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/tests/test_conversation_history.py

@@ -139,6 +139,8 @@ class ConversationRecorderTest(unittest.IsolatedAsyncioTestCase):
                 input_id="input_photo",
                 timestamp="2026-08-05T10:00:00+08:00",
                 mime_type="image/jpeg",
+                content="帮我看看",
+                source="uploaded_asset",
             )
 
         self.assertTrue(session.committed)
@@ -146,7 +148,9 @@ class ConversationRecorderTest(unittest.IsolatedAsyncioTestCase):
         message, artifact = session.added
         self.assertEqual(message.content_type, "image")
         self.assertEqual(message.role, "user")
+        self.assertEqual(message.content, "帮我看看")
         self.assertEqual(message.extra["input_id"], "input_photo")
+        self.assertEqual(message.extra["source"], "uploaded_asset")
         self.assertEqual(artifact.message_id, message.id)
         self.assertEqual(artifact.kind, "image")
         self.assertEqual(artifact.size_bytes, len(b"jpeg-data"))

backend/tests/test_input_assets.py

@@ -0,0 +1,58 @@
+from __future__ import annotations
+
+import unittest
+from io import BytesIO
+from unittest.mock import patch
+
+from PIL import Image
+
+from services.input_assets import consume_input_image, store_input_image
+
+
+def png_bytes(size: tuple[int, int] = (32, 24)) -> bytes:
+    output = BytesIO()
+    Image.new("RGBA", size, (20, 80, 160, 180)).save(output, format="PNG")
+    return output.getvalue()
+
+
+class InputAssetTests(unittest.TestCase):
+    def test_store_normalizes_and_consume_removes_temporary_object(self):
+        with (
+            patch("services.input_assets.put_object") as put_object,
+            patch("services.input_assets.get_object") as get_object,
+            patch("services.input_assets.delete_object") as delete_object,
+        ):
+            stored = store_input_image(png_bytes())
+
+            key, normalized, mime_type = put_object.call_args.args
+            self.assertTrue(key.startswith("conversation-inputs/"))
+            self.assertEqual(mime_type, "image/jpeg")
+            self.assertTrue(normalized.startswith(b"\xff\xd8"))
+            self.assertEqual((stored.width, stored.height), (32, 24))
+            self.assertEqual(stored.size_bytes, len(normalized))
+
+            get_object.return_value = normalized
+            consumed = consume_input_image(stored.token)
+
+        self.assertEqual(consumed, normalized)
+        get_object.assert_called_once_with(key)
+        delete_object.assert_called_once_with(key)
+
+    def test_tampered_token_is_rejected_before_storage_read(self):
+        with patch("services.input_assets.put_object"):
+            stored = store_input_image(png_bytes())
+        tampered = f"{stored.token}x"
+
+        with patch("services.input_assets.get_object") as get_object:
+            with self.assertRaisesRegex(ValueError, "签名无效"):
+                consume_input_image(tampered)
+
+        get_object.assert_not_called()
+
+    def test_non_image_is_rejected(self):
+        with self.assertRaisesRegex(ValueError, "可识别的图片"):
+            store_input_image(b"not an image")
+
+
+if __name__ == "__main__":
+    unittest.main()

backend/tests/test_user_input.py

@@ -32,7 +32,7 @@ class UserInputParserTests(unittest.TestCase):
         self.assertIsNotNone(value)
         self.assertEqual(value.text, "帮我看看")
         self.assertTrue(value.has_camera_frame)
-        self.assertEqual(value.transcript_text, "帮我看看")
+        self.assertEqual(value.transcript_text, "")
 
     def test_rejects_legacy_and_unsupported_image_sources(self):
         self.assertIsNone(parse_user_input({"type": "user-text", "text": "旧协议"}))
@@ -46,14 +46,40 @@ class UserInputParserTests(unittest.TestCase):
                         {
                             "type": "input_image",
                             "source": {
-                                "type": "uploaded_asset",
-                                "asset_id": "asset_1",
+                                "type": "remote_url",
+                                "url": "https://example.com/image.jpg",
                             },
                         }
                     ],
                 }
             )
 
+    def test_parses_uploaded_image_asset(self):
+        value = parse_user_input(
+            {
+                "type": "user-input",
+                "schema_version": 1,
+                "input_id": "input_upload",
+                "parts": [
+                    {"type": "input_text", "text": "这是什么？"},
+                    {
+                        "type": "input_image",
+                        "source": {
+                            "type": "uploaded_asset",
+                            "asset_token": "signed-token",
+                        },
+                    },
+                ],
+            }
+        )
+
+        self.assertIsNotNone(value)
+        self.assertTrue(value.has_image)
+        self.assertFalse(value.has_camera_frame)
+        self.assertEqual(value.image_asset_token, "signed-token")
+        self.assertEqual(value.prompt_text, "这是什么？")
+        self.assertEqual(value.transcript_text, "")
+
     def test_image_only_input_has_no_synthetic_chat_text(self):
         value = parse_user_input(
             {