api has llm response event

core/duplex_pipeline.py

@@ -426,6 +426,15 @@ class DuplexPipeline:
                 sentence_buffer += text_chunk
                 await self.conversation.update_assistant_text(text_chunk)
                 
+                # Send LLM response streaming event to client
+                await self.transport.send_event({
+                    "event": "llmResponse",
+                    "trackId": self.session_id,
+                    "text": text_chunk,
+                    "isFinal": False,
+                    "timestamp": self._get_timestamp_ms()
+                })
+                
                 # Check for sentence completion - synthesize immediately for low latency
                 while any(end in sentence_buffer for end in sentence_ends):
                     # Find first sentence end
@@ -454,6 +463,16 @@ class DuplexPipeline:
                     else:
                         break
             
+            # Send final LLM response event
+            if full_response and not self._interrupt_event.is_set():
+                await self.transport.send_event({
+                    "event": "llmResponse",
+                    "trackId": self.session_id,
+                    "text": full_response,
+                    "isFinal": True,
+                    "timestamp": self._get_timestamp_ms()
+                })
+            
             # Speak any remaining text
             if sentence_buffer.strip() and not self._interrupt_event.is_set():
                 if not first_audio_sent: