feat: add guided exploration and share user-started MCP backend

artemis/graph/perception.py

@@ -125,6 +125,18 @@ async def perception_node(state: State, ctx: ArtemisContext) -> dict:
         ctx.data_engine.allocate_step_id()
         logger.info(f"Pre-allocated step ID in Perception Node: {ctx.data_engine.current_step_id}")
 
+    controller = UnifiedMobileController(ctx)
+
+    # Observe before acknowledging guidance. A failed capture must not remove
+    # instructions that the Operator has never seen.
+    if not _should_skip_settling(state):
+        await asyncio.sleep(0.4)
+    from artemis.exploration.runtime import observe_or_pause
+
+    device_data = await observe_or_pause(ctx, controller)
+    latest_screenshot_b64 = device_data.base64
+    xml_hierarchy = device_data.elements
+
     # Check for injected instruction without blocking event loop
     injected_instruction = None
     user_stop_requested = False
@@ -144,21 +156,6 @@ async def perception_node(state: State, ctx: ArtemisContext) -> dict:
                 # this moment stay protected until the next instruction.
                 _unprotect_current_check_lines(ctx)
 
-    controller = UnifiedMobileController(ctx)
-
-    # 1. Capture screenshot with hybrid settling strategy
-    if _should_skip_settling(state):
-        logger.info("Skipping settling. Retrieving screen data directly...")
-        device_data = await controller.get_screen_data()
-        latest_screenshot_b64 = device_data.base64
-        xml_hierarchy = device_data.elements
-    else:
-        logger.info("Screen settling required. Delaying 0.4s and capturing screen data...")
-        await asyncio.sleep(0.4)
-        device_data = await controller.get_screen_data()
-        latest_screenshot_b64 = device_data.base64
-        xml_hierarchy = device_data.elements
-
     latest_screenshot_bytes = base64.b64decode(latest_screenshot_b64)
 
     # Update context device dimensions dynamically