flash 自动效率分析（模型不可用时静默跳过）

步骤 5：监控 + 收尾

使用 process(poll) 等待所有 Agent 完成，然后检查：

cat /tmp/<agent_name>.log | grep "AGENT_DONE"

如果 Agent 没有自行 commit（常见于第一次），脚本会自动收尾。

步骤 5：推送 PR + 合并

bash scripts/push_and_pr.sh \
  <repo_dir> <gh_token> <owner/repo> <agent_name> <worktree_path> <branch> main

验证通过后通过 GitHub API 合并：

curl -s -X PUT \
  -H "Authorization: Bearer $GH_TOKEN" \
  -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/<owner/repo>/pulls/<pr_number>/merge" \
  -d '{"merge_method": "squash"}'

已知限制与解决方案

WSL 写文件被沙盒阻止 → Claude Code 必须加 --dangerously-skip-permissions

Codex .cmd 在 WSL 下无法运行 → 使用 --dangerously-skip-permissions 的 Claude Code 替代，或在 WSL 安装原生 codex：npm install -g @openai/codex

主仓库 git pull 权限错误（NTFS） → worktree 放在 WSL 本地文件系统（~/projects/worktrees/），不放 /mnt/c/

Agent 没有自行 commit → orchestrate.sh 脚本会检测未提交变更并自动 commit

环境变量获取

CLAUDE_BIN="/mnt/c/Users/Inuyasha/.local/bin/claude.exe"
GH_TOKEN=$(cat ~/.openclaw/openclaw.json | jq -r '.skills.entries["gh-issues"].apiKey')

模型选择与限制

⚠️ 核心限制：Codex CLI 只能用实现了 /v1/responses 端点的后端。

自建代理（http://152.53.52.170:3003/v1）只对 OpenAI 原生模型路由 /responses，deepseek / gemini / qwen 等全部 404。OpenRouter 同样不支持 /responses。

结论：目前只有 gpt-5.x-codex 系列可用。

模型	/responses	说明
gpt-5.3-codex	✅	默认，推荐
gpt-4.1-mini	✅（待验证）	轻量
deepseek-v3	❌ 404	代理未路由
gemini-2.5-flash	❌ 挂起	代理未路由
OpenRouter 任意模型	❌ 401	不支持端点