Agent/examples/research/run.py

"""
浏览器调研示例 (交互增强版)

功能：
1. Agent 模式自动化调研
2. 手动接管：随时按 [Enter] 键暂停 Agent 并手动操作浏览器
3. 自动清理：无论成功或崩溃，均安全关闭浏览器进程
"""

import os
import sys
import asyncio
import logging
import re
import uuid
from pathlib import Path
from datetime import datetime
from argparse import Namespace

# 添加项目根目录到 Python 路径
sys.path.insert(0, str(Path(__file__).parent.parent.parent))

from dotenv import load_dotenv
load_dotenv()

# --- 日志配置 ---
logging.basicConfig(level=logging.WARNING)
logging.getLogger("agent.core.message_manager").setLevel(logging.INFO)
logging.getLogger("tools").setLevel(logging.INFO)

from agent.llm.prompts import SimplePrompt
from agent.core.runner import AgentRunner, RunConfig
from agent.trace import FileSystemTraceStore, Trace, Message
from agent.llm import create_openrouter_llm_call
from agent.tools.builtin.browser.baseClass import kill_browser_session

# ===== 全局交互控制 =====
pause_event = asyncio.Event()

async def listen_for_interrupt():
    """后台协程：监听标准输入，按下回车即触发暂停"""
    while True:
        # 在执行器中运行同步的 readline，避免阻塞事件循环
        await asyncio.get_event_loop().run_in_executor(None, sys.stdin.readline)
        if not pause_event.is_set():
            print("\n" + "!" * 40)
            print("🛑 检测到手动干预请求！")
            print("Agent 将在完成当前动作后暂停，请准备接管浏览器。")
            print("!" * 40 + "\n")
            pause_event.set()

# ===== 核心逻辑 =====

async def main():
    # 1. 环境准备
    base_dir = Path(__file__).parent
    project_root = base_dir.parent.parent
    trace_dir = project_root / ".trace"
    prompt_path = base_dir / "test.prompt"
    output_dir = base_dir / "output"
    output_dir.mkdir(exist_ok=True)

    print("=" * 60)
    print("🚀 交互式浏览器调研 Agent")
    print("👉 操作指南：")
    print("   - 运行中随时按下 [Enter] 键进入手动接管模式")
    print("   - 在浏览器完成操作后，点击页面上的 'Done' 或回车返回")
    print("=" * 60 + "\n")

    # 2. 加载任务
    prompt = SimplePrompt(prompt_path)
    system_prompt = prompt._messages.get("system", "")
    user_task = prompt._messages.get("user", "")
    # 默认使用 cheap 模型进行调研，如 gemini-3-flash-preview
    model_name = prompt.config.get('model', 'gemini-3-flash-preview')
    temperature = float(prompt.config.get('temperature', 0.3))

    messages = prompt.build_messages()

    # 3. 初始化 Runner
    # 注意：确保你的 openrouter 配置正确
    runner = AgentRunner(
        trace_store=FileSystemTraceStore(base_path=str(trace_dir)),
        llm_call=create_openrouter_llm_call(model=f"google/{model_name}"),
        skills_dir=None,
        debug=True 
    )

    # 4. 启动监听任务
    interrupt_task = asyncio.create_task(listen_for_interrupt())
    
    final_response = ""
    current_trace_id = None

    try:
        # 启动 Agent 迭代
        agent_stream = runner.run(
            messages=messages,
            config=RunConfig(
                system_prompt=system_prompt,
                model=f"google/{model_name}",
                temperature=temperature,
                max_iterations=30,
                name=user_task[:50],
            ),
        )

        async for item in agent_stream:
            # --- 检查手动暂停信号 ---
            if pause_event.is_set():
                print("\n" + "🛠️" * 20)
                print(">>> 人工接管模式激活 <<<")
                print("1. 请在浏览器窗口进行必要操作（登录、过验证码等）")
                print("2. 操作完成后，请在终端按 [Enter] 或在页面点击交互按钮继续")
                
                try:
                    # 调用内置的等待交互工具
                    await runner.tools.execute(
                        "browser_wait_for_user_action",
                        {"message": "人工干预中，请完成操作后恢复 Agent"},
                        uid="human_admin",
                        context={"runner": runner}
                    )
                except Exception as e:
                    print(f"⚠️ 交互工具调用失败: {e}")
                
                print(">>> 交互结束，交还控制权给 Agent <<<")
                print("🛠️" * 20 + "\n")
                pause_event.clear()

            # --- 正常处理 Agent 消息输出 ---
            if isinstance(item, Trace):
                current_trace_id = item.trace_id
                if item.status == "running":
                    print(f"[{datetime.now().strftime('%H:%M:%S')}] 🛰️ Trace 启动: {item.trace_id[:8]}")
                elif item.status == "completed":
                    print(f"\n✅ 任务圆满完成！Cost: ${item.total_cost:.4f}")

            elif isinstance(item, Message):
                if item.role == "assistant":
                    content = item.content
                    if isinstance(content, dict):
                        text = content.get("text", "")
                        tool_calls = content.get("tool_calls")
                        if text:
                            # 打印摘要，带点 Wit
                            print(f"\n🤖 Agent: {text[:200]}..." if len(text) > 200 else f"\n🤖 Agent: {text}")
                        if tool_calls:
                            for tc in tool_calls:
                                t_name = tc.get("function", {}).get("name", "unknown")
                                print(f"   🛠️  执行工具: {t_name}")
                
                elif item.role == "tool":
                    t_content = item.content
                    if isinstance(t_content, dict):
                        t_name = t_content.get("tool_name", "unknown")
                        print(f"   ✅ 工具返回: {t_name}")

    except Exception as e:
        print(f"\n🔥 发生严重错误: {e}")
        import traceback
        traceback.print_exc()

    finally:
        # 停止监听协程
        interrupt_task.cancel()
        
        # 5. 强制清理浏览器环境
        print("\n" + "·" * 40)
        print("🧹 正在执行环境清理...")
        try:
            await kill_browser_session()
            print("✨ 浏览器进程已安全终止。")
        except Exception as err:
            print(f"❌ 清理失败: {err}")
        print("·" * 40 + "\n")

    # 6. 结果展示
    if current_trace_id:
        print(f"🔍 任务 Trace ID: {current_trace_id}")
        print(f"📊 访问可视化面板查看详情。")

if __name__ == "__main__":
    try:
        asyncio.run(main())
    except KeyboardInterrupt:
        print("\n👋 收到退出信号，程序已停止。")