feat: add DeepSeek official API support and capture reasoning content

- Add DeepSeek provider routing: deepseek-chat/deepseek-reasoner use
  api.deepseek.com, other models use SiliconFlow
- Add stream_with_reasoning() using raw OpenAI SDK to capture
  reasoning_content (langchain_openai strips this field)
- RAG chain and conversation chain both use stream_with_reasoning
  for proper reasoning display in thinking models
- Frontend model selector: grouped by provider (DeepSeek official +
  SiliconFlow), default changed to deepseek-chat
- Regenerate message converted to streaming with reasoning capture
- Minor UI: globals.css additions, chat store refactoring

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/src/core/config.py

@@ -23,6 +23,10 @@ class Settings(BaseSettings):
     siliconflow_base_url: str = "https://api.siliconflow.cn/v1"
     siliconflow_model: str = "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B"
 
+    # DeepSeek官方API配置
+    deepseek_api_key: str = ""
+    deepseek_base_url: str = "https://api.deepseek.com"
+    
     # 数据库配置
     database_url: str = "sqlite:///../data/database/course_agent.db"
     

backend/src/llm/siliconflow.py

@@ -1,43 +1,62 @@
 """
-硅基流动大模型API集成
+大模型API集成 — 支持 SiliconFlow 和 DeepSeek 官方
 """
 import os
-from typing import List, Dict, Any, Optional, AsyncGenerator
+from typing import List, Dict, Any, Optional, AsyncGenerator, Tuple
 from langchain_openai import ChatOpenAI
 from langchain_core.messages import BaseMessage, HumanMessage, AIMessage, SystemMessage
 from langchain_core.callbacks.base import BaseCallbackHandler
 from langchain_core.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
+import openai
 
 from ..core.config import get_settings
 
 settings = get_settings()
 
+# DeepSeek 官方模型 ID 前缀（用于自动路由）
+DEEPSEEK_OFFICIAL_MODELS = {
+    "deepseek-chat",
+    "deepseek-reasoner",
+}
+
+# 模型 ID → (api_key, base_url) 的路由映射
+def _resolve_provider(model: str) -> tuple[str, str]:
+    """根据模型 ID 选择 API provider，返回 (api_key, base_url)"""
+    # DeepSeek 官方模型（不带 siliconflow 前缀的纯 deepseek-xxx）
+    if model in DEEPSEEK_OFFICIAL_MODELS or model.startswith("deepseek/"):
+        if not settings.deepseek_api_key:
+            raise ValueError(
+                f"模型 {model} 需要 DeepSeek 官方 API Key，"
+                "请在 .env 中配置 DEEPSEEK_API_KEY"
+            )
+        actual_model = model.replace("deepseek/", "")
+        return settings.deepseek_api_key, settings.deepseek_base_url, actual_model
+
+    # 默认走 SiliconFlow
+    return settings.siliconflow_api_key, settings.siliconflow_base_url, model
+
 
 class SiliconFlowLLM:
-    """硅基流动大模型客户端"""
+    """大模型客户端 — 自动路由 SiliconFlow / DeepSeek 官方"""
 
     def __init__(self, model: Optional[str] = None):
-        """初始化LLM客户端"""
+        raw_model = model or settings.siliconflow_model
-        # 设置环境变量
+        api_key, base_url, resolved_model = _resolve_provider(raw_model)
-        os.environ["OPENAI_API_KEY"] = settings.siliconflow_api_key
-        os.environ["OPENAI_API_BASE"] = settings.siliconflow_base_url
 
-        # 使用传入的模型或默认模型
+        self.model_name = resolved_model
-        self.model_name = model or settings.siliconflow_model
+        print(f"[LLM] 模型: {resolved_model}, API: {base_url}")
-        print(f"[DEBUG-LLM] 初始化LLM客户端，使用模型: {self.model_name} (传入参数: {model}, 默认配置: {settings.siliconflow_model})")
 
-        # 创建LLM实例
         self.llm = ChatOpenAI(
-            model=self.model_name,
+            model=resolved_model,
-            api_key=settings.siliconflow_api_key,
+            api_key=api_key,
-            base_url=settings.siliconflow_base_url,
+            base_url=base_url,
             temperature=0.7,
             max_tokens=2000,
             streaming=True
         )
-        # 检查实际使用的模型名称
+
-        actual_model = getattr(self.llm, 'model_name', None) or getattr(self.llm, 'model', None) or str(self.llm)
+        self._api_key = api_key
-        print(f"[DEBUG-LLM] ChatOpenAI实例创建完成，实际模型: {actual_model}")
+        self._base_url = base_url
 
     def chat(self, messages: List[BaseMessage], **kwargs) -> str:
         """同步聊天"""
@@ -64,6 +83,49 @@ class SiliconFlowLLM:
         except Exception as e:
             raise Exception(f"LLM流式调用失败: {str(e)}")
 
+    async def stream_with_reasoning(
+        self, messages: List[BaseMessage], **kwargs
+    ) -> AsyncGenerator[Tuple[str, str], None]:
+        """流式调用（直接使用 OpenAI SDK，捕获推理内容）
+
+        Yields:
+            (type, content) — type 为 "reasoning" 或 "content"
+        """
+        client = openai.AsyncOpenAI(
+            api_key=self._api_key,
+            base_url=self._base_url
+        )
+
+        openai_messages = []
+        for msg in messages:
+            if isinstance(msg, SystemMessage):
+                openai_messages.append({"role": "system", "content": msg.content})
+            elif isinstance(msg, HumanMessage):
+                openai_messages.append({"role": "user", "content": msg.content})
+            elif isinstance(msg, AIMessage):
+                openai_messages.append({"role": "assistant", "content": msg.content})
+            else:
+                openai_messages.append({"role": "user", "content": str(msg.content)})
+
+        stream = await client.chat.completions.create(
+            model=self.model_name,
+            messages=openai_messages,
+            stream=True,
+            **kwargs
+        )
+
+        async for chunk in stream:
+            if not chunk.choices:
+                continue
+            delta = chunk.choices[0].delta
+
+            rc = getattr(delta, 'reasoning_content', None)
+            if rc:
+                yield ("reasoning", rc)
+
+            if delta.content:
+                yield ("content", delta.content)
+
     def create_messages(
         self,
         user_message: str,
@@ -73,11 +135,9 @@ class SiliconFlowLLM:
         """创建消息列表"""
         messages = []
 
-        # 添加系统提示
         if system_prompt:
             messages.append(SystemMessage(content=system_prompt))
 
-        # 添加聊天历史
         if chat_history:
             for msg in chat_history:
                 if msg["role"] == "user":
@@ -85,7 +145,6 @@ class SiliconFlowLLM:
                 elif msg["role"] == "assistant":
                     messages.append(AIMessage(content=msg["content"]))
 
-        # 添加当前用户消息
         messages.append(HumanMessage(content=user_message))
 
         return messages
@@ -104,5 +163,4 @@ def get_llm_client(model: Optional[str] = None) -> SiliconFlowLLM:
     if model is None:
         return llm_client
     else:
-        # 为指定模型创建新实例
         return SiliconFlowLLM(model=model)

backend/src/rag/chains.py

@@ -32,7 +32,9 @@ class RAGChain:
             model: 可选的模型名称，如 deepseek-ai/DeepSeek-V3, Qwen/QwQ-32B
         """
         print(f"[DEBUG-RAGChain] 初始化，knowledge_base_ids: {knowledge_base_ids}, model: {model}")
-        self.llm = get_llm_client(model=model).llm
+        _sf_client = get_llm_client(model=model)
+        self.llm = _sf_client.llm
+        self.client = _sf_client
         self.vector_store = get_vector_store()
         self.knowledge_base_ids = knowledge_base_ids
         
@@ -116,7 +118,8 @@ class RAGChain:
         context = "\n\n".join(doc.page_content for doc in docs)
         
         # 构建prompt
-        messages = await self.prompt.ainvoke({"context": context, "question": question})
+        prompt_value = await self.prompt.ainvoke({"context": context, "question": question})
+        messages = prompt_value.to_messages()
         
         # 流式生成答案
         async for chunk in self.llm.astream(messages):
@@ -161,24 +164,17 @@ class RAGChain:
 
         # 2. 构建prompt
         yield {"type": "thinking", "stage": "generating", "message": f"基于 {len(docs)} 篇文档生成回答..."}
-        messages = await self.prompt.ainvoke({"context": context, "question": question})
+        prompt_value = await self.prompt.ainvoke({"context": context, "question": question})
+        messages = prompt_value.to_messages()
 
-        # 3. 流式生成答案（捕获推理内容）
+        # 3. 流式生成答案（使用原始 OpenAI SDK 捕获推理内容）
         answer_chunks = []
         reasoning_parts = []
-        async for chunk in self.llm.astream(messages):
+        async for event_type, content in self.client.stream_with_reasoning(messages):
-            # 捕获推理内容（DeepSeek-R1/QwQ 等推理模型）
+            if event_type == "reasoning":
-            if hasattr(chunk, 'additional_kwargs') and 'reasoning_content' in chunk.additional_kwargs:
+                reasoning_parts.append(content)
-                reasoning_text = chunk.additional_kwargs['reasoning_content']
+                yield {"type": "thinking", "stage": "reasoning", "message": content}
-                if reasoning_text:
+            elif event_type == "content":
-                    reasoning_parts.append(reasoning_text)
-                    yield {"type": "thinking", "stage": "reasoning", "message": reasoning_text}
-            elif hasattr(chunk, 'reasoning_content') and chunk.reasoning_content:
-                reasoning_parts.append(chunk.reasoning_content)
-                yield {"type": "thinking", "stage": "reasoning", "message": chunk.reasoning_content}
-
-            content = chunk.content if hasattr(chunk, 'content') else str(chunk)
-            if content:
                 answer_chunks.append(content)
                 yield {"type": "chunk", "content": content}
         

backend/src/rag/conversation_chains.py

@@ -24,7 +24,9 @@ class ConversationChain:
             system_prompt: 系统提示词
             model: 可选的模型名称，如 deepseek-ai/DeepSeek-V3, Qwen/QwQ-32B
         """
-        self.llm = get_llm_client(model=model).llm
+        _sf_client = get_llm_client(model=model)
+        self.llm = _sf_client.llm
+        self.client = _sf_client
         self.system_prompt = system_prompt or "你是一个专业的国土空间规划知识问答助手。请基于你的知识回答用户的问题。"
         
         # 创建带历史的Prompt模板
@@ -94,24 +96,18 @@ class ConversationChain:
 
         yield {"type": "thinking", "stage": "generating", "message": "正在生成回答..."}
 
-        # 直接流式调用 LLM 以捕获推理内容
+        # 直接使用原始 OpenAI SDK 捕获推理内容
-        prompt_messages = await self.prompt.ainvoke({
+        prompt_value = await self.prompt.ainvoke({
             "question": question,
             "chat_history": history_messages
         })
+        prompt_messages = prompt_value.to_messages()
 
         content_started = False
-        async for chunk in self.llm.astream(prompt_messages):
+        async for event_type, content in self.client.stream_with_reasoning(prompt_messages):
-            # 捕获推理内容（DeepSeek-R1/QwQ 等推理模型）
+            if event_type == "reasoning":
-            if hasattr(chunk, 'additional_kwargs') and 'reasoning_content' in chunk.additional_kwargs:
+                yield {"type": "thinking", "stage": "reasoning", "message": content}
-                reasoning_text = chunk.additional_kwargs['reasoning_content']
+            elif event_type == "content":
-                if reasoning_text:
-                    yield {"type": "thinking", "stage": "reasoning", "message": reasoning_text}
-            elif hasattr(chunk, 'reasoning_content') and chunk.reasoning_content:
-                yield {"type": "thinking", "stage": "reasoning", "message": chunk.reasoning_content}
-
-            content = chunk.content if hasattr(chunk, 'content') else str(chunk)
-            if content:
                 if not content_started:
                     content_started = True
                 yield {"type": "chunk", "content": content}

env.example

@@ -3,6 +3,10 @@ SILICONFLOW_API_KEY=your-api-key-here
 SILICONFLOW_BASE_URL=https://api.siliconflow.cn/v1
 SILICONFLOW_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507
 
+# DeepSeek官方API配置（可选，不配则走硅基流动）
+# DEEPSEEK_API_KEY=your-deepseek-api-key-here
+DEEPSEEK_BASE_URL=https://api.deepseek.com
+
 # 数据库配置（本地开发默认使用SQLite）
 DATABASE_URL=sqlite:///../data/database/course_agent.db
 # Docker部署使用PostgreSQL

web/src/app/globals.css

@@ -322,6 +322,62 @@ button {
   cursor: pointer;
 }
 
+/* Prose 修复 — 确保有序/无序列表样式正确 */
+.prose ol {
+  list-style-type: decimal !important;
+  list-style-position: outside !important;
+  padding-left: 1.625em !important;
+  margin-top: 0;
+  margin-bottom: 0;
+}
+.prose ol li {
+  display: list-item !important;
+  margin-top: 0.25em;
+  margin-bottom: 0.25em;
+}
+.prose ol li::before {
+  content: none !important;
+}
+.prose ul {
+  list-style-type: disc !important;
+  list-style-position: outside !important;
+  padding-left: 1.625em !important;
+}
+.prose ul li {
+  display: list-item !important;
+  margin-top: 0.25em;
+  margin-bottom: 0.25em;
+}
+.prose ul li::before {
+  content: none !important;
+}
+.prose p {
+  margin-top: 0.5em;
+  margin-bottom: 0.5em;
+}
+.prose h1, .prose h2, .prose h3, .prose h4 {
+  font-family: inherit;
+  letter-spacing: normal;
+  margin-top: 1em;
+  margin-bottom: 0.5em;
+  font-weight: 600;
+}
+.prose h2 { font-size: 1.25em; }
+.prose h3 { font-size: 1.1em; }
+.prose h4 { font-size: 1em; }
+.prose strong { font-weight: 700; }
+.prose em { font-style: italic; }
+.prose blockquote {
+  border-left: 3px solid hsl(var(--border));
+  padding-left: 1em;
+  color: hsl(var(--muted-foreground));
+  font-style: italic;
+}
+.prose hr {
+  border-color: hsl(var(--border));
+  margin: 1.5em 0;
+}
+
 @media (prefers-reduced-motion: reduce) {
   *, *::before, *::after {
     animation-duration: 0.01ms !important;

web/src/components/chat/chat-interface.tsx

@@ -15,7 +15,7 @@ import { knowledgeBaseAPI } from "@/lib/api";
 export default function ChatInterface() {
   const [inputMessage, setInputMessage] = useState("");
   const [isComposing, setIsComposing] = useState(false);
-  const [selectedModel, setSelectedModel] = useState("deepseek-ai/DeepSeek-V3");
+  const [selectedModel, setSelectedModel] = useState("deepseek-chat");
   const [selectedKnowledgeBases, setSelectedKnowledgeBases] = useState<string[]>([]);
   const [systemKnowledgeBases, setSystemKnowledgeBases] = useState<KnowledgeBase[]>([]);
   const [userKnowledgeBases, setUserKnowledgeBases] = useState<KnowledgeBase[]>([]);

web/src/components/chat/model-selector.tsx

@@ -8,9 +8,10 @@ import {
   DropdownMenuItem,
   DropdownMenuTrigger,
   DropdownMenuSeparator,
+  DropdownMenuLabel,
 } from "@/components/ui/dropdown-menu";
 import { Badge } from "@/components/ui/badge";
-import { ChevronDown, Cpu, Zap, Sparkles } from "lucide-react";
+import { ChevronDown, Cpu, Zap, Sparkles, Brain } from "lucide-react";
 import { cn } from "@/lib/utils";
 
 export interface ModelOption {
@@ -18,6 +19,7 @@ export interface ModelOption {
   name: string;
   description: string;
   provider: string;
+  providerLabel: string;
   icon?: React.ComponentType<{ className?: string }>;
 }
 
@@ -27,31 +29,64 @@ interface ModelSelectorProps {
   className?: string;
 }
 
-const models: ModelOption[] = [
+const modelGroups = [
+  {
+    provider: "deepseek-official",
+    label: "DeepSeek 官方",
+    models: [
+      {
+        id: "deepseek-chat",
+        name: "DeepSeek-V3",
+        description: "DeepSeek 最新通用模型，速度快、能力强",
+        provider: "deepseek-official",
+        providerLabel: "DeepSeek",
+        icon: Sparkles,
+      },
+      {
+        id: "deepseek-reasoner",
+        name: "DeepSeek-R1",
+        description: "深度推理模型，适合复杂分析任务",
+        provider: "deepseek-official",
+        providerLabel: "DeepSeek",
+        icon: Brain,
+      },
+    ],
+  },
+  {
+    provider: "siliconflow",
+    label: "SiliconFlow（硅基流动）",
+    models: [
       {
         id: "deepseek-ai/DeepSeek-V3",
         name: "DeepSeek-V3",
-    description: "DeepSeek 最新版本，强大的推理能力",
+        description: "通过硅基流动调用，稳定的推理能力",
-    provider: "DeepSeek",
+        provider: "siliconflow",
+        providerLabel: "硅基流动",
         icon: Sparkles,
       },
       {
         id: "Qwen/QwQ-32B",
         name: "QwQ-32B",
-    description: "Qwen 量子化模型，高效推理",
+        description: "Qwen 推理模型，高效准确",
-    provider: "Qwen",
+        provider: "siliconflow",
+        providerLabel: "硅基流动",
         icon: Zap,
       },
+    ],
+  },
 ];
 
+const allModels = modelGroups.flatMap((g) => g.models);
+
 export default function ModelSelector({
   selectedModel,
   onModelChange,
-  className 
+  className,
 }: ModelSelectorProps) {
   const [isOpen, setIsOpen] = useState(false);
 
-  const selectedModelData = models.find(model => model.id === selectedModel) || models[0];
+  const selectedModelData =
+    allModels.find((model) => model.id === selectedModel) || allModels[0];
   const IconComponent = selectedModelData.icon || Cpu;
 
   return (
@@ -68,16 +103,21 @@ export default function ModelSelector({
         >
           <IconComponent className="w-4 h-4 mr-2" />
           <span className="hidden sm:inline">{selectedModelData.name}</span>
-          <span className="sm:hidden">{selectedModelData.name.split(' ')[0]}</span>
+          <span className="sm:hidden">{selectedModelData.name.split(" ")[0]}</span>
           <ChevronDown className="w-3 h-3 ml-1 opacity-50" />
         </Button>
       </DropdownMenuTrigger>
-      <DropdownMenuContent align="start" className="w-64">
+      <DropdownMenuContent align="start" className="w-72">
         <div className="px-2 py-1.5 text-xs font-medium text-muted-foreground">
           选择AI模型
         </div>
         <DropdownMenuSeparator />
-        {models.map((model) => {
+        {modelGroups.map((group) => (
+          <div key={group.provider}>
+            <DropdownMenuLabel className="text-xs text-muted-foreground/70 font-normal px-2 pt-2">
+              {group.label}
+            </DropdownMenuLabel>
+            {group.models.map((model) => {
               const ModelIcon = model.icon || Cpu;
               const isSelected = model.id === selectedModel;
 
@@ -106,15 +146,12 @@ export default function ModelSelector({
                     <p className="text-xs text-muted-foreground mt-0.5">
                       {model.description}
                     </p>
-                <div className="flex items-center space-x-1 mt-1">
-                  <span className="text-xs text-muted-foreground">
-                    {model.provider}
-                  </span>
-                </div>
                   </div>
                 </DropdownMenuItem>
               );
             })}
+          </div>
+        ))}
       </DropdownMenuContent>
     </DropdownMenu>
   );

web/src/store/chat.ts

@@ -255,6 +255,36 @@ export const useChatStore = create<ChatStore>((set, get) => ({
     let chunkCount = 0;
     let totalChars = 0;
     let thinkingSteps: ThinkingStep[] = [];
+    let contentBuffer = '';
+    let thinkingBuffer: ThinkingStep[] = [];
+    let rafId: number | null = null;
+
+    const flushBuffer = () => {
+      rafId = null;
+      const bufferedContent = contentBuffer;
+      const bufferedThinking = thinkingBuffer.length > 0 ? [...thinkingBuffer] : null;
+      contentBuffer = '';
+      thinkingBuffer = [];
+
+      if (!bufferedContent && !bufferedThinking) return;
+
+      set((state) => ({
+        messages: state.messages.map(msg => {
+          if (msg.id !== assistantMessage.id) return msg;
+          return {
+            ...msg,
+            ...(bufferedContent ? { content: msg.content + bufferedContent } : {}),
+            ...(bufferedThinking ? { thinking: [...(msg.thinking || []), ...bufferedThinking] } : {}),
+          };
+        }),
+      }));
+    };
+
+    const scheduleFlush = () => {
+      if (rafId === null) {
+        rafId = requestAnimationFrame(flushBuffer);
+      }
+    };
 
     try {
       await chatAPI.streamMessage(
@@ -264,59 +294,37 @@ export const useChatStore = create<ChatStore>((set, get) => ({
         knowledgeBaseIds,
         abortController.signal, // 新增参数
         (chunk: string) => {
-          // 解析chunk
+          if (chunk.startsWith('{')) {
             try {
               const data = JSON.parse(chunk);
 
               if (data.type === 'thinking') {
-              // 处理思考过程
+                const step: ThinkingStep = {
-              thinkingSteps.push({
                   stage: data.stage,
                   message: data.message,
                   doc_count: data.doc_count,
                   time: data.time
-              });
+                };
-              
+                thinkingSteps.push(step);
-              // 立即更新思考过程到UI
+                thinkingBuffer.push(step);
-              requestAnimationFrame(() => {
+                scheduleFlush();
-                set((state) => ({
+                return;
-                  messages: state.messages.map(msg =>
-                    msg.id === assistantMessage.id
-                      ? { ...msg, thinking: [...thinkingSteps] }
-                      : msg
-                  ),
-                }));
-              });
               } else if (data.type === 'chunk') {
-              // 处理内容chunk
                 chunkCount++;
                 totalChars += data.content.length;
-
+                contentBuffer += data.content;
-              requestAnimationFrame(() => {
+                scheduleFlush();
-                set((state) => ({
+                return;
-                  messages: state.messages.map(msg =>
-                    msg.id === assistantMessage.id
-                      ? { ...msg, content: msg.content + data.content }
-                      : msg
-                  ),
-                }));
-              });
               }
             } catch (e) {
-            // 向后兼容：纯文本chunk
+              // JSON 解析失败，按纯文本处理
+            }
+          }
+
           chunkCount++;
           totalChars += chunk.length;
-
+          contentBuffer += chunk;
-            requestAnimationFrame(() => {
+          scheduleFlush();
-              set((state) => ({
-                messages: state.messages.map(msg =>
-                  msg.id === assistantMessage.id
-                    ? { ...msg, content: msg.content + chunk }
-                    : msg
-                ),
-              }));
-            });
-          }
         },
         (_sessionId: number, messageId?: number, userMessageId?: number) => {
           if (messageId) {
@@ -438,10 +446,37 @@ export const useChatStore = create<ChatStore>((set, get) => ({
       abortController,
     });
 
-    // 判断模式：检查之前的消息中是否有知识库相关内容
     const mode = "normal";
 
     let thinkingSteps: any[] = [];
+    let regenContentBuffer = '';
+    let regenThinkingBuffer: any[] = [];
+    let regenRafId: number | null = null;
+
+    const regenFlush = () => {
+      regenRafId = null;
+      const bc = regenContentBuffer;
+      const bt = regenThinkingBuffer.length > 0 ? [...regenThinkingBuffer] : null;
+      regenContentBuffer = '';
+      regenThinkingBuffer = [];
+      if (!bc && !bt) return;
+      set((state) => ({
+        messages: state.messages.map(msg => {
+          if (msg.id !== assistantPlaceholder.id) return msg;
+          return {
+            ...msg,
+            ...(bc ? { content: msg.content + bc } : {}),
+            ...(bt ? { thinking: [...(msg.thinking || []), ...bt] } : {}),
+          };
+        }),
+      }));
+    };
+
+    const regenSchedule = () => {
+      if (regenRafId === null) {
+        regenRafId = requestAnimationFrame(regenFlush);
+      }
+    };
 
     try {
       await chatAPI.streamMessage(
@@ -451,48 +486,32 @@ export const useChatStore = create<ChatStore>((set, get) => ({
         undefined, // knowledgeBaseIds
         abortController.signal,
         (chunk: string) => {
-          // 处理 chunk
+          if (chunk.startsWith('{')) {
             try {
               const data = JSON.parse(chunk);
               if (data.type === 'thinking') {
-              thinkingSteps.push({
+                const step = {
                   stage: data.stage,
                   message: data.message,
                   doc_count: data.doc_count,
                   time: data.time
-              });
+                };
-              requestAnimationFrame(() => {
+                thinkingSteps.push(step);
-                set((state) => ({
+                regenThinkingBuffer.push(step);
-                  messages: state.messages.map(msg =>
+                regenSchedule();
-                    msg.id === assistantPlaceholder.id
+                return;
-                      ? { ...msg, thinking: [...thinkingSteps] }
-                      : msg
-                  ),
-                }));
-              });
               } else if (data.type === 'chunk') {
-              requestAnimationFrame(() => {
+                regenContentBuffer += data.content;
-                set((state) => ({
+                regenSchedule();
-                  messages: state.messages.map(msg =>
+                return;
-                    msg.id === assistantPlaceholder.id
-                      ? { ...msg, content: msg.content + data.content }
-                      : msg
-                  ),
-                }));
-              });
               }
             } catch {
-            // 纯文本 chunk
+              // JSON 解析失败，按纯文本处理
-            requestAnimationFrame(() => {
-              set((state) => ({
-                messages: state.messages.map(msg =>
-                  msg.id === assistantPlaceholder.id
-                    ? { ...msg, content: msg.content + chunk }
-                    : msg
-                ),
-              }));
-            });
             }
+          }
+
+          regenContentBuffer += chunk;
+          regenSchedule();
         },
         (_sessionId: number, messageId?: number) => {
           // onComplete: 只替换 assistant 占位符 ID，不重新加载