feat: support uploaded images in debug voice preview

Allow paste/drag temporary image assets so vision turns work without a live camera frame.

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/services/pipecat/pipeline.py

@@ -27,6 +27,7 @@ from services.pipecat.service_factory import (
 )
 from db.session import SessionLocal
 from services.knowledge import search as search_knowledge
+from services.input_assets import consume_input_image
 from services.client_tools import ClientToolBroker
 from services.tool_policy import policy_for_tool
 from services.vision import (
@@ -36,6 +37,7 @@ from services.vision import (
     config_with_main_llm_as_vision,
     config_with_vision_resource,
     image_data_uri,
+    image_frame_from_jpeg,
     image_jpeg_bytes,
 )
 from services.workflow_engine import WorkflowEngine
@@ -734,7 +736,7 @@ async def run_pipeline(
     greeting = await brain.greeting(cfg)
 
     async def submit_user_input(value: UserInput) -> None:
-        if not value.has_camera_frame:
+        if not value.has_image:
             if not value.run_immediately:
                 brain.record_user_message(value.text)
             await worker.queue_frame(
@@ -749,28 +751,49 @@ async def run_pipeline(
             raise ValueError("P0 图片输入必须立即触发回复")
         if not vision_enabled:
             raise ValueError("当前助手未启用视觉能力")
-        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")
         native_vision = active_vision_uses_main_llm()
         analysis_cfg = None if native_vision else active_vision_config()
 
-        request = UserImageRequestFrame(
-            user_id=user_id,
-            text=value.prompt_text,
-            append_to_context=False,
-        )
-        try:
-            image_frame = await vision_capture.request_image(llm, request)
-        except asyncio.TimeoutError as exc:
-            raise ValueError("等待摄像头视频帧超时") from exc
+        if value.image_asset_token:
+            try:
+                image_bytes = await asyncio.to_thread(
+                    consume_input_image,
+                    value.image_asset_token,
+                )
+                image_frame = await asyncio.to_thread(
+                    image_frame_from_jpeg,
+                    image_bytes,
+                )
+            except ValueError:
+                raise
+            except Exception as exc:
+                raise ValueError("上传图片暂时无法读取") from exc
+        else:
+            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")
+            request = UserImageRequestFrame(
+                user_id=user_id,
+                text=value.prompt_text,
+                append_to_context=False,
+            )
+            try:
+                image_frame = await vision_capture.request_image(llm, request)
+            except asyncio.TimeoutError as exc:
+                raise ValueError("等待摄像头视频帧超时") from exc
+            image_bytes = await asyncio.to_thread(image_jpeg_bytes, image_frame)
 
         if recorder:
-            image_bytes = await asyncio.to_thread(image_jpeg_bytes, image_frame)
             recorder.record_image_later(
                 image_bytes,
                 input_id=value.input_id,
                 timestamp=time_now_iso8601(),
+                content=value.text,
+                source=(
+                    "uploaded_asset"
+                    if value.image_asset_token
+                    else "camera_capture"
+                ),
             )
 
         if native_vision: