feat(workflow): enhance message stages and image routing

backend/services/pipecat/pipeline.py

@@ -173,6 +173,35 @@ def _image_data_uri(frame: UserImageRawFrame) -> str:
     return f"data:image/jpeg;base64,{encoded}"
 
 
+def _multimodal_user_input_frame(
+    image_frame: UserImageRawFrame,
+    prompt_text: str,
+) -> LLMMessagesAppendFrame:
+    """Submit an explicit camera capture through the normal user-turn path.
+
+    ``UserImageRawFrame`` is appended by Pipecat's assistant-side aggregator,
+    which pushes context upstream directly into the LLM.  Workflow routing sits
+    on the downstream user-turn path, so queuing the raw frame would let the
+    Agent see the image while skipping edge evaluation.  A standard multimodal
+    user message keeps text and image turns on the same routing path.
+    """
+    return LLMMessagesAppendFrame(
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": prompt_text},
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": _image_data_uri(image_frame)},
+                    },
+                ],
+            }
+        ],
+        run_llm=True,
+    )
+
+
 async def _analyze_image_with_vision_model(
     cfg: AssistantConfig,
     frame: UserImageRawFrame,
@@ -783,10 +812,12 @@ async def run_pipeline(
             raise ValueError("等待摄像头视频帧超时") from exc
 
         if native_vision:
-            image_frame.text = value.prompt_text
-            image_frame.append_to_context = True
-            image_frame.request = None
-            await worker.queue_frame(image_frame)
+            input_frame = await asyncio.to_thread(
+                _multimodal_user_input_frame,
+                image_frame,
+                value.prompt_text,
+            )
+            await worker.queue_frame(input_frame)
             return
 
         try: