标题生成强制关闭 think 模式,回答速度指示器默认开启

- TitleMiddleware 给模型实例注入 extra_body 关 think(覆盖 Zhipu GLM 与 Qwen/vLLM 两套写法),并兜底剥离不闭合/残留的 <think> 标签,避免 GLM5 起标题时长时间思考与标签泄漏。
- 前端 streamSpeedMeterEnabled 默认改为 true,设置项描述同步为"默认开启"。

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

frontend-web/src/components/workspace/settings/appearance-settings-page.tsx

@@ -122,7 +122,7 @@ export function AppearanceSettingsPage() {
             <div className="text-sm font-medium">回答速度指示器</div>
             <p className="text-muted-foreground mt-1 max-w-md text-xs leading-relaxed">
               在生成回答时显示一个小胶囊，实时估算当前 token/s
-              并用绿到红的渐变条直观反映快慢。默认关闭。
+              并用绿到红的渐变条直观反映快慢。默认开启。
             </p>
           </div>
           <Switch

frontend-web/src/core/settings/local.ts

@@ -5,7 +5,7 @@ export const DEFAULT_LOCAL_SETTINGS: LocalSettings = {
     enabled: true,
   },
   display: {
-    streamSpeedMeterEnabled: false,
+    streamSpeedMeterEnabled: true,
   },
   context: {
     model_name: undefined,
@@ -29,7 +29,7 @@ export interface LocalSettings {
   display: {
     /**
      * Whether to render a small token/s speed meter while a streaming
-     * AI reply is in flight. Default off; controlled by the user from
+     * AI reply is in flight. Default on; controlled by the user from
      * the appearance settings page.
      */
     streamSpeedMeterEnabled: boolean;

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/title_middleware.py

@@ -105,8 +105,16 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
         return prompt, user_msg
 
     def _strip_think_tags(self, text: str) -> str:
-        """Remove <think>...</think> blocks emitted by reasoning models (e.g. minimax, DeepSeek-R1)."""
+        """Remove <think>...</think> blocks emitted by reasoning models (e.g. minimax, GLM, DeepSeek-R1).
-        return re.sub(r"<think>[\s\S]*?</think>", "", text, flags=re.IGNORECASE).strip()
+
+        Also handles malformed cases: an unclosed `<think>` (response truncated mid-reasoning)
+        or a stray `</think>` (open tag eaten by upstream parsing) — both can otherwise leak
+        raw reasoning into the title.
+        """
+        text = re.sub(r"<think>[\s\S]*?</think>", "", text, flags=re.IGNORECASE)
+        text = re.sub(r"<think>[\s\S]*$", "", text, flags=re.IGNORECASE)
+        text = re.sub(r"^[\s\S]*?</think>", "", text, flags=re.IGNORECASE)
+        return text.strip()
 
     def _parse_title(self, content: object) -> str:
         """Normalize model output into a clean title string."""
@@ -146,6 +154,36 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
         _, user_msg = self._build_title_prompt(state)
         return {"title": self._fallback_title(user_msg)}
 
+    def _force_disable_thinking(self, model: object) -> None:
+        """Force-disable thinking on the title model via `extra_body` overrides.
+
+        Title generation never needs reasoning. The model factory only emits
+        disable params when the model config declares `when_thinking_disabled`
+        or a matching `when_thinking_enabled` shape — many user-added reasoning
+        models (e.g. GLM-4.5/5) only set `supports_thinking: true` and would
+        otherwise keep thinking on, which is slow and leaks <think> tags into
+        the title prompt response. Injecting both common OpenAI-compatible
+        disable shapes covers Zhipu GLM (`thinking.type`) and Qwen/vLLM
+        (`chat_template_kwargs.enable_thinking`); other providers ignore
+        unknown body fields.
+        """
+        if not hasattr(model, "extra_body"):
+            return
+        current = getattr(model, "extra_body", None) or {}
+        merged_chat_template_kwargs = {
+            **(current.get("chat_template_kwargs") or {}),
+            "enable_thinking": False,
+        }
+        merged = {
+            **current,
+            "thinking": {**(current.get("thinking") or {}), "type": "disabled"},
+            "chat_template_kwargs": merged_chat_template_kwargs,
+        }
+        try:
+            model.extra_body = merged
+        except Exception:
+            logger.debug("Could not override extra_body to disable thinking on title model", exc_info=True)
+
     async def _agenerate_title_result(self, state: TitleMiddlewareState) -> dict | None:
         """Generate a title asynchronously and fall back locally on failure."""
         if not self._should_generate_title(state):
@@ -162,6 +200,7 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
                 model = create_chat_model(name=config.model_name, **model_kwargs)
             else:
                 model = create_chat_model(**model_kwargs)
+            self._force_disable_thinking(model)
             response = await model.ainvoke(prompt, config=self._get_runnable_config())
             title = self._parse_title(response.content)
             if title: