Enhance DuplexPipeline to track assistant activity and improve presence probe logic. Introduce a method to update assistant activity timestamp and modify presence probe conditions to consider both user and assistant activity. Update tests to validate new behavior and ensure contextual prompts are generated correctly.

engine/core/duplex_pipeline.py

@@ -382,6 +382,7 @@ class DuplexPipeline:
         self._last_llm_delta_emit_ms: float = 0.0
         now_ms = time.monotonic() * 1000.0
         self._last_user_activity_ms: float = now_ms
+        self._last_assistant_activity_ms: float = now_ms
         self._last_presence_probe_ms: float = 0.0
         self._presence_probe_attempts: int = 0
         self._presence_probe_seq: int = 0
@@ -947,6 +948,9 @@ class DuplexPipeline:
         self._last_user_activity_ms = time.monotonic() * 1000.0
         self._active_presence_probe_call_id = None
 
+    def _touch_assistant_activity(self) -> None:
+        self._last_assistant_activity_ms = time.monotonic() * 1000.0
+
     def _presence_probe_in_progress(self) -> bool:
         return bool(self._active_presence_probe_call_id)
 
@@ -970,7 +974,8 @@ class DuplexPipeline:
             return False
         if self.conversation.turn_count <= 0:
             return False
-        if now_ms - self._last_user_activity_ms < self._presence_probe_idle_ms():
+        last_activity_ms = max(self._last_user_activity_ms, self._last_assistant_activity_ms)
+        if now_ms - last_activity_ms < self._presence_probe_idle_ms():
             return False
         if self._last_presence_probe_ms > 0.0 and now_ms - self._last_presence_probe_ms < self._presence_probe_cooldown_ms():
             return False
@@ -984,12 +989,52 @@ class DuplexPipeline:
             return compact
         return f"{compact[:self._PRESENCE_PROBE_CONTEXT_CHARS]}..."
 
-    def _build_presence_probe_question(self) -> str:
+    async def _build_presence_probe_question(self) -> str:
         manual_question = str(self._runtime_presence_probe.get("question") or "").strip()
         if manual_question:
             return manual_question
 
         include_context = bool(self._runtime_presence_probe.get("includeContext", True))
+        if include_context and self.llm_service:
+            last_user = self._clip_presence_context(self.conversation.last_user_text or "")
+            last_assistant = self._clip_presence_context(self.conversation.last_assistant_text or "")
+            context_lines: List[str] = []
+            if last_user:
+                context_lines.append(f"用户：{last_user}")
+            if last_assistant:
+                context_lines.append(f"助手：{last_assistant}")
+            context_blob = "\n".join(context_lines).strip()
+            try:
+                generated = await self.llm_service.generate(
+                    [
+                        LLMMessage(
+                            role="system",
+                            content=(
+                                "你是语音助手。目标是在用户长时间静默时发起一次自然、礼貌、简短的在线确认。"
+                                "请只输出一句中文问句，不要解释，不要使用引号，不要使用 markdown。"
+                                "优先沿用最近上下文，不要机械重复固定模板。"
+                            ),
+                        ),
+                        LLMMessage(
+                            role="user",
+                            content=(
+                                "请基于以下最近对话，生成一句在线确认问句（不超过22个汉字）。\n"
+                                f"{context_blob or '（无明显上下文）'}"
+                            ),
+                        ),
+                    ],
+                    temperature=0.7,
+                    max_tokens=64,
+                )
+                cleaned = str(generated or "").strip().strip('"').strip("'")
+                if cleaned:
+                    cleaned = cleaned.replace("\n", " ").strip()
+                    if len(cleaned) > 60:
+                        cleaned = cleaned[:60]
+                    return cleaned
+            except Exception as exc:
+                logger.warning(f"Presence probe LLM question generation failed: {exc}")
+
         if include_context:
             last_assistant = self._clip_presence_context(self.conversation.last_assistant_text or "")
             if last_assistant:
@@ -1010,10 +1055,11 @@ class DuplexPipeline:
         self._last_presence_probe_ms = current_ms
         self._presence_probe_attempts += 1
 
-        question = self._build_presence_probe_question()
+        question = await self._build_presence_probe_question()
         probe_turn_id = self._start_turn()
         probe_response_id = self._start_response()
         try:
+            await self.conversation.add_assistant_turn(question)
             await self._send_event(
                 {
                     **ev(
@@ -1028,11 +1074,12 @@ class DuplexPipeline:
             )
             if self._tts_output_enabled():
                 await self._speak(question, audio_event_priority=30)
+            self._touch_assistant_activity()
 
             logger.info(
                 "[PresenceProbe] sent probe_id={} idle_ms={} question={}",
                 probe_id,
-                int(max(0.0, current_ms - self._last_user_activity_ms)),
+                int(max(0.0, current_ms - max(self._last_user_activity_ms, self._last_assistant_activity_ms))),
                 question,
             )
             return True
@@ -1300,6 +1347,7 @@ class DuplexPipeline:
         if self._tts_output_enabled() and not used_preloaded_audio:
             # Keep opener text ahead of opener voice start.
             await self._speak(greeting_to_speak, audio_event_priority=30)
+        self._touch_assistant_activity()
 
     async def _play_preloaded_opener_audio(self) -> bool:
         """
@@ -2564,6 +2612,7 @@ class DuplexPipeline:
             self._barge_in_silence_frames = 0
             self._current_response_id = None
             self._current_tts_id = None
+            self._touch_assistant_activity()
 
     async def _speak_sentence(
         self,

engine/tests/test_tool_call_flow.py

@@ -62,6 +62,12 @@ class _FakeLLM:
         self._rounds = rounds
         self._call_index = 0
 
+    async def generate(self, messages, temperature=0.7, max_tokens=None):
+        prompt = " ".join([str(getattr(m, "content", "")) for m in messages])
+        if "订单号" in prompt:
+            return "关于订单号这块，你还在线吗？"
+        return "你还在线吗？"
+
     async def generate_stream(self, _messages, temperature=0.7, max_tokens=None):
         idx = self._call_index
         self._call_index += 1
@@ -384,7 +390,9 @@ async def test_presence_probe_emits_contextual_direct_prompt(monkeypatch):
     )
     await pipeline._shutdown_presence_probe_task()
     await pipeline.conversation.add_assistant_turn("请把你的订单号告诉我，我继续帮你处理。")
-    pipeline._last_user_activity_ms = (time.monotonic() * 1000.0) - 8000.0
+    now_ms = time.monotonic() * 1000.0
+    pipeline._last_user_activity_ms = now_ms - 8000.0
+    pipeline._last_assistant_activity_ms = now_ms - 8000.0
 
     fired = await pipeline._run_presence_probe_once()
 
@@ -392,6 +400,7 @@ async def test_presence_probe_emits_contextual_direct_prompt(monkeypatch):
     probe_text_events = [e for e in events if e.get("type") == "assistant.response.final"]
     assert probe_text_events
     assert "订单号" in str(probe_text_events[-1].get("text") or "")
+    assert "订单号" in str(pipeline.conversation.last_assistant_text or "")
     assert any(e.get("type") == "output.audio.start" for e in events)
     assert not any(e.get("type") == "assistant.tool_call" for e in events)
 
@@ -412,7 +421,9 @@ async def test_presence_probe_respects_max_prompts_limit(monkeypatch):
     )
     await pipeline._shutdown_presence_probe_task()
     await pipeline.conversation.add_assistant_turn("我们继续。")
-    pipeline._last_user_activity_ms = (time.monotonic() * 1000.0) - 8000.0
+    now_ms = time.monotonic() * 1000.0
+    pipeline._last_user_activity_ms = now_ms - 8000.0
+    pipeline._last_assistant_activity_ms = now_ms - 8000.0
 
     first_fired = await pipeline._run_presence_probe_once()
     second_fired = await pipeline._run_presence_probe_once(
@@ -441,7 +452,9 @@ async def test_presence_probe_text_mode_emits_text_only(monkeypatch):
     )
     await pipeline._shutdown_presence_probe_task()
     await pipeline.conversation.add_assistant_turn("我们继续。")
-    pipeline._last_user_activity_ms = (time.monotonic() * 1000.0) - 8000.0
+    now_ms = time.monotonic() * 1000.0
+    pipeline._last_user_activity_ms = now_ms - 8000.0
+    pipeline._last_assistant_activity_ms = now_ms - 8000.0
 
     fired = await pipeline._run_presence_probe_once()
 
@@ -451,6 +464,28 @@ async def test_presence_probe_text_mode_emits_text_only(monkeypatch):
     assert not any(e.get("type") == "output.audio.start" for e in events)
 
 
+@pytest.mark.asyncio
+async def test_presence_probe_does_not_count_assistant_speaking_time_as_idle(monkeypatch):
+    pipeline, _events = _build_pipeline(monkeypatch, [[LLMStreamEvent(type="done")]])
+    pipeline.apply_runtime_overrides(
+        {
+            "presenceProbe": {
+                "enabled": True,
+                "idleSeconds": 10,
+                "cooldownSeconds": 5,
+                "maxPrompts": 1,
+            }
+        }
+    )
+    await pipeline._shutdown_presence_probe_task()
+    await pipeline.conversation.add_assistant_turn("我们继续。")
+    now_ms = time.monotonic() * 1000.0
+    pipeline._last_user_activity_ms = now_ms - 30_000.0
+    pipeline._last_assistant_activity_ms = now_ms - 2_000.0
+
+    assert pipeline._presence_probe_due(now_ms) is False
+
+
 @pytest.mark.asyncio
 async def test_server_calculator_emits_tool_result(monkeypatch):
     pipeline, events = _build_pipeline(

web/pages/Assistants.tsx

@@ -3815,8 +3815,15 @@ export const DebugDrawer: React.FC<{
             if (!finalText) return prev;
             const last = prev[prev.length - 1];
             if (last?.role === 'model') {
-              if (last.text === finalText) return prev;
+              const sameResponse = Boolean(
-              if (finalText.startsWith(last.text) || last.text.startsWith(finalText)) {
+                responseId
+                && last.responseId
+                && responseId === last.responseId
+              );
+              const bothWithoutResponseId = !responseId && !last.responseId;
+              const canMergeTail = sameResponse || bothWithoutResponseId;
+              if (canMergeTail && last.text === finalText) return prev;
+              if (canMergeTail && (finalText.startsWith(last.text) || last.text.startsWith(finalText))) {
                 const next = [...prev];
                 next[next.length - 1] = { ...last, text: finalText };
                 return next;