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,