refactor: remove legacy think tags

core/main_terminal.py

@@ -696,25 +696,15 @@ class MainTerminal:
             
             # ===== 统一保存到对话历史（关键修复） =====
             
-            # 1. 构建完整的assistant消息内容
+            # 1. 构建助手回复内容（思考内容通过 reasoning_content 存储）
-            assistant_content_parts = []
+            assistant_content = final_response or "已完成操作。"
-            
-            # 添加思考内容
-            if final_thinking:
-                assistant_content_parts.append(f"<think>\n{final_thinking}\n</think>")
-            
-            # 添加回复内容
-            if final_response:
-                assistant_content_parts.append(final_response)
-            
-            # 合并内容
-            assistant_content = "\n".join(assistant_content_parts) if assistant_content_parts else "已完成操作。"
             
             # 2. 保存assistant消息（包含tool_calls但不包含结果）
             self.context_manager.add_conversation(
                 "assistant",
                 assistant_content,
-                collected_tool_calls if collected_tool_calls else None
+                collected_tool_calls if collected_tool_calls else None,
+                reasoning_content=final_thinking or None
             )
             
             # 3. 保存独立的tool消息

doc/frontend/legacy_ui_spec.md

@@ -77,7 +77,7 @@
      - `tool`：卡片包含工具名、图标（`TOOL_ICON_MAP`），参数快照/状态消息/结果（JSON、plain text、命令输出 `pre`），流式时加 loading 条；`action.tool.awaiting_content` 控制“等待正文”占位。工具 `status` 颜色区分 `preparing/running/completed/error`。
      - `append_payload`/`modify_payload`：专门显示文件路径、强制标志、成功/失败行数。
      - `system`：用于 `role=system` 文本，折叠块 `system-message`。
-- **特殊处理**：`renderHistoryMessages()` 会解析 `<think>` 块、自定义 metadata（`append_payload`, `modify_payload`），并避免 LLm 输出的 `<<<APPEND`/`<<<MODIFY>>>` 内容直接进入 UI。
+- **特殊处理**：`renderHistoryMessages()` 会读取消息的 `reasoning_content`、自定义 metadata（`append_payload`, `modify_payload`），并避免 LLM 输出的 `<<<APPEND`/`<<<MODIFY>>>` 内容直接进入 UI。
 - **滚动策略**：`messagesArea` 的 scroll 事件更新 `userScrolling/autoScrollEnabled`，`scroll-lock-toggle` 按钮切换锁定；思考块单独监听 `scroll` 以保持 pinned 行为。
 
 #### 3.3 输入区（`.compact-input-area`）

static/src/app.ts

@@ -1387,23 +1387,7 @@ const appOptions = {
                         }
                         
                         const content = message.content || '';
-                        let reasoningText = (message.reasoning_content || '').trim();
+                        const reasoningText = (message.reasoning_content || '').trim();
-                        
-                        if (!reasoningText) {
-                            const thinkPatterns = [
-                                /<think>([\s\S]*?)<\/think>/g,
-                                /<thinking>([\s\S]*?)<\/thinking>/g
-                            ];
-                            
-                            let extracted = '';
-                            for (const pattern of thinkPatterns) {
-                                let match;
-                                while ((match = pattern.exec(content)) !== null) {
-                                    extracted += (match[1] || '').trim() + '\n';
-                                }
-                            }
-                            reasoningText = extracted.trim();
-                        }
                         
                         if (reasoningText) {
                             const blockId = `history-thinking-${Date.now()}-${Math.random().toString(36).slice(2)}`;
@@ -1426,15 +1410,7 @@ const appOptions = {
                         const isModifyMessage = message.name === 'modify_file';
                         const containsAppendMarkers = /<<<\s*(APPEND|MODIFY)/i.test(content || '') || /<<<END_\s*(APPEND|MODIFY)>>>/i.test(content || '');
                         
-                        let textContent = content;
+                        const textContent = content.trim();
-                        if (!message.reasoning_content) {
-                            textContent = textContent
-                                .replace(/<think>[\s\S]*?<\/think>/g, '')
-                                .replace(/<thinking>[\s\S]*?<\/thinking>/g, '')
-                                .trim();
-                        } else {
-                            textContent = textContent.trim();
-                        }
                             
                         if (appendPayloadMeta) {
                             currentAssistantMessage.actions.push({

sub_agent/core/main_terminal.py

@@ -592,25 +592,15 @@ class MainTerminal:
             
             # ===== 统一保存到对话历史（关键修复） =====
             
-            # 1. 构建完整的assistant消息内容
+            # 1. 构建助手回复内容（思考内容通过 reasoning_content 保存）
-            assistant_content_parts = []
+            assistant_content = final_response or "已完成操作。"
-            
-            # 添加思考内容
-            if final_thinking:
-                assistant_content_parts.append(f"<think>\n{final_thinking}\n</think>")
-            
-            # 添加回复内容
-            if final_response:
-                assistant_content_parts.append(final_response)
-            
-            # 合并内容
-            assistant_content = "\n".join(assistant_content_parts) if assistant_content_parts else "已完成操作。"
             
             # 2. 保存assistant消息（包含tool_calls但不包含结果）
             self.context_manager.add_conversation(
                 "assistant",
                 assistant_content,
-                collected_tool_calls if collected_tool_calls else None
+                collected_tool_calls if collected_tool_calls else None,
+                reasoning_content=final_thinking or None
             )
             
