Merge pull request #130 from aiastia-dockerhub/pr-to-upstream

fix: MCP插件TimeoutError修复 + 多项Bug修复和性能优化

Dockerfile

@@ -60,10 +60,14 @@ RUN apt-get update && apt-get install -y \
 # 复制后端依赖文件
 COPY backend/requirements.txt ./
 
-# 安装 Python 依赖（包含 torch，避免单独安装造成重复层）
+# 安装 Python 依赖
+# 先安装 torch CPU版本（~200MB vs 完整版~2GB，节省90%下载时间）
+# 对于embedding场景，CPU版本完全够用
 RUN if [ "$USE_CN_MIRROR" = "true" ]; then \
+        pip install --no-cache-dir torch==2.8.0 --index-url https://mirrors.aliyun.com/pypi/simple/ --extra-index-url https://download.pytorch.org/whl/cpu && \
         pip install --no-cache-dir -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/; \
     else \
+        pip install --no-cache-dir torch==2.8.0 --index-url https://download.pytorch.org/whl/cpu && \
         pip install --no-cache-dir -r requirements.txt; \
     fi
 

backend/app/api/careers.py

@@ -7,7 +7,7 @@ import json
 from typing import AsyncGenerator
 
 from app.database import get_db
-from app.utils.sse_response import SSEResponse, create_sse_response, WizardProgressTracker
+from app.utils.sse_response import SSEResponse, create_sse_response, WizardProgressTracker, wrap_stream_with_heartbeat, HEARTBEAT
 from app.models.career import Career, CharacterCareer
 from app.models.character import Character
 from app.models.project import Project
@@ -25,6 +25,7 @@ from app.schemas.career import (
     CareerStage
 )
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.logger import get_logger
 from app.api.settings import get_user_ai_service
 from app.api.common import verify_project_access
@@ -155,14 +156,10 @@ async def create_career(
         raise HTTPException(status_code=500, detail=f"创建职业失败: {str(e)}")
 
 
-@router.get("/generate-system", summary="AI生成新职业（增量式，流式）")
+@router.post("/generate-system", summary="AI生成新职业（增量式，流式）")
 async def generate_career_system(
-    project_id: str,
-    main_career_count: int = 3,
-    sub_career_count: int = 6,
-    user_requirements: str = "",
-    enable_mcp: bool = False,
-    http_request: Request = None,
+    request_data: CareerGenerateRequest,
+    http_request: Request,
     db: AsyncSession = Depends(get_db),
     user_ai_service: AIService = Depends(get_user_ai_service)
 ):
@@ -176,6 +173,10 @@ async def generate_career_system(
         try:
             # 验证用户权限和项目是否存在
             user_id = getattr(http_request.state, 'user_id', None)
+            project_id = request_data.project_id
+            main_career_count = request_data.main_career_count
+            sub_career_count = request_data.sub_career_count
+            user_requirements = request_data.user_requirements
             project = await verify_project_access(project_id, user_id, db)
             
             yield await tracker.start()
@@ -316,7 +317,15 @@ async def generate_career_system(
                 chunk_count = 0
                 estimated_total = max(3000, len(prompt) * 8)
                 
-                async for chunk in user_ai_service.generate_text_stream(prompt=prompt):
+                async for chunk in wrap_stream_with_heartbeat(
+                    user_ai_service.generate_text_stream(prompt=prompt),
+                    heartbeat_interval=15.0
+                ):
+                    # 心跳哨兵：发送心跳保活，不混入AI响应
+                    if chunk is HEARTBEAT:
+                        yield await tracker.heartbeat()
+                        continue
+
                     chunk_count += 1
                     ai_response += chunk
                     
@@ -345,7 +354,7 @@ async def generate_career_system(
             # 清洗并解析JSON
             try:
                 cleaned_response = user_ai_service._clean_json_response(ai_response)
-                career_data = json.loads(cleaned_response)
+                career_data = loads_json(cleaned_response)
                 logger.info(f"✅ 职业体系JSON解析成功")
             except json.JSONDecodeError as e:
                 logger.error(f"❌ 职业体系JSON解析失败: {e}")

backend/app/api/characters.py

@@ -7,7 +7,7 @@ import json
 from typing import AsyncGenerator
 
 from app.database import get_db
-from app.utils.sse_response import SSEResponse, create_sse_response, WizardProgressTracker
+from app.utils.sse_response import SSEResponse, create_sse_response, WizardProgressTracker, wrap_stream_with_heartbeat, HEARTBEAT
 from app.models.character import Character
 from app.models.project import Project
 from app.models.generation_history import GenerationHistory
@@ -20,6 +20,7 @@ from app.schemas.character import (
     CharacterGenerateRequest
 )
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.services.prompt_service import prompt_service, PromptService
 from app.services.import_export_service import ImportExportService
 from app.schemas.import_export import CharactersExportRequest, CharactersImportResult
@@ -947,10 +948,18 @@ async def generate_character_stream(
                 logger.info(f"🎯 开始生成角色（流式模式）...")
                 yield await tracker.generating(0, estimated_total, "开始生成角色...")
                 
-                async for chunk in user_ai_service.generate_text_stream(
+                async for chunk in wrap_stream_with_heartbeat(
+                    user_ai_service.generate_text_stream(
                         prompt=prompt,
                         tool_choice="required",
+                    ),
+                    heartbeat_interval=15.0
                 ):
+                    # 心跳哨兵：发送心跳保活，不混入AI响应
+                    if chunk is HEARTBEAT:
+                        yield await tracker.heartbeat()
+                        continue
+
                     # chunk 现在可能是 dict 或 str，提取 content 字段
                     if isinstance(chunk, dict):
                         content = chunk.get("content", "")
@@ -987,7 +996,7 @@ async def generate_character_stream(
             # ✅ 使用统一的 JSON 清洗方法
             try:
                 cleaned_response = user_ai_service._clean_json_response(ai_response)
-                character_data = json.loads(cleaned_response)
+                character_data = loads_json(cleaned_response)
                 logger.info(f"✅ 角色JSON解析成功")
             except json.JSONDecodeError as e:
                 logger.error(f"❌ 角色JSON解析失败: {e}")

backend/app/api/inspiration.py

@@ -6,6 +6,7 @@ import json
 
 from app.database import get_db
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.api.settings import get_user_ai_service
 from app.services.prompt_service import PromptService
 from app.logger import get_logger
@@ -166,7 +167,7 @@ async def generate_options(
                 # 使用统一的JSON清洗方法
                 cleaned_content = ai_service._clean_json_response(content)
                 
-                result = json.loads(cleaned_content)
+                result = loads_json(cleaned_content)
                 
                 # 校验返回格式
                 is_valid, error_msg = validate_options_response(result, step)
@@ -343,7 +344,7 @@ async def refine_options(
             # 解析JSON
             try:
                 cleaned_content = ai_service._clean_json_response(content)
-                result = json.loads(cleaned_content)
+                result = loads_json(cleaned_content)
                 
                 # 校验返回格式
                 is_valid, error_msg = validate_options_response(result, step)
@@ -466,7 +467,7 @@ async def quick_generate(
             # 使用统一的JSON清洗方法
             cleaned_content = ai_service._clean_json_response(content)
             
-            result = json.loads(cleaned_content)
+            result = loads_json(cleaned_content)
             
             # 合并用户已提供的信息（用户输入优先）
             final_result = {

backend/app/api/mcp_plugins.py

@@ -54,14 +54,24 @@ async def _register_plugin_background(
     plugin_type: str,
     server_url: str,
     headers: Optional[dict],
-    config: Optional[dict]
+    config: Optional[dict],
+    max_retries: int = 2,
+    retry_delay: float = 3.0
 ):
     """
-    后台任务：注册MCP插件并更新数据库状态
+    后台任务：注册MCP插件并更新数据库状态（带重试）
 
-    在独立的任务中执行MCP连接，避免阻塞请求处理
+    在独立的任务中执行MCP连接，避免阻塞请求处理。
+    连接失败时会自动重试，提高对临时网络问题的容错性。
     """
+    last_error = None
+    
+    for attempt in range(max_retries + 1):
         try:
+            if attempt > 0:
+                logger.info(f"后台注册MCP插件重试 ({attempt}/{max_retries}): {plugin_name}")
+                await asyncio.sleep(retry_delay)
+            else:
                 logger.info(f"后台注册MCP插件: {plugin_name}")
 
             if plugin_type in HTTP_PLUGIN_TYPES and server_url:
@@ -77,29 +87,29 @@ async def _register_plugin_background(
             else:
                 success = False
 
-        # 更新数据库状态
+            if success:
+                # 更新数据库状态为active
                 engine = await get_engine(user_id)
                 AsyncSessionLocal = async_sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
-
                 async with AsyncSessionLocal() as db:
                     stmt = (
                         update(MCPPlugin)
                         .where(MCPPlugin.user_id == user_id, MCPPlugin.plugin_name == plugin_name)
-                .values(
-                    status="active" if success else "error",
-                    last_error=None if success else "连接失败"
-                )
+                        .values(status="active", last_error=None)
                     )
                     await db.execute(stmt)
                     await db.commit()
-
-        if success:
                 logger.info(f"后台注册MCP插件成功: {plugin_name}")
+                return
             else:
-            logger.warning(f"后台注册MCP插件失败: {plugin_name}")
+                last_error = "连接失败"
                 
         except Exception as e:
-        logger.error(f"后台注册MCP插件异常: {plugin_name}, 错误: {e}")
+            last_error = str(e)
+            logger.warning(f"后台注册MCP插件异常 (尝试 {attempt + 1}/{max_retries + 1}): {plugin_name}, 错误: {e}")
+    
+    # 所有重试都失败，更新数据库状态为error
+    logger.error(f"后台注册MCP插件最终失败 (已重试{max_retries}次): {plugin_name}, 错误: {last_error}")
     try:
         engine = await get_engine(user_id)
         AsyncSessionLocal = async_sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
@@ -107,7 +117,7 @@ async def _register_plugin_background(
             stmt = (
                 update(MCPPlugin)
                 .where(MCPPlugin.user_id == user_id, MCPPlugin.plugin_name == plugin_name)
-                    .values(status="error", last_error=str(e))
+                .values(status="error", last_error=str(last_error)[:500] if last_error else "连接失败")
             )
             await db.execute(stmt)
             await db.commit()
@@ -215,22 +225,26 @@ async def create_plugin(
         **plugin_data
     )
     
+    # 如果启用，设为pending状态等待后台连接
+    if plugin.enabled:
+        plugin.status = "pending"
+    
     db.add(plugin)
     await db.commit()
     await db.refresh(plugin)
     
-    # 如果启用，注册到统一门面
+    # 如果启用，后台注册到统一门面（避免MCP操作阻塞导致超时）
     if plugin.enabled:
-        success = await _register_plugin_to_facade(plugin, user.user_id)
-        if success:
-            plugin.status = "active"
-        else:
-            plugin.status = "error"
-            plugin.last_error = "加载失败"
-        await db.commit()
-        await db.refresh(plugin)
+        asyncio.create_task(_register_plugin_background(
+            user_id=user.user_id,
+            plugin_name=plugin.plugin_name,
+            plugin_type=plugin.plugin_type,
+            server_url=plugin.server_url,
+            headers=plugin.headers,
+            config=plugin.config
+        ))
     
-    logger.info(f"用户 {user.user_id} 创建插件: {plugin.plugin_name}")
+    logger.info(f"用户 {user.user_id} 创建插件: {plugin.plugin_name}（MCP注册在后台执行）")
     return plugin
 
 
@@ -438,15 +452,29 @@ async def update_plugin(
     for key, value in update_data.items():
         setattr(plugin, key, value)
     
+    # 如果启用，设为pending状态等待后台连接
+    if plugin.enabled:
+        plugin.status = "pending"
+        plugin.last_error = None
+    
     await db.commit()
     await db.refresh(plugin)
     
-    # 如果插件已启用，重新注册
+    # 如果插件已启用，后台重新注册MCP连接
     if plugin.enabled:
-        await mcp_client.unregister(user.user_id, plugin.plugin_name)
-        await _register_plugin_to_facade(plugin, user.user_id)
+        # 先后台注销旧连接
+        asyncio.create_task(_unregister_plugin_safe(user.user_id, plugin.plugin_name))
+        # 再后台注册新连接
+        asyncio.create_task(_register_plugin_background(
+            user_id=user.user_id,
+            plugin_name=plugin.plugin_name,
+            plugin_type=plugin.plugin_type,
+            server_url=plugin.server_url,
+            headers=plugin.headers,
+            config=plugin.config
+        ))
     
-    logger.info(f"用户 {user.user_id} 更新插件: {plugin.plugin_name}")
+    logger.info(f"用户 {user.user_id} 更新插件: {plugin.plugin_name}（MCP操作在后台执行）")
     return plugin
 
 
@@ -470,15 +498,19 @@ async def delete_plugin(
     if not plugin:
         raise HTTPException(status_code=404, detail="插件不存在")
     
-    # 从统一门面注销
-    await mcp_client.unregister(user.user_id, plugin.plugin_name)
+    # 保存插件信息用于后台注销
+    plugin_name = plugin.plugin_name
+    user_id = user.user_id
     
-    # 删除数据库记录
+    # 先删除数据库记录
     await db.delete(plugin)
     await db.commit()
     
-    logger.info(f"用户 {user.user_id} 删除插件: {plugin.plugin_name}")
-    return {"message": "插件已删除", "plugin_name": plugin.plugin_name}
+    # 后台从统一门面注销（避免MCP操作阻塞导致超时）
+    asyncio.create_task(_unregister_plugin_safe(user_id, plugin_name))
+    
+    logger.info(f"用户 {user.user_id} 删除插件: {plugin_name}（MCP注销在后台执行）")
+    return {"message": "插件已删除", "plugin_name": plugin_name}
 
 
 @router.post("/{plugin_id}/toggle", response_model=MCPPluginResponse)
@@ -490,6 +522,10 @@ async def toggle_plugin(
 ):
     """
     启用或禁用插件
+    
+    启用时：先更新数据库状态为pending，再通过后台任务注册MCP连接，
+    避免长时间持有数据库会话导致超时。
+    禁用时：先更新数据库状态，再通过后台任务注销MCP连接。
     """
     result = await db.execute(
         select(MCPPlugin).where(
@@ -509,51 +545,35 @@ async def toggle_plugin(
     headers = plugin.headers
     config = plugin.config
     
-    # 先更新数据库状态
+    # 更新数据库状态
     plugin.enabled = enabled
-    if not enabled:
+    if enabled:
+        # 启用时先设为pending状态，等待后台MCP连接完成
+        plugin.status = "pending"
+        plugin.last_error = None
+    else:
         plugin.status = "inactive"
     
     await db.commit()
     await db.refresh(plugin)
     
-    # 数据库操作完成后，再进行MCP操作
+    # 数据库操作完成后，通过后台任务进行MCP操作（避免长时间持有数据库会话）
     if enabled:
-        # 启用：注册到统一门面
-        try:
-            if plugin_type in HTTP_PLUGIN_TYPES and server_url:
-                server_url = _validate_mcp_server_url(plugin_type, server_url)
-                success = await mcp_client.register(MCPPluginConfig(
+        # 启用：后台注册到统一门面
+        asyncio.create_task(_register_plugin_background(
             user_id=user.user_id,
             plugin_name=plugin_name,
-                    url=server_url,
             plugin_type=plugin_type,
+            server_url=server_url,
             headers=headers,
-                    timeout=config.get('timeout', 60.0) if config else 60.0
+            config=config
         ))
     else:
-                success = False
-            
-            # 更新状态
-            plugin.status = "active" if success else "error"
-            plugin.last_error = None if success else "加载失败"
-            await db.commit()
-            await db.refresh(plugin)
-        except Exception as e:
-            logger.error(f"注册插件失败: {plugin_name}, 错误: {e}")
-            plugin.status = "error"
-            plugin.last_error = str(e)
-            await db.commit()
-            await db.refresh(plugin)
-    else:
-        # 禁用：从统一门面注销（不影响数据库状态）
-        try:
-            await mcp_client.unregister(user.user_id, plugin_name)
-        except Exception as e:
-            logger.warning(f"注销插件时出错（可忽略）: {plugin_name}, 错误: {e}")
+        # 禁用：后台从统一门面注销（不影响数据库状态）
+        asyncio.create_task(_unregister_plugin_safe(user.user_id, plugin_name))
     
     action = "启用" if enabled else "禁用"
-    logger.info(f"用户 {user.user_id} {action}插件: {plugin_name}")
+    logger.info(f"用户 {user.user_id} {action}插件: {plugin_name}（MCP操作在后台执行）")
     return plugin
 
 

backend/app/api/organizations.py

@@ -7,7 +7,7 @@ from pydantic import BaseModel, Field
 import json
 
 from app.database import get_db
-from app.utils.sse_response import SSEResponse, create_sse_response, WizardProgressTracker
+from app.utils.sse_response import SSEResponse, create_sse_response, WizardProgressTracker, wrap_stream_with_heartbeat, HEARTBEAT
 from app.models.relationship import Organization, OrganizationMember
 from app.models.character import Character
 from app.models.project import Project
@@ -24,6 +24,7 @@ from app.schemas.relationship import (
 )
 from app.schemas.character import CharacterResponse
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.services.prompt_service import prompt_service, PromptService
 from app.logger import get_logger
 from app.api.settings import get_user_ai_service
@@ -500,7 +501,15 @@ async def generate_organization_stream(
                 chunk_count = 0
                 estimated_total = max(3000, len(prompt) * 8)
                 
-                async for chunk in user_ai_service.generate_text_stream(prompt=prompt):
+                async for chunk in wrap_stream_with_heartbeat(
+                    user_ai_service.generate_text_stream(prompt=prompt),
+                    heartbeat_interval=15.0
+                ):
+                    # 心跳哨兵：发送心跳保活，不混入AI响应
+                    if chunk is HEARTBEAT:
+                        yield await tracker.heartbeat()
+                        continue
+
                     chunk_count += 1
                     ai_content += chunk
                     
@@ -529,7 +538,7 @@ async def generate_organization_stream(
             # ✅ 使用统一的 JSON 清洗方法
             try:
                 cleaned_response = user_ai_service._clean_json_response(ai_content)
-                organization_data = json.loads(cleaned_response)
+                organization_data = loads_json(cleaned_response)
                 logger.info(f"✅ 组织JSON解析成功")
             except json.JSONDecodeError as e:
                 logger.error(f"❌ 组织JSON解析失败: {e}")

backend/app/api/outlines.py

@@ -27,6 +27,7 @@ from app.schemas.outline import (
     CreateChaptersFromPlansResponse
 )
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.services.prompt_service import prompt_service, PromptService
 from app.services.memory_service import memory_service
 from app.services.plot_expansion_service import PlotExpansionService
@@ -850,7 +851,7 @@ def _parse_ai_response(ai_response: str, raise_on_error: bool = False) -> list:
         ai_service_temp = AIService()
         cleaned_text = ai_service_temp._clean_json_response(ai_response)
         
-        outline_data = json.loads(cleaned_text)
+        outline_data = loads_json(cleaned_text)
         
         # 确保是列表格式
         if not isinstance(outline_data, list):
@@ -1447,6 +1448,31 @@ async def continue_outline_generator(
                 message=f"🤖 调用AI生成第{str(batch_num + 1)}批..."
             )
             
+            # 获取伏笔提醒信息（用于大纲续写）
+            foreshadow_reminders_text = "暂无需要关注的伏笔"
+            try:
+                foreshadow_context = await foreshadow_service.build_chapter_context(
+                    db=db,
+                    project_id=project_id,
+                    chapter_number=current_start_chapter,
+                    include_pending=False,
+                    include_overdue=True,
+                    lookahead=10
+                )
+                if foreshadow_context and foreshadow_context.get("context_text"):
+                    foreshadow_reminders_text = foreshadow_context["context_text"]
+                    logger.info(f"✅ 大纲续写获取到伏笔提醒: {len(foreshadow_reminders_text)}字符")
+                    # 追加伏笔统计信息
+                    foreshadow_stats = await foreshadow_service.get_stats(db, project_id)
+                    if foreshadow_stats:
+                        planted = foreshadow_stats.get('planted', 0)
+                        resolved = foreshadow_stats.get('resolved', 0)
+                        partial = foreshadow_stats.get('partially_resolved', 0)
+                        pending = foreshadow_stats.get('pending', 0)
+                        foreshadow_reminders_text += f"\n【📊 伏笔统计】已埋设:{planted} 已回收:{resolved} 部分回收:{partial} 待埋入:{pending}"
+            except Exception as e:
+                logger.warning(f"⚠️ 获取大纲续写伏笔提醒失败: {str(e)}")
+
             # 使用标准续写提示词模板（简化版）
             template = await PromptService.get_template("OUTLINE_CONTINUE", user_id, db)
             prompt = PromptService.format_prompt(
@@ -1463,6 +1489,8 @@ async def continue_outline_generator(
                 # 上下文信息
                 recent_outlines=context['recent_outlines'],
                 characters_info=context['characters_info'],
+                # 伏笔提醒
+                foreshadow_reminders=foreshadow_reminders_text,
                 # 续写参数
                 chapter_count=current_batch_size,
                 start_chapter=current_start_chapter,

backend/app/api/wizard_stream.py

@@ -16,6 +16,7 @@ from app.models.relationship import CharacterRelationship, Organization, Organiz
 from app.models.writing_style import WritingStyle
 from app.models.project_default_style import ProjectDefaultStyle
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.services.prompt_service import prompt_service, PromptService
 from app.services.plot_expansion_service import PlotExpansionService
 from app.logger import get_logger
@@ -169,7 +170,7 @@ async def world_building_generator(
                     logger.info(f"✅ JSON清洗完成，清洗后长度: {len(cleaned_text)}")
                     logger.info(f"   清洗后预览: {cleaned_text[:300]}...")
                     
-                    world_data = json.loads(cleaned_text)
+                    world_data = loads_json(cleaned_text)
                     logger.info(f"✅ 世界观JSON解析成功（尝试{world_retry_count+1}/{MAX_WORLD_RETRIES}）")
                     world_generation_success = True  # 解析成功，标记完成
                             
@@ -433,7 +434,7 @@ async def career_system_generator(
                 # 清洗并解析JSON
                 try:
                     cleaned_response = user_ai_service._clean_json_response(career_response)
-                    career_data = json.loads(cleaned_response)
+                    career_data = loads_json(cleaned_response)
                     logger.info(f"✅ 职业体系JSON解析成功（尝试{career_retry_count+1}/{MAX_CAREER_RETRIES}）")
                     
                     yield await tracker.saving("保存职业数据...")
@@ -771,7 +772,7 @@ async def characters_generator(
                     
                     # 解析批次结果 - 使用统一的JSON清洗方法
                     cleaned_text = user_ai_service._clean_json_response(accumulated_text)
-                    characters_data = json.loads(cleaned_text)
+                    characters_data = loads_json(cleaned_text)
                     if not isinstance(characters_data, list):
                         characters_data = [characters_data]
                     
@@ -1362,7 +1363,7 @@ async def outline_generator(
         
         try:
             cleaned_text = user_ai_service._clean_json_response(accumulated_text)
-            outline_data = json.loads(cleaned_text)
+            outline_data = loads_json(cleaned_text)
             if not isinstance(outline_data, list):
                 outline_data = [outline_data]
         except json.JSONDecodeError as e:
@@ -1668,7 +1669,7 @@ async def world_building_regenerate_generator(
                     cleaned_text = user_ai_service._clean_json_response(accumulated_text)
                     logger.info(f"✅ JSON清洗完成，清洗后长度: {len(cleaned_text)}")
                     
-                    world_data = json.loads(cleaned_text)
+                    world_data = loads_json(cleaned_text)
                     logger.info(f"✅ 世界观重新生成JSON解析成功（尝试{world_retry_count+1}/{MAX_WORLD_RETRIES}）")
                     world_generation_success = True
                             

backend/app/mcp/facade.py

@@ -316,6 +316,9 @@ class MCPClientFacade:
             if key in self._sessions:
                 await self._close_session_unsafe(key)
 
+            stream_ctx = None
+            session = None
+            
             try:
                 logger.info(f"🔗 连接MCP服务器: {config.plugin_name} -> {config.url} (类型: {config.plugin_type})")
 
@@ -365,11 +368,19 @@ class MCPClientFacade:
                     error_details.append(f"{type(exc).__name__}: {exc}")
                 error_msg = "; ".join(error_details)
                 logger.error(f"❌ MCP连接失败 {key}: TaskGroup异常 - {error_msg}")
+                
+                # 在同一任务中清理已创建的上下文，避免跨任务清理cancel scope
+                await self._cleanup_contexts_in_task(session, stream_ctx)
+                
                 await self._emit_status_change(config.user_id, config.plugin_name, "inactive", "error", error_msg)
                 return False
 
             except Exception as e:
                 logger.error(f"❌ MCP连接失败 {key}: {type(e).__name__}: {e}")
+                
+                # 在同一任务中清理已创建的上下文，避免跨任务清理cancel scope
+                await self._cleanup_contexts_in_task(session, stream_ctx)
+                
                 await self._emit_status_change(config.user_id, config.plugin_name, "inactive", "error", str(e))
                 return False
     
@@ -392,6 +403,27 @@ class MCPClientFacade:
         
         await self._emit_status_change(user_id, plugin_name, old_status, "inactive", "已注销")
     
+    async def _cleanup_contexts_in_task(self, session, stream_ctx):
+        """在当前任务中清理已创建的上下文（异步方法）
+        
+        当MCP连接失败时，上下文（cancel scope）必须在与创建时相同的任务中清理。
+        由于异常处理和上下文创建在同一个任务中，这里可以安全地await __aexit__。
+        """
+        # 先清理session，再清理stream（LIFO顺序）
+        if session is not None:
+            try:
+                await session.__aexit__(None, None, None)
+            except Exception as e:
+                logger.debug(f"清理session上下文: {e}")
+        
+        if stream_ctx is not None:
+            try:
+                await stream_ctx.__aexit__(None, None, None)
+            except Exception as e:
+                logger.debug(f"清理stream上下文: {e}")
+        
+        logger.debug("已在当前任务中清理MCP上下文")
+    
     async def _close_session_unsafe(self, key: str):
         """关闭会话（不加用户锁，需要调用者确保线程安全）"""
         async with self._session_lock:

backend/app/schemas/career.py

@@ -78,6 +78,7 @@ class CareerGenerateRequest(BaseModel):
     project_id: str = Field(..., description="项目ID")
     main_career_count: int = Field(5, description="主职业数量", ge=1, le=20)
     sub_career_count: int = Field(8, description="副职业数量", ge=0, le=30)
+    user_requirements: str = Field("", description="用户额外要求")
     enable_mcp: bool = Field(False, description="是否启用MCP工具增强")
 
 

backend/app/services/foreshadow_service.py

@@ -1283,6 +1283,10 @@ class ForeshadowService:
             # 预先获取所有已埋入的伏笔，用于内容匹配
             planted_foreshadows = await self.get_planted_foreshadows_for_analysis(db, project_id)
             
+            # 每章最多创建的新伏笔数量
+            MAX_NEW_FORESHADOWS_PER_CHAPTER = 2
+            new_foreshadow_count = 0
+
             for fs_data in analysis_foreshadows:
                 try:
                     fs_type = fs_data.get("type", "planted")
@@ -1416,6 +1420,11 @@ class ForeshadowService:
                             logger.info(f"📝 更新已存在伏笔（避免重复）: {fs_title} (ID: {existing_fs.id})")
                         else:
                             # 创建新伏笔
+                            # 检查每章新伏笔数量上限
+                            if new_foreshadow_count >= MAX_NEW_FORESHADOWS_PER_CHAPTER:
+                                logger.info(f"🚫 已达每章新伏笔上限({MAX_NEW_FORESHADOWS_PER_CHAPTER}个)，跳过: {fs_title}")
+                                continue
+                            
                             # 不再为 estimated_resolve_chapter 设置默认值，避免误报"超期"
                             estimated_resolve = fs_data.get("estimated_resolve_chapter")
                             if estimated_resolve is None:
@@ -1448,10 +1457,11 @@ class ForeshadowService:
                             db.add(new_foreshadow)
                             await db.flush()
                             
+                            new_foreshadow_count += 1
                             stats["planted_count"] += 1
                             stats["created_count"] += 1
                             stats["created_ids"].append(new_foreshadow.id)
-                            logger.info(f"✅ 自动创建伏笔: {fs_title} (ID: {new_foreshadow.id})")
+                            logger.info(f"✅ 自动创建伏笔: {fs_title} (ID: {new_foreshadow.id}) [{new_foreshadow_count}/{MAX_NEW_FORESHADOWS_PER_CHAPTER}]")
                     
                 except Exception as item_error:
                     error_msg = f"处理伏笔时出错: {str(item_error)}"

backend/app/services/json_helper.py

@@ -4,9 +4,154 @@ import re
 from typing import Any, Dict, List, Union
 from app.logger import get_logger
 
+try:
+    import json5
+    HAS_JSON5 = True
+except ImportError:
+    HAS_JSON5 = False
+
 logger = get_logger(__name__)
 
 
+# 中文引号/括号到ASCII的映射
+_QUOTE_MAP = {
+    '\u201c': '"',  # " → "
+    '\u201d': '"',  # " → "
+    '\u2018': "'",  # ' → '
+    '\u2019': "'",  # ' → '
+    '\u300e': '"',  # 『 → "
+    '\u300f': '"',  # 』 → "
+    '\u300c': '"',  # 「 → "
+    '\u300d': '"',  # 」 → "
+}
+
+
+def _fix_json_string_values(text: str) -> str:
+    """
+    修复JSON字符串值中的常见问题：
+    1. 裸换行符/制表符 → 转义
+    2. 字符串值内的中文引号 → 转义为ASCII引号（避免破坏JSON结构）
+    3. 结构位置的中文引号 → 直接替换为ASCII引号
+    
+    AI生成的JSON常在字符串值中插入未转义的换行符和中文引号。
+    此函数遍历文本，区分字符串内外，分别处理。
+    """
+    if not text or '"' not in text:
+        return text
+    
+    result = []
+    i = 0
+    in_string = False
+    fixed_count = 0
+    
+    while i < len(text):
+        c = text[i]
+        
+        if c == '"' and not in_string:
+            # 进入字符串
+            in_string = True
+            result.append(c)
+            i += 1
+            continue
+        
+        if in_string:
+            if c == '\\':
+                # 转义字符，检查下一个字符是否合法
+                if i + 1 < len(text):
+                    next_c = text[i + 1]
+                    # JSON 合法转义：\" \\ \/ \b \f \n \r \t \uXXXX
+                    if next_c in ('"', '\\', '/', 'b', 'f', 'n', 'r', 't'):
+                        # 合法转义，直接保留
+                        result.append(c)
+                        result.append(next_c)
+                        i += 2
+                        continue
+                    elif next_c == 'u':
+                        # Unicode 转义 \uXXXX，检查是否有4个十六进制字符
+                        if i + 5 < len(text) and all(text[i+2+k] in '0123456789abcdefABCDEF' for k in range(4)):
+                            result.append(text[i:i+6])
+                            i += 6
+                            continue
+                        else:
+                            # 不完整的unicode转义，去掉反斜杠
+                            result.append(next_c)
+                            fixed_count += 1
+                            i += 2
+                            continue
+                    else:
+                        # 非法转义字符（如 \c \p \d 等），去掉反斜杠只保留字符
+                        result.append(next_c)
+                        fixed_count += 1
+                        i += 2
+                        continue
+                else:
+                    # 末尾孤立的反斜杠，去掉
+                    fixed_count += 1
+                    i += 1
+                    continue
+            
+            if c == '"':
+                # 字符串结束
+                in_string = False
+                result.append(c)
+                i += 1
+                continue
+            
+            if c == '\n':
+                # 裸换行符 → 替换为转义换行
+                result.append('\\')
+                result.append('n')
+                fixed_count += 1
+                i += 1
+                continue
+            
+            if c == '\r':
+                # 裸回车符 → 忽略或替换
+                if i + 1 < len(text) and text[i + 1] == '\n':
+                    result.append('\\')
+                    result.append('n')
+                    fixed_count += 1
+                    i += 2
+                else:
+                    result.append('\\')
+                    result.append('n')
+                    fixed_count += 1
+                    i += 1
+                continue
+            
+            if c == '\t':
+                # 裸制表符 → 替换为转义制表符
+                result.append('\\')
+                result.append('t')
+                fixed_count += 1
+                i += 1
+                continue
+            
+            # 字符串值内的中文引号 → 转义为 \"（避免破坏JSON结构）
+            if c in _QUOTE_MAP:
+                result.append('\\')
+                result.append(_QUOTE_MAP[c])
+                fixed_count += 1
+                i += 1
+                continue
+        
+        # 非字符串内的字符
+        # 结构位置的中文引号 → 直接替换
+        if not in_string and c in _QUOTE_MAP:
+            result.append(_QUOTE_MAP[c])
+            fixed_count += 1
+            i += 1
+            continue
+        
+        result.append(c)
+        i += 1
+    
+    if fixed_count > 0:
+        logger.debug(f"✅ 修复了{fixed_count}个JSON问题（裸控制字符/中文引号）")
+    
+    return ''.join(result)
+
+
 def clean_json_response(text: str) -> str:
     """清洗 AI 返回的 JSON（改进版 - 流式安全）"""
     try:
@@ -17,6 +162,13 @@ def clean_json_response(text: str) -> str:
         original_length = len(text)
         logger.debug(f"🔍 开始清洗JSON，原始长度: {original_length}")
         
