菜单
Delegate a task's implementation wholesale to OpenAI Codex (`codex-run work`) — the whole requirements document in one long autonomous run — keeping policy, orchestration, and acceptance in the main loop. Use at the Implement step of the Orchestrator loop, for any multi-part build.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run a PR's full review-fix cycle — wait for review and CI, fix failures, evaluate and reply to comments, then commit and push. Use whenever a PR was just opened or updated — the post-Ship phase of the Orchestrator loop.
Push the current branch and create a pull request with a title and description derived from the commit history. Use when the user asks to open a PR — the Ship step of the Orchestrator loop, always user-triggered.
Cross-model second opinion from OpenAI Codex — `codex-run advise --context` reviews the current session, or pipe a brief. Triggers on "ask codex", "codex review", "second opinion", or the "Cross-model review" rules.
Define requirements before implementation — clarify constraints and design policy, then write a requirements spec and task breakdown for the Implementer to a gitignored plan file. Use at the Plan step of the Orchestrator loop, or whenever a request's scope needs pinning down before building. Never implements; the plan file is its only write.
Root-cause investigation — enumerate competing hypotheses, verify each with evidence, suggest a fix but never apply it. Use when an error, failing test, or anomaly needs diagnosing before any fix is written; implementation hands off to the Implementer.
Create, improve, and evaluate skills. Use when the user wants to create a skill from scratch, edit or optimize an existing one, run evals or benchmarks on it, or optimize its description for better triggering accuracy.
| name | codex-work |
| description | Delegate a task's implementation wholesale to OpenAI Codex (`codex-run work`) — the whole requirements document in one long autonomous run — keeping policy, orchestration, and acceptance in the main loop. Use at the Implement step of the Orchestrator loop, for any multi-part build. |
| argument-hint | [task-description] |
Dispatch the Implementer: one autonomous Codex run through the whole plan, committing as it goes. Flags: codex-run --help.
codex-run work -C <repo> --log /tmp/codex-work-<slug>.jsonl < brief.md # one run per plan; unique log path
codex-run work --resume <id> -C <repo> < followup.md # send fixes, finish, or steer (id in the banner)
pkill -INT -f 'codex exec .*<repo>' # interrupt a live run before resuming it
/define plan path plus only what the repo can't tell Codex: in-session decisions, branch, scope bounds. Never restate ~/.codex/AGENTS.md or the wrapper's preamble.$HOME needs a git worktree; the wrapper refuses a writable run there.--log. Don't edit the checkout; give parallel runs separate worktrees.--resume with the correction. History survives, so write "drop X, do Y instead", not a re-brief; you lose only the unfinished turn.--resume. Other non-zero = turn failed; check git status for partial edits. Never resume a session another run is still writing.