src/app.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
AI大模型辩论系统Web版本
基于FastAPI的Web应用，提供图形用户界面的辩论系统
"""

import os
import sys
import json
import logging
from datetime import datetime, timezone, timedelta
from typing import Optional, Dict, Any
import importlib.util
import asyncio
import socket
import uuid
import math
import re
from fastapi import Query
from fastapi.responses import HTMLResponse, StreamingResponse
from urllib.parse import quote # 新增：导入URL编码工具

# 在代码开头强制设置终端编码为UTF-8（仅在Windows执行）
if os.name == 'nt':
    os.system('chcp 65001 > nul')

# 获取当前脚本文件所在目录的绝对路径
current_script_dir = os.path.dirname(os.path.abspath(__file__))

# 项目根目录 (即 '20250907_大模型辩论' 目录, 是 'src' 的上一级)
project_root = os.path.dirname(current_script_dir)

# 定义数据输入、输出和日志目录
DATA_DIR = os.path.join(project_root, 'data')
OUTPUT_DIR = os.path.join(project_root, 'output')
LOGS_DIR = os.path.join(project_root, 'logs')

# 确保目录存在
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(LOGS_DIR, exist_ok=True)

# 配置日志
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
    handlers=[
        logging.FileHandler(os.path.join(LOGS_DIR, '对话系统Web日志.log'), encoding='utf-8'),
        logging.StreamHandler(sys.stdout)
    ]
)
logger = logging.getLogger(__name__)

# 捕获警告并记录到日志
logging.captureWarnings(True)

# 导入FastAPI相关模块
try:
    from fastapi import FastAPI, Request, WebSocket, WebSocketDisconnect
    from fastapi.responses import HTMLResponse
    from fastapi.staticfiles import StaticFiles
    from fastapi.templating import Jinja2Templates
    import uvicorn
    from uvicorn.middleware.proxy_headers import ProxyHeadersMiddleware
    logger.info("FastAPI模块导入成功")
except ImportError as e:
    logger.error(f"导入FastAPI模块失败: {str(e)}")
    print("请安装FastAPI: pip install fastapi uvicorn")
    sys.exit(1)

# 导入自定义模块
try:
    # 动态导入模型接口模块
    model_interface_path = os.path.join(current_script_dir, "model_interface.py")
    model_interface_spec = importlib.util.spec_from_file_location("model_interface", model_interface_path)
    model_interface = importlib.util.module_from_spec(model_interface_spec)
    model_interface_spec.loader.exec_module(model_interface)
    
    # 动态导入对话控制器模块
    debate_controller_path = os.path.join(current_script_dir, "debate_controller.py")
    debate_controller_spec = importlib.util.spec_from_file_location("debate_controller", debate_controller_path)
    debate_controller = importlib.util.module_from_spec(debate_controller_spec)
    debate_controller_spec.loader.exec_module(debate_controller)
    
    # 从模块中获取需要的类
    ModelManager = model_interface.ModelManager
    ConversationMessage = debate_controller.ConversationMessage
    ConversationSession = debate_controller.ConversationSession
    
except Exception as e:
    logger.error(f"导入模块失败: {str(e)}")
    sys.exit(1)

# 创建FastAPI应用
app = FastAPI(title="AI大模型对话系统", description="基于FastAPI的AI大模型对话系统Web版本")

# 添加中间件以处理反向代理头信息
app.add_middleware(ProxyHeadersMiddleware, trusted_hosts="*")

# 设置静态文件目录
static_dir = os.path.join(project_root, "static")
app.mount("/static", StaticFiles(directory=static_dir), name="static")

# 设置模板目录
templates_dir = os.path.join(project_root, "templates")
templates = Jinja2Templates(directory=templates_dir)

# 补回：连接管理器
class ConnectionManager:
    """连接管理器，负责管理所有WebSocket连接及其状态"""
    def __init__(self):
        self.active_connections: Dict[WebSocket, Dict[str, Any]] = {}

    async def connect(self, websocket: WebSocket):
        await websocket.accept()
        self.active_connections[websocket] = {"session": None, "task": None, "judge_task": None, "conv_id": None, "judge_id": None}
        logger.info(f"新连接建立: {websocket.client}. 当前总连接数: {len(self.active_connections)}")

    def disconnect(self, websocket: WebSocket):
        if websocket in self.active_connections:
            task = self.active_connections[websocket].get("task")
            if task and not task.done():
                task.cancel()
                logger.info(f"连接 {websocket.client} 的对话任务已被取消。")
            jtask = self.active_connections[websocket].get("judge_task")
            if jtask and not jtask.done():
                jtask.cancel()
                logger.info(f"连接 {websocket.client} 的评判任务已被取消。")
            del self.active_connections[websocket]
            logger.info(f"连接断开: {websocket.client}. 当前总连接数: {len(self.active_connections)}")

    def get_session(self, websocket: WebSocket) -> Optional[ConversationSession]:
        return self.active_connections.get(websocket, {}).get("session")

    def get_task(self, websocket: WebSocket) -> Optional[asyncio.Task]:
        return self.active_connections.get(websocket, {}).get("task")

    def get_judge_task(self, websocket: WebSocket) -> Optional[asyncio.Task]:
        return self.active_connections.get(websocket, {}).get("judge_task")

    def set_conv_id(self, websocket: WebSocket, conv_id: str):
        if websocket in self.active_connections:
            self.active_connections[websocket]["conv_id"] = conv_id

    def get_conv_id(self, websocket: WebSocket) -> Optional[str]:
        return self.active_connections.get(websocket, {}).get("conv_id")

    def set_judge_id(self, websocket: WebSocket, judge_id: str):
        if websocket in self.active_connections:
            self.active_connections[websocket]["judge_id"] = judge_id

    def get_current_judge_id(self, websocket: WebSocket) -> Optional[str]:
        return self.active_connections.get(websocket, {}).get("judge_id")

    def set_conversation(self, websocket: WebSocket, session: ConversationSession, task: asyncio.Task):
        if websocket in self.active_connections:
            self.active_connections[websocket]["session"] = session
            self.active_connections[websocket]["task"] = task

    def set_judge_task(self, websocket: WebSocket, task: asyncio.Task):
        if websocket in self.active_connections:
            self.active_connections[websocket]["judge_task"] = task

    def clear_conversation(self, websocket: WebSocket):
        if websocket in self.active_connections:
            self.active_connections[websocket]["session"] = None
            self.active_connections[websocket]["task"] = None

    def cancel_all(self, websocket: WebSocket):
        if websocket in self.active_connections:
            t = self.active_connections[websocket].get("task")
            if t and not t.done():
                t.cancel()
            jt = self.active_connections[websocket].get("judge_task")
            if jt and not jt.done():
                jt.cancel()
            # 刷新流ID，确保旧任务输出被丢弃
            self.active_connections[websocket]["conv_id"] = str(uuid.uuid4())
            self.active_connections[websocket]["judge_id"] = str(uuid.uuid4())


def _get_api_key_from_env() -> str:
    """从多种环境变量名中获取API Key"""
    candidate_keys = [
        "MODELSCOPE_API_KEY", "MODELSCOPE_TOKEN", "MS_API_KEY",
        "MS_TOKEN", "API_KEY"
    ]
    for k in candidate_keys:
        v = os.environ.get(k)
        if v:
            logger.info(f"已从环境变量 {k} 读取到API Key（不显示具体值）")
            return v
    logger.warning("未在环境变量中找到ModelScope API Key，请在Space Secrets中设置 MODELSCOPE_API_KEY")
    return ""

# 实例化连接管理器和模型管理器
manager = ConnectionManager()
model_manager = ModelManager(_get_api_key_from_env())
# 新增：内存中的会话备份，避免必须写磁盘
recent_sessions: Dict[str, ConversationSession] = {}
# 新增：内存缓存最近一次评判结果
recent_judges: Dict[str, str] = {}

@app.on_event("startup")
async def startup_event():
    """应用启动时执行的事件"""
    os.makedirs(static_dir, exist_ok=True)
    os.makedirs(templates_dir, exist_ok=True)
    create_templates()

@app.get("/", response_class=HTMLResponse)
async def read_root(request: Request):
    """主页路由，返回Web界面"""
    return templates.TemplateResponse("index.html", {
        "request": request,
        "title": "AI大模型对话系统"
    })

@app.get("/api/status")
async def get_status():
    """获取系统状态"""
    return {
        "status": "running",
        "timestamp": datetime.now().isoformat(),
        "models": ["glm45", "deepseek_v31", "qwen", "qwen_instruct"],
        "active_connections": len(manager.active_connections),
        "has_api_key": bool(_get_api_key_from_env())
    }

@app.websocket("/ws")
async def websocket_endpoint(websocket: WebSocket):
    """WebSocket端点，用于实时通信"""
    await manager.connect(websocket)
    try:
        while True:
            data = await websocket.receive_text()
            await handle_websocket_message(websocket, data)
    except WebSocketDisconnect:
        manager.disconnect(websocket)
    except Exception as e:
        logger.error(f"WebSocket处理错误: {str(e)}", exc_info=True)
        manager.disconnect(websocket)

async def handle_websocket_message(websocket: WebSocket, message: str):
    """处理WebSocket消息"""
    try:
        data = json.loads(message)
        action = data.get("action")
        
