feat: route workflow image inputs natively

backend/services/pipecat/pipeline.py

@@ -141,6 +141,25 @@ def _vision_uses_main_llm(cfg: AssistantConfig) -> bool:
     return not cfg.vision_model_resource_id and cfg.llm_support_image_input
 
 
+def _workflow_vision_uses_main_llm(
+    cfg: AssistantConfig,
+    scope: dict[str, Any],
+) -> bool:
+    """Resolve native versus auxiliary vision for the active Workflow Agent."""
+    if not scope.get("enabled"):
+        raise ValueError("当前 Workflow Agent 节点未启用视觉能力")
+    if scope.get("vision_model_resource_id"):
+        return False
+
+    llm_resource_id = str(scope.get("llm_resource_id") or "")
+    resource = cfg.workflow_model_resources.get(llm_resource_id)
+    if not resource:
+        raise ValueError(f"当前 Workflow Agent 的 LLM 资源未加载:{llm_resource_id}")
+    if not resource.support_image_input:
+        raise ValueError("当前 Workflow Agent 的 LLM 不支持图片输入")
+    return True
+
+
 def _image_data_uri(frame: UserImageRawFrame) -> str:
     if not frame.format:
         raise ValueError("摄像头图片帧缺少 format,无法编码给视觉模型")
@@ -388,6 +407,11 @@ async def run_pipeline(
         "vision_model_resource_id": None,
         "llm_resource_id": None,
     }
+
+    def active_vision_uses_main_llm() -> bool:
+        if cfg.type == "workflow":
+            return _workflow_vision_uses_main_llm(cfg, workflow_vision_scope)
+        return vision_native_mode
     vision_schema = FunctionSchema(
         name=VISION_TOOL_NAME,
         description=(
@@ -510,6 +534,27 @@ async def run_pipeline(
         raise ValueError(f"视觉模型资源未加载:{vision_resource_id}")
 
     if cfg.type == "workflow" and vision_enabled:
+        async def native_flow_fetch_user_image(params: FunctionCallParams) -> None:
+            question = str(params.arguments.get("question") or "请描述当前画面。")
+            user_id = vision_state.get("client_id")
+            if not user_id:
+                await params.result_callback(
+                    {
+                        "status": "no_video_client",
+                        "message": "当前还没有可用的摄像头视频流。",
+                    }
+                )
+                return
+            request = UserImageRequestFrame(
+                user_id=user_id,
+                text=question,
+                append_to_context=True,
+                function_name=params.function_name,
+                tool_call_id=params.tool_call_id,
+                result_callback=params.result_callback,
+            )
+            await params.llm.push_frame(request, FrameDirection.UPSTREAM)
+
         async def flow_fetch_user_image(args, _flow_manager):
             if not workflow_vision_scope.get("enabled"):
                 return {
@@ -547,6 +592,20 @@ async def run_pipeline(
                 logger.warning(f"Workflow 视觉理解失败:{exc}")
                 return {"status": "error", "message": "视觉理解暂时不可用。"}
 
+        # ConfiguredFlowManager keeps the Flows handler for auxiliary models,
+        # but uses Pipecat's native function-call image path when the active
+        # Agent selected its own visual-capable LLM.
+        setattr(
+            flow_fetch_user_image,
+            "_workflow_native_vision_handler",
+            native_flow_fetch_user_image,
+        )
+        setattr(
+            flow_fetch_user_image,
+            "_workflow_native_vision_enabled",
+            active_vision_uses_main_llm,
+        )
+
         workflow_vision_function = FlowsFunctionSchema(
             name=VISION_TOOL_NAME,
             description=vision_schema.description,
@@ -710,7 +769,8 @@ async def run_pipeline(
         user_id = vision_state.get("client_id")
         if not user_id:
             raise ValueError("当前没有可用的摄像头视频流")
-        analysis_cfg = None if vision_native_mode else active_vision_config()
+        native_vision = active_vision_uses_main_llm()
+        analysis_cfg = None if native_vision else active_vision_config()
 
         request = UserImageRequestFrame(
             user_id=user_id,
@@ -722,7 +782,7 @@ async def run_pipeline(
         except asyncio.TimeoutError as exc:
             raise ValueError("等待摄像头视频帧超时") from exc
 
-        if vision_native_mode:
+        if native_vision:
             image_frame.text = value.prompt_text
             image_frame.append_to_context = True
             image_frame.request = None