fix: bound task failures and release stalled workers

tests/unit/agents/test_flash_runner.py

@@ -15,6 +15,9 @@
 """Unit tests for Universal FlashRunner."""
 
 import base64
+from io import BytesIO
+from PIL import Image
+from artemis.mcp.observation import DeviceObservationError
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, Mock, patch
 import pytest
@@ -72,9 +75,18 @@ def test_flash_runner_tools_initialization(mock_context):
 # --- Post-action screenshot fallback (failed action / unknown tool) ----------------
 
 
+def _post_image():
+    buffer = BytesIO()
+    Image.new("RGB", (64, 64), "white").save(buffer, format="PNG")
+    return buffer.getvalue()
+
+
+_POST_IMG = _post_image()
+
+
 def _screen(elements):
     return SimpleNamespace(
-        base64=base64.b64encode(b"POST-IMG").decode("ascii"),
+        base64=base64.b64encode(_POST_IMG).decode("ascii"),
         elements=elements,
         width=1080,
         height=2400,
@@ -120,7 +132,7 @@ async def test_fallback_observation_renders_the_indexed_ui_list(mock_context, tm
             exec_result, "click", frozenset({"click"}), state
         )
 
-    assert post == b"POST-IMG"
+    assert post == _POST_IMG
     assert isinstance(exec_result.ui_elements_text, str)
     assert exec_result.ui_elements_text == "[1] Text: 'Settings' | Bounds: [37,50][962,112]"
     assert "{'text'" not in exec_result.ui_elements_text
@@ -154,8 +166,8 @@ async def test_fallback_observation_is_skipped_for_non_action_tools(mock_context
 
 
 @pytest.mark.asyncio
-async def test_fallback_observation_keeps_the_executors_ui_list(mock_context, tmp_path):
-    """An action that reported a UI list but no screenshot keeps its own list."""
+async def test_fallback_observation_replaces_stale_executor_ui_list(mock_context, tmp_path):
+    """Fresh screenshots and element lists must describe the same observation."""
     with patch("artemis.controllers.unified_controller.get_driver"):
         runner = FlashRunner(mock_context, goal="g")
     runner.controller = Mock()
@@ -175,12 +187,13 @@ async def test_fallback_observation_keeps_the_executors_ui_list(mock_context, tm
             exec_result, "click", frozenset({"click"}), None
         )
 
-    assert post == b"POST-IMG"
-    assert exec_result.ui_elements_text == "[1] Text: 'Mine' | Bounds: [0,0][10,10]"
+    assert post == _POST_IMG
+    assert "Other" in exec_result.ui_elements_text
+    assert "Mine" not in exec_result.ui_elements_text
 
 
 @pytest.mark.asyncio
-async def test_fallback_observation_failure_is_contained(mock_context):
+async def test_fallback_observation_failure_stops_after_bounded_recovery(mock_context):
     with patch("artemis.controllers.unified_controller.get_driver"):
         runner = FlashRunner(mock_context, goal="g")
     runner.controller = Mock()
@@ -189,11 +202,13 @@ async def test_fallback_observation_failure_is_contained(mock_context):
     state.latest_screenshot = "old.jpg"
     exec_result = _failed_click()
 
-    post = await runner._capture_post_screenshot(exec_result, "click", frozenset({"click"}), state)
-
-    assert post is None
-    assert exec_result.ui_elements_text is None
-    assert state.latest_screenshot == "old.jpg"
+    with patch("artemis.mcp.observation.settings.ARTEMIS_OBSERVATION_MAX_ATTEMPTS", 3):
+        with pytest.raises(DeviceObservationError, match="adb gone"):
+            await runner._capture_post_screenshot(exec_result, "click", frozenset({"click"}), state)
+    assert runner.controller.get_screen_data.await_count == 3
+    assert state.latest_screenshot is None
+    assert state.indexed_elements == []
+    assert state.indexed_points == []
 
 
 def test_flash_runner_executor_follows_the_flash_explorer_tier(mock_context):