        if action == "start_conversation":
            # 强行停止正在进行的对话与评判
            manager.cancel_all(websocket)
            # 启动新对话（后台任务）
            new_task = asyncio.create_task(start_conversation(websocket, data))
            # 任务将在 start_conversation 内部注册到 manager 中
        elif action == "stop_conversation":
            await stop_conversation(websocket)
        elif action == "judge_debate":
            # 评判改为后台任务，并记录以便可取消
            jtask = asyncio.create_task(judge_debate(websocket, data))
            manager.set_judge_task(websocket, jtask)
        elif action == "summarize_collaboration":
            # 协作总结改为后台任务，并记录以便可取消
            stask = asyncio.create_task(summarize_collaboration(websocket, data))
            manager.set_judge_task(websocket, stask)
        else:
            await websocket.send_text(json.dumps({
                "type": "error",
                "message": f"未知操作: {action}"
            }))
    except json.JSONDecodeError:
        await websocket.send_text(json.dumps({
            "type": "error",
            "message": "无效的JSON格式"
        }))
    except Exception as e:
        logger.error(f"处理WebSocket消息时出错: {str(e)}", exc_info=True)
        await websocket.send_text(json.dumps({
            "type": "error",
            "message": f"处理消息时出错: {str(e)}"
        }))

async def start_conversation(websocket: WebSocket, data: dict):
    """开始一场独立的对话"""
    loop = asyncio.get_event_loop()
    session = None
    try:
        topic = data.get("topic", "真与善谁更重要？")
        mode = data.get("mode", "debate")
        rounds = int(data.get("rounds", 3))
        pro_model_name = data.get("pro_model", "deepseek_v31")
        con_model_name = data.get("con_model", "qwen_instruct")
        initial_prompt = data.get("initial_prompt", "").strip()
        initial_prompt_mode = data.get("initial_prompt_mode", "append")
        save_enabled = bool(data.get("save_records", False))
        
        session = ConversationSession(
            topic=topic, mode=mode, max_rounds=rounds, 
            pro_model=pro_model_name, con_model=con_model_name, 
            initial_prompt=initial_prompt, initial_prompt_mode=initial_prompt_mode
        )
        # 将新创建的session和当前任务关联到这个websocket连接
        manager.set_conversation(websocket, session, asyncio.current_task())
        # 刷新对话流ID
        conv_stream_id = str(uuid.uuid4())
        manager.set_conv_id(websocket, conv_stream_id)

        await websocket.send_text(json.dumps({ 
            "type": "conversation_started", "message": "对话已开始", 
            "topic": topic, "mode": mode, "rounds": rounds, 
            "pro_model": pro_model_name, "con_model": con_model_name,
            "debate_id": session.debate_id
        }))
        
        model_a = model_manager.get_model(pro_model_name)
        model_b = model_manager.get_model(con_model_name)
        session.start_time = datetime.now()
        speakers = [(pro_model_name, model_a), (con_model_name, model_b)]
        
        for i in range(rounds * 2):
            round_num = (i // 2) + 1
            if i % 2 == 0:
                session.current_round = round_num
                await websocket.send_text(json.dumps({ "type": "round_info", "message": f"—— 第（{round_num}/{rounds}）轮 ——" }))

            speaker_name, speaker_model = speakers[i % 2]
            is_pro = (i % 2 == 0)
            role = ("正方" if mode == 'debate' else "AI 1") if is_pro else ("反方" if mode == 'debate' else "AI 2")
            
            await websocket.send_text(json.dumps({ "type": "model_speaking", "model": speaker_name, "role": role }))
            prompt = session.generate_prompt(speaker_name)
            
            response_content = ""
            def stream_callback(content):
                nonlocal response_content
                # 丢弃旧流输出
                if manager.get_conv_id(websocket) != conv_stream_id:
                    return
                response_content += content
                asyncio.run_coroutine_threadsafe(
                    websocket.send_text(json.dumps({"type": "stream_content", "content": content})), loop
                )

            await loop.run_in_executor(None, speaker_model.chat_stream, session.get_messages_for_model(speaker_name) + [{"role": "user", "content": prompt}], stream_callback)
            await websocket.send_text(json.dumps({"type": "stream_end", "model": speaker_name}))

            session.add_message(ConversationMessage("user", prompt, "system"))
            session.add_message(ConversationMessage("assistant", response_content, speaker_name))
            save_conversation_record(session, save_enabled)
            # 新增：实时更新内存备份
            recent_sessions[session.debate_id] = session

        if session.is_active:
            session.end_time = datetime.now()
            session.is_active = False
            await websocket.send_text(json.dumps({ "type": "conversation_ended", "message": "=== 对话结束 ===" }))
            save_conversation_record(session, save_enabled)
            # 新增：结束时更新内存备份
            recent_sessions[session.debate_id] = session
            logger.info(f"对话 {session.debate_id} 正常结束。")
        
    except asyncio.CancelledError:
        if session:
            logger.info(f"对话任务 {session.debate_id} 被取消。")
            session.is_active = False
            session.end_time = datetime.now()
            save_conversation_record(session, bool(data.get("save_records", False)))
            # 新增：取消时也更新内存备份
            recent_sessions[session.debate_id] = session
            await websocket.send_text(json.dumps({
                "type": "conversation_stopped", "message": "对话已停止"
            }))
        raise
    except Exception as e:
        logger.error(f"对话过程中出错: {str(e)}", exc_info=True)
        await websocket.send_text(json.dumps({ "type": "error", "message": f"对话过程中出错: {str(e)}" }))
    finally:
        manager.clear_conversation(websocket)
        logger.info("连接的对话会话清理完毕。")

async def judge_debate(websocket: WebSocket, data: dict):
    """评判一场指定的对话"""
    loop = asyncio.get_event_loop()
    try:
        judge_model_name = data.get("judge_model", "qwen_instruct")
        debate_id = data.get("debate_id")
        save_enabled = bool(data.get("save_records", False))
        if not debate_id:
            await websocket.send_text(json.dumps({"type": "error", "message": "缺少对话ID无法评判。"}))
            return
        
        file_path = os.path.join(OUTPUT_DIR, "对话记录", f"{debate_id}.json")
        session_to_judge = None
        if os.path.exists(file_path):
            session_to_judge = ConversationSession.load_from_file(file_path)
        else:
            # 新增：优先使用内存备份，避免必须写磁盘
            session_to_judge = recent_sessions.get(debate_id)
            if not session_to_judge:
                await websocket.send_text(json.dumps({"type": "error", "message": f"找不到对话记录: {debate_id}（内存与磁盘均不存在）"}))
                return # 修复：此 return 必须在 if 块内部
        
        judge_model = model_manager.get_model(judge_model_name)
        
        assistant_messages = [msg for msg in session_to_judge.messages if msg.role == 'assistant']
        # 构造带轮次的实录
        total_msgs = len(assistant_messages)
        actual_rounds = math.ceil(total_msgs / 2) if total_msgs > 0 else 0
        transcript_parts = []
        total_rounds = session_to_judge.max_rounds
        for r in range(actual_rounds):
            transcript_parts.append(f"—— 第（{r+1}/{total_rounds}）轮 ——")
            pro_idx = r * 2
            con_idx = r * 2 + 1
            if pro_idx < total_msgs:
                pro_model = session_to_judge.pro_model
                transcript_parts.append(f"正方 ({pro_model}): {assistant_messages[pro_idx].content}")
            if con_idx < total_msgs:
                con_model = session_to_judge.con_model
                transcript_parts.append(f"反方 ({con_model}): {assistant_messages[con_idx].content}")
        debate_transcript = "\n\n".join(transcript_parts)
        
        if session_to_judge.mode == 'debate':
            judge_prompt = (
                f"你是一位专业的辩论评审。请基于以下对话实录，对双方的辩论表现进行专业评判。\n\n"
                f"对话模式：辩论\n"
                f"辩论话题：{session_to_judge.topic}\n"
                f"正方（AI 1）：{session_to_judge.pro_model}\n"
                f"反方（AI 2）：{session_to_judge.con_model}\n\n"
                f"实际轮次：{actual_rounds}（禁止杜撰额外轮次）\n\n"
                f"对话实录：\n{debate_transcript}\n\n"
                f"请严格按照以下步骤进行分析并输出：\n"
                f"1.  **结论先行（1-2句）**：直接指出哪一方更胜一筹，并给出最关键的1-2条理由。\n"
                f"2.  **维度对比表（Markdown 表格）**：从至少六个维度对双方评分/评述：立场清晰度、论据扎实度、反驳力度、逻辑结构、证据引用/事实性、聚焦度（针对性）。最后一列写明该维度的优势方。\n"
                f"3.  **证据引用**：逐点引用原文并标注\"第X轮-正/反\"来支撑判定，禁止引用不存在的轮次。\n"
                f"4.  **改进建议**：分别给正反双方各2-3条可执行的改进建议。"
            )
        else:
            judge_prompt = (
                    f"你是一位专业的协作评审。请基于以下对话实录，评估两位 AI 的协作质量与产出。\n\n"
                    f"对话模式：协作讨论\n"
                    f"协作任务：{session_to_judge.topic}\n"
                    f"AI 1：{session_to_judge.pro_model}\n"
                    f"AI 2：{session_to_judge.con_model}\n\n"
                    f"实际轮次：{actual_rounds}（禁止杜撰额外轮次）\n\n"
                    f"对话实录：\n{debate_transcript}\n\n"
                    f"请严格按照以下步骤进行分析并输出：\n"
                    f"1.  **总体评估（1-2句）**：先给出任务完成度与协作有效性的总体判断。\n"
                    f"2.  **协作维度表（Markdown 表格）**：从至少六个维度评述：目标对齐、信息共享/互补、方案可行性、风险识别、推进计划（时间/里程碑）、个人贡献度。最后一列说明哪一方贡献更关键。\n"
                    f"3.  **行动计划**：给出一份精炼的下一步行动清单（里程碑+负责人+时间节点）。引用原文时请标注\"第X轮-参与者\"。\n"
                    f"4.  **改进建议**：指出影响协作效率的关键瓶颈，并给出2-3条可落地的改进建议。"
            )
        
        # 刷新评判流ID
        judge_stream_id = str(uuid.uuid4())
        manager.set_judge_id(websocket, judge_stream_id)
        await websocket.send_text(json.dumps({"type": "judge_started", "model": judge_model_name}))
        
        response_content = ""
        def stream_callback(content):
            nonlocal response_content
            # 丢弃旧流输出
            if manager.get_current_judge_id(websocket) != judge_stream_id:
                return
            response_content += content
            asyncio.run_coroutine_threadsafe(
                websocket.send_text(json.dumps({"type": "judge_stream_content", "content": content})), loop
            )
        
        await loop.run_in_executor(None, judge_model.chat_stream, [{"role": "user", "content": judge_prompt}], stream_callback)
        await websocket.send_text(json.dumps({"type": "judge_stream_end", "model": judge_model_name}))
        logger.info(f"模型 {judge_model_name} 已完成评判。")

        # 可选：保存评判结果
        if save_enabled:
            try:
                judge_dir = os.path.join(OUTPUT_DIR, "评判记录")
                os.makedirs(judge_dir, exist_ok=True)
                judge_file = os.path.join(judge_dir, f"{debate_id}_judge_{judge_model_name}.md")
                with open(judge_file, "w", encoding="utf-8") as jf:
                    jf.write(f"# 评判结果\n\n对话ID: {debate_id}\n\n评判模型: {judge_model_name}\n\n---\n\n")
                    jf.write(response_content)
                logger.info(f"评判结果已保存: {judge_file}")
            except Exception as e:
                logger.error(f"保存评判结果失败: {str(e)}")
        # 新增：写入内存缓存，便于导出
        recent_judges[debate_id] = response_content

    except Exception as e:
        logger.error(f"评判过程中出错: {str(e)}", exc_info=True)
        await websocket.send_text(json.dumps({"type": "error", "message": f"评判过程中出错: {str(e)}"}))

async def summarize_collaboration(websocket: WebSocket, data: dict):
    """总结协作任务的对话内容"""
    loop = asyncio.get_event_loop()
    try:
        summary_model_name = data.get("summary_model", "qwen_instruct")
        debate_id = data.get("debate_id")
        save_enabled = bool(data.get("save_records", False))
        if not debate_id:
            await websocket.send_text(json.dumps({"type": "error", "message": "缺少对话ID无法总结。"}))
            return
        
        file_path = os.path.join(OUTPUT_DIR, "对话记录", f"{debate_id}.json")
        session_to_summarize = None
        if os.path.exists(file_path):
            session_to_summarize = ConversationSession.load_from_file(file_path)
        else:
            # 优先使用内存备份，避免必须写磁盘
            session_to_summarize = recent_sessions.get(debate_id)
            if not session_to_summarize:
                await websocket.send_text(json.dumps({"type": "error", "message": f"找不到对话记录: {debate_id}（内存与磁盘均不存在）"}))
                return
        
        summary_model = model_manager.get_model(summary_model_name)
        
        assistant_messages = [msg for msg in session_to_summarize.messages if msg.role == 'assistant']
        # 构造带轮次的实录
        total_msgs = len(assistant_messages)
        actual_rounds = math.ceil(total_msgs / 2) if total_msgs > 0 else 0
        transcript_parts = []
        total_rounds = session_to_summarize.max_rounds
        for r in range(actual_rounds):
            transcript_parts.append(f"—— 第（{r+1}/{total_rounds}）轮 ——")
            pro_idx = r * 2
            con_idx = r * 2 + 1
            if pro_idx < total_msgs:
                pro_model = session_to_summarize.pro_model
                transcript_parts.append(f"AI 1 ({pro_model}): {assistant_messages[pro_idx].content}")
            if con_idx < total_msgs:
                con_model = session_to_summarize.con_model
                transcript_parts.append(f"AI 2 ({con_model}): {assistant_messages[con_idx].content}")
        collaboration_transcript = "\n\n".join(transcript_parts)
        
        summary_prompt = (
            f"你是一位专业的会议记录员和内容总结专家。请基于以下协作对话实录，对两位AI的讨论内容进行全面总结。\n\n"
            f"对话模式：协作讨论\n"
            f"协作任务：{session_to_summarize.topic}\n"
            f"AI 1：{session_to_summarize.pro_model}\n"
            f"AI 2：{session_to_summarize.con_model}\n\n"
            f"实际轮次：{actual_rounds}（禁止杜撰额外轮次）\n\n"
            f"对话实录：\n{collaboration_transcript}\n\n"
            f"请严格按照以下步骤进行分析并输出：\n"
            f"1.  **任务概述**：简要描述协作任务的目标和背景。\n"
            f"2.  **主要观点**：总结两位AI在讨论中提出的主要观点和想法，按主题分类。\n"
            f"3.  **达成共识**：列出两位AI在讨论中达成的一致意见和共识。\n"
            f"4.  **分歧点**：指出两位AI在讨论中存在的不同意见或分歧。\n"
            f"5.  **最终结论**：总结协作讨论的最终结论或成果。\n"
            f"6.  **后续建议**：提出基于当前讨论的后续行动建议或需要进一步探讨的问题。"
        )
        
        # 刷新总结流ID
        summary_stream_id = str(uuid.uuid4())
        manager.set_judge_id(websocket, summary_stream_id)
        await websocket.send_text(json.dumps({"type": "summary_started", "model": summary_model_name}))
        
        response_content = ""
        def stream_callback(content):
            nonlocal response_content
            # 丢弃旧流输出
            if manager.get_current_judge_id(websocket) != summary_stream_id:
                return
            response_content += content
            asyncio.run_coroutine_threadsafe(
                websocket.send_text(json.dumps({"type": "summary_stream_content", "content": content})), loop
            )
        
        await loop.run_in_executor(None, summary_model.chat_stream, [{"role": "user", "content": summary_prompt}], stream_callback)
        await websocket.send_text(json.dumps({"type": "summary_stream_end", "model": summary_model_name}))
        logger.info(f"模型 {summary_model_name} 已完成协作总结。")

        # 可选：保存总结结果
        if save_enabled:
            try:
                summary_dir = os.path.join(OUTPUT_DIR, "总结记录")
                os.makedirs(summary_dir, exist_ok=True)
                summary_file = os.path.join(summary_dir, f"{debate_id}_summary_{summary_model_name}.md")
                with open(summary_file, "w", encoding="utf-8") as sf:
                    sf.write(f"# 协作总结\n\n对话ID: {debate_id}\n\n总结模型: {summary_model_name}\n\n---\n\n")
                    sf.write(response_content)
                logger.info(f"协作总结已保存: {summary_file}")
            except Exception as e:
                logger.error(f"保存协作总结失败: {str(e)}")
        # 写入内存缓存，便于导出
        recent_judges[debate_id] = response_content

    except Exception as e:
        logger.error(f"总结过程中出错: {str(e)}", exc_info=True)
        await websocket.send_text(json.dumps({"type": "error", "message": f"总结过程中出错: {str(e)}"}))

async def stop_conversation(websocket: WebSocket):
    """停止当前连接的对话"""
    task = manager.get_task(websocket)
    if task and not task.done():
        task.cancel()
        logger.info("发送取消请求到对话任务。")
    else:
        logger.warning("请求停止对话，但没有活动的任务。")
        await websocket.send_text(json.dumps({"type": "info", "message": "没有正在进行的对话可供停止。"}))

def save_conversation_record(session: ConversationSession, save_enabled: bool):
    """保存指定的对话记录（按需）"""
    if not save_enabled:
        return
    if session:
        try:
            output_dir = os.path.join(OUTPUT_DIR, "对话记录")
            os.makedirs(output_dir, exist_ok=True)
            file_path = os.path.join(output_dir, f"{session.debate_id}.json")
            session.save_to_file(file_path)
            logger.info(f"对话记录已保存: {file_path}")
        except Exception as e:
            logger.error(f"保存对话记录时出错: {str(e)}")

def create_templates():
    """创建HTML模板文件"""
    template_path = os.path.join(templates_dir, "index.html")
    # 始终覆盖模板文件，确保其与代码同步
    index_html = """
<!DOCTYPE html>
<html lang="zh-CN">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>AI大模型对话系统</title>
    <link rel="preconnect" href="https://fonts.googleapis.com">
    <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
    <link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=Noto+Sans+SC:wght@400;500;700&display=swap" rel="stylesheet">
    <link rel="stylesheet" href="{{ url_for('static', path='css/style.css') }}?v=20250908a">
    <script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
    <script src="https://cdn.jsdelivr.net/npm/dompurify/dist/purify.min.js"></script>

</head>
<body>
    <div class="container">
        <div class="header">
            <div class="header-title">
                <h1>AI大模型对话系统</h1>
                <p>观看两个AI大模型实时对话</p>
            </div>
            <div class="header-controls">
                <select id="exportFormat">
                    <option value="md" selected>Markdown (.md)</option>
                    <option value="json">JSON (.json)</option>
                    <option value="txt">文本 (.txt)</option>
                </select>
                <button id="exportAllBtn" disabled>导出记录</button>
            </div>
        </div>
        
        <div class="main-layout">
            <div class="sidebar">
                <h3 class="panel-title">设置</h3>
                <div class="control-group">
                    <label>对话模式</label>
                    <div class="radio-group">
                        <input type="radio" id="modeDebate" name="mode" value="debate" checked>
                        <label for="modeDebate">辩论</label>
                        <input type="radio" id="modeDiscussion" name="mode" value="discussion">
                        <label for="modeDiscussion">协作讨论</label>
                    </div>
                </div>
                <div class="control-group"><label for="topic">对话任务/话题</label><input type="text" id="topic" value="真与善谁更重要？"></div>
                <div class="control-group">
                    <label for="initialPrompt">自定义初始提示 (可选)</label>
                    <textarea id="initialPrompt" rows="6" placeholder="默认提示示例：'你将作为正方，就[话题]进行辩论...'。你可以在此输入额外指示（默认追加），或选择覆盖默认提示。"></textarea>
                </div>
                <div class="control-group prompt-mode-group">
                    <label>提示词模式</label>
                    <div class="radio-group">
                        <input type="radio" id="promptAppend" name="promptMode" value="append" checked>
                        <label for="promptAppend">追加</label>
                        <input type="radio" id="promptOverride" name="promptMode" value="override">
                        <label for="promptOverride">覆盖</label>
                    </div>
                </div>
                <div class="control-group"><label for="rounds">轮数</label><input type="number" id="rounds" min="1" max="10" value="3"></div>
                <div class="control-group">
                    <label for="proModel">AI 1 (正方)</label>
                    <select id="proModel">
                        <option value="deepseek_v31" selected>deepseek-ai/DeepSeek-V3.1</option>
                        <option value="qwen_instruct">Qwen/Qwen3-235B-Instruct</option>
                        <option value="qwen">Qwen/Qwen3-235B-Thinking</option>
                        <option value="glm45">ZhipuAI/GLM-4.5</option>
                    </select>
                </div>
                <div class="control-group">
                    <label for="conModel">AI 2 (反方)</label>
                    <select id="conModel">
                        <option value="deepseek_v31">deepseek-ai/DeepSeek-V3.1</option>
                        <option value="qwen_instruct" selected>Qwen/Qwen3-235B-Instruct</option>
                        <option value="qwen">Qwen/Qwen3-235B-Thinking</option>
                        <option value="glm45">ZhipuAI/GLM-4.5</option>
                    </select>
                </div>
                <div class="controls">
                    <button id="startBtn" disabled>开始对话</button>
                    <button id="stopBtn" disabled>停止对话</button>
                </div>
            </div>
            <div class="chat-area">
                <div class="conversation-wrapper">
                    <h3 class="panel-title">对话区</h3>
                    <div id="output" class="output-container"></div>
                </div>
                <div id="judge-section" class="judge-section">
                    <h3 class="panel-title" id="judgeSectionTitle">评判区</h3>
                    <div class="judge-controls">
                        <label for="judgeModel">选择总结模型</label>
                        <select id="judgeModel">
                            <option value="deepseek_v31">deepseek-ai/DeepSeek-V3.1</option>
                            <option value="qwen_instruct" selected>Qwen/Qwen3-235B-Instruct</option>
                            <option value="qwen">Qwen/Qwen3-235B-Thinking</option>
                            <option value="glm45">ZhipuAI/GLM-4.5</option>
                        </select>
                        <button id="judgeBtn" disabled>评判双方辩论表现</button>
                        <button id="summaryBtn" disabled style="display:none;">总结对话</button>
                    </div>
                    <div id="judge-output" class="output-container judge-output"></div>
                </div>
            </div>
        </div>
    </div>
    <script src="{{ url_for('static', path='js/script.js') }}?v=20250908a"></script>
</body>
</html>"""
    with open(template_path, "w", encoding="utf-8") as f:
        f.write(index_html)
    logger.info("Web应用模板文件 'index.html' 已被强制更新。")


def get_local_ip():
    """获取本机局域网IP地址"""
    try:
        with socket.socket(socket.AF_INET, socket.SOCK_DGRAM) as s:
            # 连接到一个公共DNS服务器的IP（不会真的发送数据）
            s.connect(("8.8.8.8", 80))
            return s.getsockname()[0]
    except Exception:
        return "127.0.0.1"  # 如果获取失败，返回本地回环地址


if __name__ == "__main__":
    # 本地运行时，FastAPI的启动事件也会触发，所以模板创建是安全的
    
    # 智能端口切换：优先使用环境变量PORT，否则默认为8000
    port = int(os.environ.get("PORT", 8000))
    local_ip = get_local_ip()
    
    logger.info("="*50)
    logger.info("AI大模型对话系统已启动")
    logger.info(f"  - 本机访问: http://localhost:{port}")
    logger.info(f"  - 局域网访问: http://{local_ip}:{port}")
    logger.info("="*50)
    
    uvicorn.run(app, host="0.0.0.0", port=port) 

@app.get("/api/export")
async def export_records(debate_id: str = Query(...), format: str = Query("md")):
    """导出指定对话的记录（对话+最近一次评判），不落盘，直接下载。"""
    fmt = (format or "md").lower()
    session = recent_sessions.get(debate_id)
    if not session:
        # 兜底：尝试磁盘
        file_path = os.path.join(OUTPUT_DIR, "对话记录", f"{debate_id}.json")
        if os.path.exists(file_path):
            session = ConversationSession.load_from_file(file_path)
    if not session:
        return HTMLResponse(content=f"<pre>找不到对话记录: {debate_id}</pre>", status_code=404)

    judge_md = recent_judges.get(debate_id, "")
    # 仅保留双方模型的回答（assistant），不导出提示词/用户消息
    assistant_messages = [m for m in session.messages if m.role == 'assistant']
    # 导出时间（优先用会话开始时间），统一转换为东八区，精确到分钟
    base_dt = session.start_time or datetime.utcnow()
    if base_dt.tzinfo is None:
        beijing_dt = base_dt + timedelta(hours=8)
    else:
        beijing_dt = base_dt.astimezone(timezone(timedelta(hours=8)))
    ts = beijing_dt.strftime('%Y-%m-%d %H:%M (UTC+8)')
    filename_ts = beijing_dt.strftime('%Y%m%d_%H%M%S')
    
