diff --git a/engine/core/duplex_pipeline.py b/engine/core/duplex_pipeline.py
index ad42d13..c99cba7 100644
--- a/engine/core/duplex_pipeline.py
+++ b/engine/core/duplex_pipeline.py
@@ -409,12 +409,13 @@ class DuplexPipeline:
 
         prompt_hint = (self._runtime_greeting or "").strip()
         system_prompt = (
-            "You generate one concise opener for a live voice call assistant. "
-            "Return plain text only, no quotes, no markdown, one sentence."
+            "你是语音通话助手的开场白生成器。"
+            "请只输出一句自然、简洁、友好的中文开场白。"
+            "不要使用引号，不要使用 markdown，不要加解释。"
         )
-        user_prompt = "Generate a friendly opening line (max 25 words)."
+        user_prompt = "请生成一句中文开场白（不超过25个汉字）。"
         if prompt_hint:
-            user_prompt += f" Style hint: {prompt_hint}"
+            user_prompt += f" 风格提示：{prompt_hint}"
 
         try:
             generated = await self.llm_service.generate(
diff --git a/web/pages/Assistants.tsx b/web/pages/Assistants.tsx
index ac11f7b..09820e6 100644
--- a/web/pages/Assistants.tsx
+++ b/web/pages/Assistants.tsx
@@ -1632,11 +1632,10 @@ export const DebugDrawer: React.FC<{
         lastUserFinalRef.current = '';
         setWsError('');
         closeWs();
-        if (textTtsEnabled) await ensureAudioContext();
-        await ensureWsSession();
-        await startVoiceCapture();
+      if (textTtsEnabled) await ensureAudioContext();
+      await ensureWsSession();
+      await startVoiceCapture();
       setCallStatus('active');
-      setMessages([]);
     } catch (e) {
       console.error(e);
       stopVoiceCapture();
@@ -1707,7 +1706,6 @@ export const DebugDrawer: React.FC<{
       if (textTtsEnabled) await ensureAudioContext();
       await ensureWsSession();
       setTextSessionStarted(true);
-      setMessages([]);
     } catch (e) {
       console.error(e);
       setWsStatus('error');