ChatGPT 网页版 / 任意 MCP 客户端        ← 操作层：派活、定约束、看结果
      │  OpenAI Secure MCP Tunnel / HTTPS / stdio
      ▼
chat2dsh（本仓库的轻量桥）              ← 桥接层：目录白名单、工具暴露、输出回传
      │  本机调用
      ▼
DeepSeek Harness (dsh headless)         ← 执行层：读改文件、跑命令、跑测试

# 1. 装 DeepSeek Harness（需要 Node ^22.19.0 或 >=24）
npm install -g @deepseek-ai/dsh

# 2. 装本仓库
git clone https://github.com/lvoliverrrr/chat2dsh.git
cd chat2dsh
npm install

# 3. 配一个 token 和白名单目录
cp .env.example .env
node -e "console.log(require('crypto').randomBytes(24).toString('hex'))"   # 填进 CHAT2DSH_TOKEN
# 再把 CHAT2DSH_ROOTS 改成你要让 agent 干活的项目目录

# 3.5 填 DSH 的 key
# 打开 .env，把 DEEPSEEK_API_KEY= 那行取消注释填上。不填也能跑，但 dsh_run 会返回
# MISSING_CREDENTIAL —— 桥是好的，只是 dsh 没有模型可用。

# 4. 起服务（HTTP 模式）
npm start
# [chat2dsh] v0.1.0 listening on http://127.0.0.1:7317/mcp
# [chat2dsh] dsh: 0.2.0-rc.2 (.../@deepseek-ai/dsh/lib/bin.js)
# [chat2dsh] roots: C:\Users\me\projects\demo
# [chat2dsh] auth: bearer token required

# 5. 自检
npm run smoke             # 握手 + tools/list + dsh_health + 越界拦截
npm run smoke -- --with-run   # 再加一次真实 dsh 任务（要花一次模型调用）

claude mcp add-json chat2dsh -s user '{"type":"stdio","command":"node","args":["<abs>/chat2dsh/src/server.mjs","--stdio","--root","<workspace>"]}'

{
  "mcpServers": {
    "chat2dsh": {
      "command": "node",
      "args": ["<abs>/chat2dsh/src/server.mjs", "--stdio", "--root", "<workspace>"],
      "env": { "DEEPSEEK_API_KEY": "..." }
    }
  }
}

node src/server.mjs --http --port 7317 --token <secret> --root <workspace>

POST http://127.0.0.1:7317/mcp
Authorization: Bearer <secret>

PASS  server boots and binds loopback — bridge 0.1.0, dsh 0.2.0-rc.2
PASS  dsh is reachable from the bridge process — resolved
PASS  unauthenticated request is rejected — HTTP 401
PASS  MCP handshake over Streamable HTTP — session 89ed50d7
PASS  tools/list returns the bridge surface — dsh_health, dsh_run
PASS  dsh_health reports a real dsh runtime — dsh 0.2.0-rc.2 @ .../@deepseek-ai/dsh/lib/bin.js
PASS  dsh_run refuses a cwd outside the roots — cwd "..." is outside the allowed roots
SKIP  dsh_run live task — pass --with-run to spend one real LLM call

7/7 checks passed

dsh: MISSING_CREDENTIAL: llm-deepseek: no API key for provider route "deepseek-official";
store DEEPSEEK_API_KEY through the credentials service (the web Models page writes it),
or export DEEPSEEK_API_KEY in the launching environment

$ curl https://<random>.trycloudflare.com/healthz
{"ok":true,"bridge":"0.1.0","dshVersion":"0.2.0-rc.2","authRequired":true}

$ curl -X POST https://<random>.trycloudflare.com/mcp -d '{"method":"tools/list",...}'
{"error":"unauthorized"}                       # HTTP 401，没带 token

$ curl -X POST https://<random>.trycloudflare.com/mcp \
    -H 'authorization: Bearer <token>' -d '{"method":"tools/list",...}'
event: message
data: {"result":{"tools":[{"name":"dsh_health",...},{"name":"dsh_run",...}]}}

# 临时域名，不用注册账号
cloudflared tunnel --url http://127.0.0.1:7317
# 它会打印 https://<random>.trycloudflare.com
# 你的 MCP 端点就是它后面加 /mcp

chat2dsh/
├── README.md        开发者视角：代码怎么跑、实测记录
├── AGENTS.md        给 AI 的入口 —— AI 进仓库先读这个
├── SECURITY.md      威胁模型、部署必做项、不提供的保证
├── llms.txt         给 LLM 的文档索引
├── src/
│   ├── server.mjs   MCP server：stdio + Streamable HTTP，两个工具
│   ├── dsh.mjs      调 dsh：入口解析、spawn、超时、截断
│   └── config.mjs   参数与环境变量、白名单校验
├── scripts/
│   └── smoke.mjs    端到端自检（真起进程、真握手）
├── docs/
│   ├── chatgpt-setup.md  路线 A：ChatGPT 网页版 + Codex Bridge 完整配置
│   └── agent-prompt.md   交给 AI 的部署任务书（让它带你一步步配）
├── .env.example
└── package.json