+        # 替换中文逗号/冒号（AI可能在JSON结构位置使用，全局替换是安全的）
+        text = text.replace('\uff0c', ',')   # ，→ ,
+        text = text.replace('\uff1a', ':')   # ：→ :
+        
+        # 修复JSON中的中文引号和裸控制字符（上下文感知，区分字符串内外）
+        text = _fix_json_string_values(text)
+        
         # 去除 markdown 代码块
         text = re.sub(r'^```json\s*\n?', '', text, flags=re.MULTILINE | re.IGNORECASE)
         text = re.sub(r'^```\s*\n?', '', text, flags=re.MULTILINE)
@@ -148,12 +300,54 @@ def clean_json_response(text: str) -> str:
 
 
 def parse_json(text: str) -> Union[Dict, List]:
-    """解析 JSON"""
-    try:
+    """解析 JSON，优先使用标准json，失败后用json5容错解析"""
     cleaned = clean_json_response(text)
+    
+    # 优先使用标准 json
+    try:
         return json.loads(cleaned)
-    except Exception as e:
-        logger.error(f"❌ parse_json 出错: {e}")
+    except (json.JSONDecodeError, Exception):
+        pass
+    
+    # json5 容错解析（处理单引号、多余逗号、宽松格式等）
+    if HAS_JSON5:
+        try:
+            logger.info("🔄 标准JSON解析失败，使用json5容错解析")
+            result = json5.loads(cleaned)
+            logger.info("✅ json5容错解析成功")
+            return result
+        except Exception as e5:
+            logger.error(f"❌ json5容错解析也失败: {e5}")
+    
+    # 最终失败
+    logger.error(f"❌ parse_json 完全失败")
     logger.error(f"   原始文本长度: {len(text) if text else 0}")
     logger.error(f"   清洗后文本长度: {len(cleaned) if cleaned else 0}")
-        raise
+    logger.debug(f"   清洗后文本预览: {cleaned[:500] if cleaned else 'None'}")
+    raise json.JSONDecodeError("JSON解析失败（标准和json5均失败）", cleaned, 0)
+
+
+def loads_json(text: str) -> Any:
+    """
+    json.loads 的容错替代品，可直接替换 json.loads()。
+    优先用标准 json.loads，失败后自动降级到 json5。
+    适用于解析 AI 返回的、可能包含不规范格式的 JSON。
+    """
+    # 优先使用标准 json
+    try:
+        return json.loads(text)
+    except (json.JSONDecodeError, Exception):
+        pass
+    
+    # json5 容错解析
+    if HAS_JSON5:
+        try:
+            logger.info("🔄 json.loads失败，使用json5容错解析")
+            result = json5.loads(text)
+            logger.info("✅ json5容错解析成功")
+            return result
+        except Exception as e5:
+            logger.error(f"❌ json5容错解析也失败: {e5}")
+    
+    # 最终失败，抛出标准异常
+    raise json.JSONDecodeError("JSON解析失败（标准和json5均失败）", text, 0)

backend/app/services/plot_analyzer.py

@@ -2,6 +2,7 @@
 from typing import Dict, Any, List, Optional, Callable, Awaitable
 from sqlalchemy.ext.asyncio import AsyncSession
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.services.prompt_service import prompt_service, PromptService
 from app.logger import get_logger
 import json
@@ -277,7 +278,7 @@ class PlotAnalyzer:
             cleaned = self.ai_service._clean_json_response(response)
             
             # 尝试解析JSON
-            result = json.loads(cleaned)
+            result = loads_json(cleaned)
             
             # 验证必要字段
             required_fields = ['hooks', 'plot_points', 'scores']

backend/app/services/plot_expansion_service.py

@@ -9,6 +9,7 @@ from app.models.project import Project
 from app.models.character import Character
 from app.models.chapter import Chapter
 from app.services.ai_service import AIService
+from app.services.json_helper import loads_json
 from app.services.prompt_service import prompt_service, PromptService
 from app.logger import get_logger
 
@@ -531,7 +532,7 @@ class PlotExpansionService:
             cleaned_text = self.ai_service._clean_json_response(ai_response)
             
             # 解析JSON
-            chapter_plans = json.loads(cleaned_text)
+            chapter_plans = loads_json(cleaned_text)
             
             # 确保是列表
             if not isinstance(chapter_plans, list):

backend/app/utils/sse_response.py

@@ -388,6 +388,42 @@ async def create_sse_generator(
         yield await SSEResponse.send_error(str(e))
 
 
+class _HeartbeatSentinel:
+    """心跳哨兵对象，用于标识心跳事件（非AI内容）"""
+    pass
+
+HEARTBEAT = _HeartbeatSentinel()
+
+
+async def wrap_stream_with_heartbeat(
+    async_gen: AsyncGenerator,
+    heartbeat_interval: float = 15.0
+) -> AsyncGenerator:
+    """
+    包装异步生成器，在等待数据时产生心跳哨兵，防止连接超时断开。
+    
+    用法：
+        async for chunk in wrap_stream_with_heartbeat(
+            ai_service.generate_text_stream(prompt), 
+            heartbeat_interval=15
+        ):
+            if chunk is HEARTBEAT:
+                yield await tracker.heartbeat()
+                continue
+            # chunk 是原始AI数据
+    """
+    ait = async_gen.__aiter__()
+    while True:
+        try:
+            item = await asyncio.wait_for(ait.__anext__(), timeout=heartbeat_interval)
+            yield item
+        except asyncio.TimeoutError:
+            # 等待超时，产生心跳哨兵
+            yield HEARTBEAT
+        except StopAsyncIteration:
+            return
+
+
 def create_sse_response(generator: AsyncGenerator[str, None]) -> StreamingResponse:
     """
     创建SSE StreamingResponse - 兼容HTTP/2协议

backend/requirements.txt

@@ -35,5 +35,8 @@ transformers==4.57.1
 # Sentence Transformers（更新到最新稳定版本以修复 FutureWarning）
 sentence-transformers==5.1.2
 
+# 宽松JSON解析
+json5==0.12.0
+
 # PyTorch 版本锁定（用于打包环境）
 torch==2.8.0

frontend/src/pages/Careers.tsx

@@ -171,47 +171,67 @@ export default function Careers() {
 
         try {
             const userRequirements = values.user_requirements?.trim() || '';
-            const eventSource = new EventSource(
-                `/api/careers/generate-system?` +
-                new URLSearchParams({
-                    project_id: projectId || '',
-                    main_career_count: values.main_career_count.toString(),
-                    sub_career_count: values.sub_career_count.toString(),
-                    user_requirements: userRequirements,
-                    enable_mcp: 'false'
-                }).toString(),
-                { withCredentials: true }
-            );
 
-            eventSource.onmessage = (event) => {
+            // 使用 fetch + POST 替代 EventSource GET，避免 URL 长度限制
+            const response = await fetch('/api/careers/generate-system', {
+                method: 'POST',
+                headers: {
+                    'Content-Type': 'application/json',
+                },
+                credentials: 'include',
+                body: JSON.stringify({
+                    project_id: projectId || '',
+                    main_career_count: values.main_career_count,
+                    sub_career_count: values.sub_career_count,
+                    user_requirements: userRequirements,
+                    enable_mcp: false
+                })
+            });
+
+            if (!response.ok || !response.body) {
+                setAiGenerating(false);
+                message.error(`请求失败: ${response.status}`);
+                return;
+            }
+
+            const reader = response.body.getReader();
+            const decoder = new TextDecoder();
+            let buffer = '';
+
+            while (true) {
+                const { done, value } = await reader.read();
+                if (done) break;
+
+                buffer += decoder.decode(value, { stream: true });
+                const lines = buffer.split('\n');
+                buffer = lines.pop() || '';
+
+                for (const line of lines) {
+                    if (line.startsWith('data: ')) {
                         try {
-                    const data = JSON.parse(event.data);
+                            const data = JSON.parse(line.slice(6));
 
                             if (data.type === 'progress') {
                                 setAiProgress(data.progress || 0);
                                 setAiMessage(data.message || '');
                             } else if (data.type === 'done') {
-                        eventSource.close();
                                 setTimeout(() => {
                                     setAiGenerating(false);
                                     message.success('AI新职业生成完成！');
                                     fetchCareers();
                                 }, 1000);
                             } else if (data.type === 'error') {
-                        eventSource.close();
                                 setAiGenerating(false);
-                        message.error(data.message || '生成失败');
+                                message.error(data.error || data.message || '生成失败');
                             }
                         } catch (e) {
-                    console.error('解析SSE数据失败:', e);
+                            // 忽略非JSON行（如心跳注释）
+                        }
+                    }
+                }
             }
-            };
 
-            eventSource.onerror = () => {
-                eventSource.close();
             setAiGenerating(false);
-                message.error('连接中断，生成失败');
-            };
         } catch (err: unknown) {
             setAiGenerating(false);
             const error = err as Error;

frontend/src/utils/sseClient.ts

@@ -151,6 +151,7 @@ export class SSEPostClient {
           headers: {
             'Content-Type': 'application/json',
           },
+          credentials: 'include',
           body: JSON.stringify(this.data),
           signal: this.abortController.signal,
         });