             # 3. 保存独立的tool消息

sub_agent/static/app.js

@@ -1984,28 +1984,18 @@ async function bootstrapApp() {
                         const appendPayloadMeta = metadata.append_payload;
                         const modifyPayloadMeta = metadata.modify_payload;
 
-                        const thinkPatterns = [/<think>([\s\S]*?)<\/think>/g, /<thinking>([\s\S]*?)<\/thinking>/g];
+                        const reasoningText = (message.reasoning_content || '').trim();
-                        let allThinkingContent = '';
+                        if (reasoningText) {
-                        for (const pattern of thinkPatterns) {
-                            let match;
-                            while ((match = pattern.exec(content)) !== null) {
-                                allThinkingContent += match[1].trim() + '\n';
-                            }
-                        }
-                        if (allThinkingContent.trim()) {
                             assistant.actions.push({
                                 id: `history-think-${Date.now()}-${Math.random()}`,
                                 type: 'thinking',
-                                content: allThinkingContent.trim(),
+                                content: reasoningText,
                                 streaming: false,
                                 timestamp: Date.now()
                             });
                         }
 
-                        let textContent = content
+                        const textContent = content.trim();
-                            .replace(/<think>[\s\S]*?<\/think>/g, '')
-                            .replace(/<thinking>[\s\S]*?<\/thinking>/g, '')
-                            .trim();
 
                         if (appendPayloadMeta) {
                             assistant.actions.push({

sub_agent/utils/api_client.py

@@ -409,10 +409,6 @@ class DeepSeekClient:
             # 构建助手消息 - 始终包含所有收集到的内容
             assistant_content_parts = []
             
-            # 添加思考内容（如果有）
-            if current_thinking:
-                assistant_content_parts.append(f"<think>\n{current_thinking}\n</think>")
-            
             # 添加正式回复内容（如果有）
             if full_response:
                 assistant_content_parts.append(full_response)

sub_agent/web_server.py

@@ -2543,7 +2543,6 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
     sender('ai_message_start', {})
     
     # 增量保存相关变量
-    has_saved_thinking = False  # 是否已保存思考内容
     accumulated_response = ""   # 累积的响应内容
     is_first_iteration = True   # 是否是第一次迭代
     
@@ -3348,13 +3347,6 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
                         sender('thinking_end', {'full_content': current_thinking})
                         await asyncio.sleep(0.1)
                     
-                        # ===== 增量保存：保存思考内容 =====
-                        if current_thinking and not has_saved_thinking and is_first_iteration:
-                            thinking_content = f"<think>\n{current_thinking}\n</think>"
-                            web_terminal.context_manager.add_conversation("assistant", thinking_content)
-                            has_saved_thinking = True
-                            debug_log(f"💾 增量保存：思考内容 ({len(current_thinking)} 字符)")
-                    
                     expecting_modify = bool(pending_modify) or bool(getattr(web_terminal, "pending_modify_request", None))
                     expecting_append = bool(pending_append) or bool(getattr(web_terminal, "pending_append_request", None))
 
@@ -3600,13 +3592,6 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
             sender('thinking_end', {'full_content': current_thinking})
             await asyncio.sleep(0.1)
         
-            # 保存思考内容
-            if current_thinking and not has_saved_thinking and is_first_iteration:
-                thinking_content = f"<think>\n{current_thinking}\n</think>"
-                web_terminal.context_manager.add_conversation("assistant", thinking_content)
-                has_saved_thinking = True
-                debug_log(f"💾 增量保存：延迟思考内容 ({len(current_thinking)} 字符)")
-        
         # 确保text_end事件被发送
         if text_started and text_has_content and not append_result["handled"] and not modify_result["handled"]:
             await flush_text_buffer(force=True)
@@ -3617,7 +3602,11 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
             
             # ===== 增量保存：保存当前轮次的文本内容 =====
             if full_response.strip():
-                web_terminal.context_manager.add_conversation("assistant", full_response)
+                web_terminal.context_manager.add_conversation(
+                    "assistant",
+                    full_response,
+                    reasoning_content=current_thinking or None
+                )
                 debug_log(f"💾 增量保存：文本内容 ({len(full_response)} 字符)")
         
         if append_result["handled"]:
@@ -3627,7 +3616,8 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
                 web_terminal.context_manager.add_conversation(
                     "assistant",
                     append_content_text,
-                    metadata=append_metadata
+                    metadata=append_metadata,
+                    reasoning_content=current_thinking or None
                 )
                 debug_log("💾 增量保存：追加正文快照")
                 
@@ -3703,7 +3693,8 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
                 web_terminal.context_manager.add_conversation(
                     "assistant",
                     modify_content_text,
-                    metadata=modify_metadata
+                    metadata=modify_metadata,
+                    reasoning_content=current_thinking or None
                 )
                 debug_log("💾 增量保存：修改正文快照")
                 
@@ -3809,9 +3800,6 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
         # 构建助手消息（用于API继续对话）
         assistant_content_parts = []
         
-        if current_thinking:
-            assistant_content_parts.append(f"<think>\n{current_thinking}\n</think>")
-        
         if full_response:
             assistant_content_parts.append(full_response)
         elif append_result["handled"] and append_result["assistant_content"]: