feat: add plugin runtime and cross-device task orchestration

artemis/agents/validator/execution_loop.py

@@ -181,6 +181,33 @@ async def _process_action(
     outcome = _ActionOutcome(action_item=action_item)
     attempts_log = []
 
+    from artemis.plugins.runtime import get_plugin_runtime
+    from artemis.runtime.task_control import TaskControl
+
+    plugins = get_plugin_runtime(node.ctx)
+    if plugins and action_name in plugins.bindings:
+        operation = plugins.execute(action_name, action_item.get("args", {}), str(uuid.uuid4()))
+        control = getattr(node.ctx, "task_control", None)
+        if isinstance(control, TaskControl):
+            interrupted, result = await control.run_interruptibly(
+                operation,
+                on_pause=getattr(node.ctx, "on_task_pause", None),
+                on_resume=getattr(node.ctx, "on_task_resume", None),
+            )
+            if interrupted:
+                outcome.success = False
+                outcome.error_msg = "插件操作已打断，结果不确定；继续前必须重新观察。"
+                return outcome
+        else:
+            result = await operation
+        outcome.success = result.status == "success"
+        outcome.error_msg = "" if outcome.success else result.text_summary
+        outcome.action_item.update(result.metadata)
+        outcome.action_item["result"] = result.text_summary
+        if result.screenshot_bytes:
+            outcome.post_screenshot_b64 = base64.b64encode(result.screenshot_bytes).decode()
+        return outcome
+
     validation_passed = True
     if not burst:
         # 1. Pre-execution validation (Safety Net)
@@ -338,6 +365,8 @@ def _model_facing_action(ctx, action_item: dict) -> dict:
     The live item carries the controller's pixels; every phrase rendered for
     a model is normalized first, and the space marker makes that idempotent.
     """
+    if action_item.get("plugin_id"):
+        return action_item
     width = getattr(getattr(ctx, "device", None), "device_width", None) or 1080
     height = getattr(getattr(ctx, "device", None), "device_height", None) or 2400
     return normalize_action_dict(action_item, width, height)
@@ -419,11 +448,19 @@ async def run_validation_loop(node, state: State) -> dict:
             execution.append(outcome.action_item)
 
             post_image_name = None
-            if outcome.post_screenshot_b64:
+            plugin_action = bool(outcome.action_item.get("plugin_id"))
+            mixed_plugin_action = (
+                plugin_action and getattr(ctx, "execution_target", "android") == "mixed"
+            )
+            if outcome.post_screenshot_b64 and not mixed_plugin_action:
                 post_image_name = _absorb_post_screenshot(ctx, state, outcome.post_screenshot_b64)
                 last_screenshot_b64 = outcome.post_screenshot_b64
                 if post_image_name:
                     last_screenshot_name = post_image_name
+            elif mixed_plugin_action:
+                # Keep the Android step observation on its own resource. Plugin
+                # images remain attached to the tool/action trace and its exports.
+                post_image_name = outcome.action_item.get("post_image_name")
 
             _record_action_end(
                 ctx,