MCP 的隐秘代价：21000 token 的工具定义，你的上下文窗口还剩多少？

{
  "name": "search_files",
  "description": "Search for files matching a pattern in the workspace. Returns file paths and metadata.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "pattern": {
        "type": "string",
        "description": "Glob pattern to match, e.g. '**/*.ts'"
      },
      "max_results": {
        "type": "integer",
        "description": "Maximum number of results to return",
        "default": 50
      },
      "include_hidden": {
        "type": "boolean",
        "description": "Whether to include hidden files",
        "default": false
      }
    },
    "required": ["pattern"]
  }
}

"""
最小 CLI 代理：模型通过执行 shell 命令完成任务，零 MCP 开销
依赖：openai >= 1.0, python-dotenv
运行前设置环境变量：OPENAI_API_KEY=sk-xxx
"""

import subprocess
import json
from openai import OpenAI

client = OpenAI()  # 从环境变量读取 API key

SYSTEM_PROMPT = """你是一个编码助手。你可以通过执行 shell 命令来完成任务。
可用指令格式：
<cmd>命令内容</cmd>

例如查看项目结构：<cmd>find . -type f -name "*.py" | head -20</cmd>
例如读取文件：<cmd>cat src/main.py</cmd>
例如运行测试：<cmd>pytest tests/ -q</cmd>

每次只输出一条命令。等待执行结果后再决定下一步。"""

def run_cli_agent(task: str, max_steps: int = 8):
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": task},
    ]

    for step in range(max_steps):
        resp = client.chat.completions.create(
            model="gpt-4o",
            messages=messages,
            temperature=0,
        )
        reply = resp.choices[0].message.content
        print(f"\n--- Step {step + 1} ---")
        print(f"模型输出:\n{reply}")

        # 提取 <cmd>...</cmd> 中的命令
        if "<cmd>" in reply and "</cmd>" in reply:
            cmd = reply.split("<cmd>")[1].split("</cmd>")[0].strip()
            print(f"执行命令: {cmd}")
            result = subprocess.run(
                cmd, shell=True, capture_output=True, text=True, timeout=30
            )
            output = result.stdout + result.stderr
            print(f"输出 (前500字):\n{output[:500]}")
            messages.append({"role": "assistant", "content": reply})
            messages.append({"role": "user", "content": f"命令输出:\n{output}"})

            if "任务完成" in reply or "DONE" in reply:
                print("\n✅ 任务完成")
                break
        else:
            print("模型未输出命令，任务结束")
            break

# 示例：让模型找到项目中的测试文件并运行
run_cli_agent("找到这个项目中所有 pytest 测试文件，然后运行它们，报告结果。")