一键导入
codex-agent
Delegate coding tasks to Codex CLI. Use when executing code changes, reviews, debugging, or parallelizing tasks via Codex.
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。
菜单
Delegate coding tasks to Codex CLI. Use when executing code changes, reviews, debugging, or parallelizing tasks via Codex.
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
Linear issue management (search, create, update, close, comment) with arbitrary GraphQL queries.
Use when asked to review or audit an implementation plan.
新功能的规划→spec→issue 生命周期。当用户说"讨论下"、"写 spec"、"开 issue"、"怎么做"、"设计一下"、"拆任务"、"评估方案"、或描述新功能需求时触发。即使用户只是随口提到一个想法或改进方向,只要涉及"该怎么做"的问题,都应触发此 skill 来结构化思考。
Release workflow for skills repo. Use when releasing, bumping version, creating tags, or publishing changes. Trigger on "release", "发布", "new version", "bump version".
Multi-lens adversarial review via Codex. Trigger on "review", "审查", "check the code".
Write HANDOFF.md before ending a complex session. Use when wrapping up, switching tasks, or when context is getting large and a fresh session would help.
| name | codex-agent |
| description | Delegate coding tasks to Codex CLI. Use when executing code changes, reviews, debugging, or parallelizing tasks via Codex. |
You are the orchestrator. Codex is the executor. You plan, decompose, and review. Codex writes code, refactors, and fixes bugs.
Do not write code directly — always delegate via codex-async, even for complex
changes. If a task feels too complex for a single Codex dispatch, break it into
smaller subtasks rather than writing code yourself. Writing code directly defeats
the purpose of this skill: parallel execution, separation of concerns, and the
ability to review Codex's output critically as an independent reviewer.
Agent Execution Instructions:
{baseDir}{baseDir}/scripts/codex-asyncchmod +x {baseDir}/scripts/codex-async once{baseDir} in this document with the actual pathScript Reference:
| Script | Purpose |
|---|---|
scripts/codex-async | Async task manager wrapping codex exec |
Verify: {baseDir}/scripts/codex-async help
Analyze the request. Decompose into discrete steps. Identify dependencies — steps without dependencies run in parallel.
# Single task
{baseDir}/scripts/codex-async run --cd "$PWD" "Implement JWT auth in src/auth.ts"
# Parallel — launch multiple in one turn
{baseDir}/scripts/codex-async run --cd "$PWD" "Implement auth module"
{baseDir}/scripts/codex-async run --cd "$PWD" "Write tests for utils"
{baseDir}/scripts/codex-async run --cd "$PWD" "Refactor error handling"
{baseDir}/scripts/codex-async wait <id1> <id2> --timeout 300
{baseDir}/scripts/codex-async result <id>
Read results critically. If not satisfied:
{baseDir}/scripts/codex-async resume <id> "Fix null input handling"After Codex completes, verify before reporting done:
git diff for unintended changesDefinition of done: tests pass, no unintended side effects, plan fulfilled.
Run {baseDir}/scripts/codex-async help for full CLI reference.
| Option | Description |
|---|---|
--model | Model: gpt-5.3-codex (default, optimized for code), gpt-5.4 (general reasoning) |
--reasoning | Reasoning effort: high (default), xhigh, medium, low |
--sandbox | Sandbox: workspace-write (default), read-only, danger-full-access |
--cd | Working directory for the task |
Use defaults unless you have a reason to change. For simple mechanical tasks
(renaming, formatting), --reasoning medium saves tokens. Use --sandbox read-only
for review-only tasks, danger-full-access only when the task needs to install
dependencies or access the network.
{baseDir}/scripts/codex-async kill <id> and retry