OpenClaw Claude Codex Workflow
ReviewAudited by ClawScan on May 1, 2026.
Overview
This skill appears to be a disclosed multi-model coding workflow, but it can let external Claude/Codex/Gemini tools change your project and handle project context.
Install this only if you intend to use Claude, Codex, and Gemini CLIs for code work. Use a clean branch or worktree, run `--dry-run` before execution, review diffs and `.claude/` artifacts before committing or attaching them to tickets, and use scoped API keys without embedding secrets in prompts or logs.
Findings (5)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
A bad prompt, bad plan, or model mistake could change project files in the working tree.
The skill intentionally delegates local code-writing/execution tasks to Codex and Gemini. This is central to the workflow and disclosed, but it is still meaningful local mutation authority.
任何写操作由 Codex/Gemini 在本地工作树执行 ... Codex：开 PTY (`codex exec --pty`)；严格按照 plan 执行命令 ... Gemini：在独立工作树或 `frontend/` 目录执行
Run with `--dry-run` first, keep work on a clean branch/worktree, inspect `git diff` before committing, and require explicit human approval for destructive, deploy, database, or push actions.
The external CLIs may use your provider accounts, quotas, billing, and access rights.
The workflow expects provider API keys to be available to subprocesses. This is expected for the integrated services, but the registry metadata does not declare credentials.
设置凭据：`CLAUDE_API_KEY`、`CODEX_API_KEY`、`GOOGLE_API_KEY` 等；在 shell profile 中 `export`，以便脚本继承。
Use least-privilege/project-specific keys, avoid storing secrets in shared shell profiles, and ensure keys are not written into `.claude/` artifacts or commits.
Installing or using the wrong CLI binary could give an unexpected program access to prompts, repository files, and API keys.
The skill relies on external CLIs installed outside the bundle and does not pin versions. This is normal for this type of workflow, but users must verify the tools' provenance.
安装 CLI 并验证 `claude --version`、`codex --version`、`gemini --version` 可用（Claude Code CLI、Codex CLI、Gemini CLI 官方安装方式自行遵循厂商文档）。
Install CLIs only from official vendor sources, verify versions, and avoid PATH entries that could shadow `claude`, `codex`, or `gemini`.
Sensitive business context or mistaken instructions could persist in the repository and be reused by later agents.
The workflow stores user requirements, ticket/spec context, plans, and logs as persistent artifacts that can later influence model actions and may be committed.
将用户需求/OPSX ticket 粘贴进 `.claude/context.md` ... 阶段日志与 `.claude/` 工件需要纳入 commit，以便追溯。
Redact secrets and confidential details before writing or committing `.claude/` files, and review generated plans/logs before using them as instructions for later phases.
Project requirements, code context, and plans may be sent to several external model providers depending on the configured CLIs.
The orchestrator passes prompts, context, and plan files across multiple model/provider CLIs. This is the intended architecture, but it creates multi-provider data-flow boundaries.
cmd+=(--context "$CONTEXT_FILE") ... "$CODEX_CLI" exec --prompt "$BACKEND_PROMPT" --plan "$PLAN_FILE" ... "$GEMINI_CLI" run --prompt "$FRONTEND_PROMPT" --plan "$PLAN_FILE"
Limit context files to information appropriate for those providers, confirm each provider's data-use policy, and avoid including secrets or customer data unless explicitly approved.