jsonlee12138/viberig · Archived

use-codex

Call Codex via its MCP server tools (codex + codex-reply) from any agent session.

First seen Jun 18, 2026

Installation

$ npx skills add jsonlee12138/viberig --skill use-codex

Summary

  • Call Codex via its MCP server tools (codex + codex-reply) from any agent session.
  • Use when delegating implementation tasks to Codex, running multi-turn Codex sessions, or orchestrating Codex as part of a multi-agent workflow.
  • Requires the Codex MCP server to be registered in the agent config.
  • Do NOT use inside a Codex session itself.

Stronger alternatives

This repository is archived — consider an actively maintained alternative.

Similar popular skills

Related neighbors and high-traction skills in the same topics — useful to compare before installing.

Also in this package

Other skills from jsonlee12138/viberig.

npx skills add jsonlee12138/viberig

Browse all from jsonlee12138/viberig

More details

Agent compatibility

Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.

Claude Code Not declared
Cursor Not declared
Codex Declared
GitHub Copilot Not declared
Windsurf Not declared
Gemini CLI Not declared
Cline Not declared
OpenCode Not declared

Repository health

Stars 4
License LICENSE
Default branch main
Open issues 0
Status Archived

Skill metadata

Parsed from SKILL.md frontmatter.

Declared agents codex

Package contents

Files included with this skill beyond the listing page.

  • skill md SKILL.md 2,564 B
  • docs SUMMARY.md 352 B

History

  1. First seen on skills.sh
  2. First recorded snapshot · 1 installs

SKILL.md

use-codex

通过已注册的 codex MCP server 调用 Codex 工具。MCP server 启动命令:codex mcp-server

两个核心工具

工具 用途
codex 启动新的 Codex 会话,返回 threadId
codex-reply 继续已有会话(需要 threadId

场景速查

场景 工具 关键参数
一次性任务(分析 / 问答) codex prompt, sandbox: read-only
写文件 / 实现代码 codex prompt, approval-policy: never, sandbox: workspace-write
多轮对话(追问 / 修改) codex-reply threadId, prompt
指定模型或配置 codex model, profile, config
切换工作目录 codex cwd

模型速查

说明
o3 最强推理,复杂任务
o4-mini 快速均衡,日常任务
(省略) 使用 Codex 本地配置默认模型

结果接收

Codex MCP 工具是请求-响应模式,必须 await 工具返回值再继续。响应结构:

# 正确:等待响应,提取内容和 threadId
response = await mcp.call("codex", { prompt: "...", sandbox: "read-only" })
text    = response.structuredContent.content   # 实际文本
threadId = response.structuredContent.threadId  # 多轮时保存此值

# 继续会话(多轮)
reply = await mcp.call("codex-reply", { threadId, prompt: "继续..." })

# 错误:fire-and-forget,结果和 threadId 丢失
mcp.call("codex", { prompt: "..." })

响应优先读 structuredContent;旧版客户端降级读 content[0].text。响应为空或工具报错时,停止并报告给调用方。

固定规则

  • MCP server 须已在 agent config 注册,key 为 codex-cli
  • 每次工具调用必须等待响应完成后再继续
  • 需要写文件时,必须同时设 approval-policy: neversandbox: workspace-write
  • 多轮会话必须保存首次响应的 structuredContent.threadId,用于后续 codex-reply
  • sandbox: danger-full-access 赋予完整系统权限,仅在用户明确授权时使用

完整参数说明见 [references/tool-specs.md](./references/tool-specs.md)。