    # 清理话题作文文件名安全的部分
    topic = session.topic or "未命名对话" # 修复：为None或空话题提供默认值
    sanitized_topic = re.sub(r'[\\/*?:"<>|]', "_", topic).replace(" ", "_")
    sanitized_topic = (sanitized_topic[:50] + '...') if len(sanitized_topic) > 50 else sanitized_topic

    if fmt == 'json':
        payload = {
            'debate_id': debate_id,
            'topic': session.topic,
            'export_time': ts,
            'messages': [
                {
                    'role': m.role,
                    'content': m.content,
                    'model': m.model_name,
                    'round': (idx // 2) + 1,
                    'round_total': session.max_rounds,
                    'round_label': f"{(idx // 2) + 1}/{session.max_rounds}"
                } for idx, m in enumerate(assistant_messages)
            ],
            'judge_markdown': judge_md
        }
        content = json.dumps(payload, ensure_ascii=False, indent=2)
        filename = f"{filename_ts}_{sanitized_topic}.json"
        media = "application/json; charset=utf-8"
    elif fmt == 'txt':
        parts = [f"对话ID: {debate_id}", f"话题: {session.topic}", f"导出时间: {ts}", "---"]
        total_rounds = session.max_rounds
        for i, m in enumerate(assistant_messages):
            round_no = (i // 2) + 1
            # 每轮开始插入分隔
            if i % 2 == 0:
                parts.append(f"—— 第（{round_no}/{total_rounds}）轮 ——")
            # 推断角色与模型
            is_pro = (i % 2 == 0)
            model_name = session.pro_model if is_pro else session.con_model
            role_cn = "正方" if is_pro else "反方"
            parts.append(f"{role_cn} ({model_name}):\n{m.content}\n")
        if judge_md:
            parts.append("\n==== 评判结果 ====")
            parts.append(md_to_text(judge_md))
        content = "\n".join(parts)
        filename = f"{filename_ts}_{sanitized_topic}.txt"
        media = "text/plain; charset=utf-8"
    else:  # md
        parts = ["# 对话记录", f"对话ID: {debate_id}", f"话题: {session.topic}", f"导出时间: {ts}", ""]
        total_rounds = session.max_rounds
        for i, m in enumerate(assistant_messages):
            round_no = (i // 2) + 1
            if i % 2 == 0:
                parts.append(f"### 第（{round_no}/{total_rounds}）轮")
                parts.append("")
            is_pro = (i % 2 == 0)
            model_name = session.pro_model if is_pro else session.con_model
            role_cn = "正方" if is_pro else "反方"
            parts.append(f"**{role_cn} ({model_name})**:")
            parts.append("")
            parts.append(m.content)
            parts.append("")
        if judge_md:
            parts += ["---", "# 评判结果", "", judge_md]
        content = "\n".join(parts)
        filename = f"{filename_ts}_{sanitized_topic}.md"
        media = "text/markdown; charset=utf-8"

    encoded_filename = quote(filename)
    headers = {"Content-Disposition": f"attachment; filename*=UTF-8''{encoded_filename}"}
    return StreamingResponse(iter([content.encode('utf-8')]), media_type=media, headers=headers) 

# 简单将Markdown转为纯文本，用于TXT导出
def md_to_text(md: str) -> str:
    if not md:
        return ""
    text = md
    # 代码块三反引号去除
    text = re.sub(r"```[\s\S]*?```", lambda m: re.sub(r"^```.*\n|\n```$", "", m.group(0)), text)
    # 行内代码
    text = re.sub(r"`([^`]+)`", r"\1", text)
    # 粗体/斜体
    text = re.sub(r"\*\*([^*]+)\*\*", r"\1", text)
    text = re.sub(r"__([^_]+)__", r"\1", text)
    text = re.sub(r"\*([^*]+)\*", r"\1", text)
    text = re.sub(r"_([^_]+)_", r"\1", text)
    # 链接与图片
    text = re.sub(r"!\[([^\]]*)\]\([^)]*\)", r"\1", text)
    text = re.sub(r"\[([^\]]+)\]\([^)]*\)", r"\1", text)
    # 标题、引用、水平线
    text = re.sub(r"^>\s*", "", text, flags=re.MULTILINE)
    text = re.sub(r"^#{1,6}\s*", "", text, flags=re.MULTILINE)
    text = re.sub(r"^\s*-{3,}\s*$", "", text, flags=re.MULTILINE)
    # 表格竖线与分隔行
    text = re.sub(r"^\|?\s*-+\s*(\|\s*-+\s*)+\|?\s*$", "", text, flags=re.MULTILINE)
    text = text.replace("|", " |")
    return text