feat: checkpoint device agent runtime milestones

tools/describe_screen.py

@@ -1,11 +1,10 @@
 from __future__ import annotations
 
-from core.device_manager import DeviceManager
 from core.models import Scene
+from device.manager import DeviceManager
+from perception.ocr import PaddleOCREngine
+from perception.provider import DefaultPerceptionProvider, PerceptionProvider
 from tools._device import get_driver
-from vision.ocr import PaddleOCREngine, run_ocr
-from vision.scene_builder import build_scene, infer_png_size
-from vision.ui_parser import parse_ui_tree
 
 
 def describe_screen(
@@ -13,15 +12,10 @@ def describe_screen(
     *,
     manager: DeviceManager | None = None,
     ocr_engine: PaddleOCREngine | None = None,
+    perception_provider: PerceptionProvider | None = None,
 ) -> Scene:
     driver = get_driver(device_id, manager=manager)
     screenshot = driver.screenshot()
     raw_tree = driver.tree()
-    width, height = infer_png_size(screenshot)
-    return build_scene(
-        screen_width=width,
-        screen_height=height,
-        ui_elements=parse_ui_tree(raw_tree),
-        ocr_elements=run_ocr(screenshot, engine=ocr_engine),
-    )
-
+    provider = perception_provider or DefaultPerceptionProvider(ocr_engine=ocr_engine)
+    return provider.build_scene(screenshot, raw_tree)