fix: bound task failures and release stalled workers

tests/unit/admin_console/test_model_settings.py

@@ -33,6 +33,12 @@ def isolated_settings(monkeypatch, tmp_path):
         "ARTEMIS_OUTPUTTER_ENABLED",
         "ARTEMIS_STEP_SUMMARIZER_ENABLED",
         "ARTEMIS_SCREENSHOT_LIMIT_ENABLED",
+        "ARTEMIS_TASK_TIMEOUT_SECONDS",
+        "ARTEMIS_OBSERVATION_TIMEOUT_SECONDS",
+        "ARTEMIS_OBSERVATION_MAX_ATTEMPTS",
+        "ARTEMIS_DEVICE_COMMAND_TIMEOUT_SECONDS",
+        "ARTEMIS_ACTION_FAILURE_ATTEMPTS",
+        "ARTEMIS_WORKER_CLEANUP_TIMEOUT_SECONDS",
     ]
     for key in keys:
         monkeypatch.setattr(settings, key, getattr(settings, key))
@@ -202,3 +208,26 @@ def test_responses_payload_has_reasoning_and_no_sdk_retry(monkeypatch, effort):
     assert factory.call_args.kwargs["use_responses_api"] is True
     ModelFactory.create_model(ModelEndpoint(api_key="test-only", reasoning_effort="default"))
     assert "reasoning_effort" not in factory.call_args.kwargs
+
+
+def test_execution_protection_saves_reload_without_touching_secret(isolated_settings):
+    result = service.save_capability_settings(
+        service.CapabilitySettings(
+            task_timeout_seconds=3600,
+            observation_timeout_seconds=45,
+            observation_max_attempts=2,
+            device_command_timeout_seconds=8,
+            action_failure_attempts=4,
+            worker_cleanup_timeout_seconds=15,
+        )
+    )
+    reloaded = Settings(_env_file=isolated_settings)
+    assert reloaded.ARTEMIS_TASK_TIMEOUT_SECONDS == result.task_timeout_seconds == 3600
+    assert reloaded.ARTEMIS_OBSERVATION_TIMEOUT_SECONDS == 45
+    assert reloaded.ARTEMIS_OBSERVATION_MAX_ATTEMPTS == 2
+    assert reloaded.ARTEMIS_DEVICE_COMMAND_TIMEOUT_SECONDS == 8
+    assert reloaded.ARTEMIS_ACTION_FAILURE_ATTEMPTS == 4
+    assert reloaded.ARTEMIS_WORKER_CLEANUP_TIMEOUT_SECONDS == 15
+    assert reloaded.OPENAI_API_KEY is None
+    assert settings.OPENAI_API_KEY.get_secret_value() == "existing-test-secret"
+    assert dotenv_values(isolated_settings)["UNRELATED"] == "keep"

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):

tests/unit/agents/test_flash_runner_ledger.py

@@ -213,7 +213,8 @@ def test_observation_tail_has_pro_shape(mock_context):
         texts = [b["text"] for b in later.content]
         assert OBSERVATION_HEADER_RE.match(texts[0])
         assert "objective" not in "".join(texts).lower()
-        assert texts[1:] == [
+        assert "当前没有可用的元素索引" in texts[1]
+        assert texts[2:] == [
             "nudge",
             "[INJECTED]",
             'User instruction: "stop"',
@@ -408,6 +409,98 @@ def _report(status="completed", tc_id="tc-report", **extra):
 _INITIAL_OBSERVATION = ("shot0.png", b"IMG0", "[1] Settings")
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("fallback_succeeds", [True, False])
+async def test_failed_actions_have_finite_fallback_and_restore_primary(
+    mock_context, monkeypatch, fallback_succeeds
+):
+    monkeypatch.setattr("artemis.agents.flash.runner.settings.ARTEMIS_ACTION_FAILURE_ATTEMPTS", 3)
+    monkeypatch.setattr(
+        "artemis.agents.flash.runner.settings.ARTEMIS_TOOL_FALLBACK_MAX_NO_PROGRESS_ATTEMPTS", 3
+    )
+    action_count = 4 if fallback_succeeds else 6
+    responses = [
+        AIMessage(
+            content="try", tool_calls=[{"name": "click", "args": {"index": 1}, "id": f"tc-{i}"}]
+        )
+        for i in range(action_count)
+    ]
+    if fallback_succeeds:
+        responses.append(_report())
+    with patch("artemis.controllers.unified_controller.get_driver"):
+        runner = _make_runner(mock_context, responses)
+    fallback = Mock()
+    count = 0
+
+    def execute(name, args, tc_id, state, **kwargs):
+        nonlocal count
+        count += 1
+        result = _exec_result(tc_id, name, b"current-screen")
+        if count <= 3 or not fallback_succeeds:
+            result.status = "error"
+            result.text_summary = "目标不存在"
+        return result
+
+    runner.executor.execute = AsyncMock(side_effect=execute)
+    with (
+        patch(
+            "artemis.agents.flash.runner.capture_screenshot_and_parse_ui",
+            AsyncMock(return_value=_INITIAL_OBSERVATION),
+        ) as observe,
+        patch("artemis.agents.flash.runner.get_llm", return_value=fallback),
+    ):
+        outcome = await runner.run(State(initial_goal="Open Settings"))
+    primary = runner._init_llm.return_value
+    assert all(call["llm"] is primary for call in runner.calls[:3])
+    assert runner.calls[3]["llm"] is fallback
+    assert observe.await_count == 2  # initial + fresh observation before fallback
+    if fallback_succeeds:
+        assert outcome["status"] == "completed"
+        assert runner.calls[-1]["llm"] is primary
+    else:
+        assert outcome["status"] == "failed"
+        assert "兜底模型连续 3" in outcome["explanation"]
+        assert "目标不存在" in outcome["explanation"]
+        assert len(runner.calls) == 6
+
+
+@pytest.mark.asyncio
+async def test_empty_current_ui_clears_previous_turn_element_list(mock_context):
+    response = AIMessage(
+        content="tap", tool_calls=[{"name": "click", "args": {"index": 1}, "id": "tc"}]
+    )
+    with patch("artemis.controllers.unified_controller.get_driver"):
+        runner = _make_runner(mock_context, [response, _report()])
+    runner.executor.execute = AsyncMock(
+        return_value=_exec_result("tc", "click", b"fresh-screen", ui="")
+    )
+    with patch(
+        "artemis.agents.flash.runner.capture_screenshot_and_parse_ui",
+        AsyncMock(return_value=_INITIAL_OBSERVATION),
+    ):
+        await runner.run(State(initial_goal="Open Settings"))
+    tail = runner.calls[-1]["messages"][-1]
+    text = "\n".join(block.get("text", "") for block in tail.content)
+    assert "当前没有可用的元素索引" in text
+    assert "[1] Settings" not in text
+
+
+@pytest.mark.asyncio
+async def test_unavailable_screen_ends_before_any_model_call(mock_context):
+    from artemis.mcp.observation import DeviceObservationError
+
+    with patch("artemis.controllers.unified_controller.get_driver"):
+        runner = _make_runner(mock_context, [])
+    with patch(
+        "artemis.agents.flash.runner.capture_screenshot_and_parse_ui",
+        AsyncMock(side_effect=DeviceObservationError("设备断连")),
+    ):
+        outcome = await runner.run(State(initial_goal="Open Settings"))
+    assert outcome["code"] == "device_observation_failed"
+    assert outcome["error"] == "设备断连"
+    runner._invoke_model.assert_not_awaited()
+
+
 @pytest.mark.asyncio
 async def test_run_builds_prompt_from_ledger_with_session_offsets(mock_context):
     """Turn 2 must see: the static system prefix, the committed turn 1 (its

tests/unit/agents/test_planner_node.py

@@ -33,6 +33,17 @@ class DummyState:
         self.operator_replan_reason = operator_replan_reason
 
 
+import base64
+from io import BytesIO
+from PIL import Image
+
+
+def _valid_screen_base64():
+    buffer = BytesIO()
+    Image.new("RGB", (64, 64), "white").save(buffer, format="PNG")
+    return base64.b64encode(buffer.getvalue()).decode("ascii")
+
+
 @pytest.fixture
 def mock_context():
     ctx = Mock(spec=ArtemisContext)
@@ -41,7 +52,7 @@ def mock_context():
     ctx.adb_client = Mock()
     ctx.ui_adb_client = Mock()
     mock_screen_data = Mock()
-    mock_screen_data.base64 = "YmFzZTY0"
+    mock_screen_data.base64 = _valid_screen_base64()
     mock_screen_data.elements = []
     mock_screen_data.width = 1080
     mock_screen_data.height = 2400

tests/unit/test_drivers.py

@@ -14,7 +14,9 @@
 
 """Unit tests for Unified Base Device Driver and implementations."""
 
+import asyncio
 import base64
+import threading
 from io import BytesIO
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock
@@ -36,6 +38,48 @@ _ONE_PIXEL_PNG = base64.b64encode(
 ).decode("ascii")
 
 
+def _valid_png():
+    buffer = BytesIO()
+    Image.new("RGB", (64, 64), "white").save(buffer, format="PNG")
+    return base64.b64encode(buffer.getvalue()).decode("ascii")
+
+
+_VALID_PNG = _valid_png()
+
+
+@pytest.mark.asyncio
+async def test_android_driver_rejects_placeholder_without_returning_stale_screen():
+    client = MagicMock()
+    client.get_screen_data.return_value = SimpleNamespace(base64=_ONE_PIXEL_PNG)
+    driver = AndroidAdbDriver("test-device", MagicMock(), ui_adb_client=client)
+    with pytest.raises(RuntimeError, match="占位图"):
+        await driver.get_screen_data(skip_settling=True)
+
+
+@pytest.mark.asyncio
+async def test_screen_capture_timeout_does_not_start_overlapping_sdk_calls(monkeypatch):
+    released = threading.Event()
+    client = MagicMock()
+
+    def blocked_capture():
+        released.wait(2)
+        return SimpleNamespace(base64=_VALID_PNG, width=64, height=64, elements=[])
+
+    client.get_screen_data.side_effect = blocked_capture
+    driver = AndroidAdbDriver("test-device", MagicMock(), ui_adb_client=client)
+    monkeypatch.setattr(
+        "artemis.drivers.android.adb_driver.settings.ARTEMIS_OBSERVATION_TIMEOUT_SECONDS", 0.02
+    )
+    try:
+        for _ in range(3):
+            with pytest.raises(TimeoutError, match="观测超时"):
+                await driver.get_screen_data(skip_settling=True)
+        assert client.get_screen_data.call_count == 1
+    finally:
+        released.set()
+        await asyncio.wait_for(driver._screen_capture_task, 2)
+
+
 def test_screenshot_limit_keeps_small_and_invalid_images_unchanged():
     small_image = base64.b64decode(_ONE_PIXEL_PNG)
     assert limit_screenshot_long_edge(small_image, 2048) is small_image
@@ -100,16 +144,16 @@ async def test_android_driver_with_mock_adb():
 
     # Test tap
     await driver.tap(540, 1200)
-    mock_adb_device.shell.assert_called_with("input tap 540 1200")
+    mock_adb_device.shell.assert_called_with("input tap 540 1200", timeout=10.0)
 
     # Test multi-tap
     await driver.tap(540, 1200, times=3, delay_ms=100)
     expected_multi = "input tap 540 1200 && sleep 0.100 && input tap 540 1200 && sleep 0.100 && input tap 540 1200"
-    mock_adb_device.shell.assert_called_with(expected_multi)
+    mock_adb_device.shell.assert_called_with(expected_multi, timeout=10.0)
 
     # Test press key
     await driver.press_key(KeyCode.BACK)
-    mock_adb_device.shell.assert_called_with("input keyevent 4")
+    mock_adb_device.shell.assert_called_with("input keyevent 4", timeout=10.0)
 
 
 @pytest.mark.asyncio
@@ -132,7 +176,7 @@ async def test_android_driver_preserves_elements_from_combined_screen_data():
         },
     ]
     mock_ui_client.get_screen_data.return_value = SimpleNamespace(
-        base64=_ONE_PIXEL_PNG,
+        base64=_VALID_PNG,
         hierarchy_xml="<hierarchy />",
         elements=expected_elements,
         width=1080,
@@ -188,7 +232,7 @@ async def test_android_driver_parses_xml_hierarchy_fallback():
     mock_adb_client = MagicMock()
     mock_ui_client = MagicMock()
     mock_ui_client.get_screen_data.return_value = SimpleNamespace(
-        base64=_ONE_PIXEL_PNG,
+        base64=_VALID_PNG,
         width=1080,
         height=2400,
     )
@@ -216,7 +260,7 @@ async def test_find_element_prefers_fresh_bounds_over_stale_center():
     driver.get_screen_data = AsyncMock(
         return_value=ScreenData(
             screenshot_bytes=base64.b64decode(_ONE_PIXEL_PNG),
-            screenshot_base64=_ONE_PIXEL_PNG,
+            screenshot_base64=_VALID_PNG,
             ui_elements=[
                 {
                     "resource_id": "app:id/digit_4",

tests/unit/test_task_limits.py

@@ -0,0 +1,194 @@
+"""Task deadlines must apply to every ingress and exclude waits."""
+
+import json
+from unittest.mock import AsyncMock, Mock
+
+import pytest
+
+from artemis.runtime.task_limits import ExecutionBudget, task_timeout
+from artemis.runtime.task_result import save_failure_result
+
+
+def test_deadline_precedence_and_single_task_default(monkeypatch):
+    monkeypatch.setattr("artemis.runtime.task_limits.settings.ARTEMIS_TASK_TIMEOUT_SECONDS", 1800)
+    assert task_timeout({}) == 1800
+    job = {"case": {"task_config": {"timeout_seconds": 120}}}
+    assert task_timeout({}, job, {"timeout_seconds": 300}) == 120
+    assert task_timeout({"runtime_overrides": {"timeout_seconds": 60}}, job) == 60
+    assert task_timeout({}, {}, {"timeout_seconds": 300}) == 300
+
+
+def test_budget_excludes_device_queue_and_user_pause(monkeypatch):
+    clock = Mock(return_value=0.0)
+    monkeypatch.setattr("artemis.runtime.task_limits.time.monotonic", clock)
+    budget = ExecutionBudget(30, _last=0)
+    budget.update(excluded=True)
+    clock.return_value = 3600
+    assert budget.update(excluded=False) == 30
+    clock.return_value = 3610
+    assert budget.update(excluded=True) == 20
+    clock.return_value = 7200
+    assert budget.update(excluded=False) == 20
+    clock.return_value = 7221
+    assert budget.update() == 0
+
+
+def test_failure_report_replaces_empty_result_atomically(tmp_path, monkeypatch):
+    monkeypatch.setattr("artemis.runtime.task_result.TRACES_PATH", tmp_path)
+    directory = tmp_path / "test-task"
+    directory.mkdir()
+    (directory / "result.json").touch()
+    save_failure_result("test-task", "设备截图失败：ADB gone", code="device_observation_failed")
+    for name in ("result.json", "run_outcome.json"):
+        payload = json.loads((directory / name).read_text(encoding="utf-8"))
+        assert payload["task_status"] == "failed"
+        assert "ADB gone" in payload["error"]
+    assert not list(directory.glob("*.tmp"))
+    (directory / "result.json").write_text(
+        '{"tests":{"checkpoint":{"passed":false}}}', encoding="utf-8"
+    )
+    save_failure_result("test-task", "任务超时", code="task_timeout")
+    payload = json.loads((directory / "result.json").read_text(encoding="utf-8"))
+    assert payload["tests"]["checkpoint"]["passed"] is False
+
+
+def test_checker_outcome_cannot_trigger_worker_cleanup(tmp_path, monkeypatch):
+    from artemis.runtime.task_result import read_worker_outcome
+
+    monkeypatch.setattr("artemis.runtime.task_result.TRACES_PATH", tmp_path)
+    directory = tmp_path / "checking"
+    directory.mkdir()
+    for status in ("failed", "completed"):
+        (directory / "run_outcome.json").write_text(
+            json.dumps({"task_status": status}), encoding="utf-8"
+        )
+        assert read_worker_outcome("checking") == {}
+    save_failure_result("checking", "设备故障")
+    assert read_worker_outcome("checking")["worker_finalized"] is True
+
+
+@pytest.mark.asyncio
+async def test_single_task_hung_worker_is_killed_and_records_timeout(tmp_path, monkeypatch):
+    import asyncio
+    from apps.admin_console.services.task_queue_service import TaskQueueService
+
+    monkeypatch.setattr("artemis.runtime.task_result.TRACES_PATH", tmp_path)
+    monkeypatch.setattr(
+        "apps.admin_console.services.automation_store.store", Mock(job=Mock(return_value=None))
+    )
+    monkeypatch.setattr("artemis.runtime.task_limits.task_timeout", lambda *args: 0)
+    monkeypatch.setattr("artemis.runtime.task_control.read_state", lambda *args: {})
+    monkeypatch.setattr(TaskQueueService, "_sync_control_states", lambda: None)
+    stopped = asyncio.Event()
+
+    async def wait(proc):
+        await stopped.wait()
+        return -9
+
+    kill = AsyncMock(side_effect=lambda proc: stopped.set())
+    monkeypatch.setattr(TaskQueueService, "_wait_for_worker_process", wait)
+    monkeypatch.setattr(TaskQueueService, "_terminate_worker_process", kill)
+    item = {"session_id": "hung-single"}
+    assert (
+        await asyncio.wait_for(
+            TaskQueueService._await_worker_with_budget(Mock(), item, "hung-single"), 2
+        )
+        == -9
+    )
+    kill.assert_awaited_once()
+    assert item["timed_out"]
+    result = json.loads((tmp_path / "hung-single" / "result.json").read_text(encoding="utf-8"))
+    assert result["code"] == "task_timeout"
+    assert "不含暂停/排队" in result["error"]
+
+
+def test_standalone_mcp_worker_deadline_kills_before_report_write(monkeypatch):
+    from mcp_server.tools import task_runner
+
+    proc = Mock(pid=999999, poll=Mock(return_value=None))
+    monkeypatch.setattr(task_runner.trace_store, "read_status", lambda *_: {"status": "running"})
+    monkeypatch.setattr(task_runner, "read_state", lambda *_: {"status": "running"})
+    monkeypatch.setattr(
+        task_runner.DeviceExecutionLock,
+        "get_active_owners",
+        lambda: {"device": Mock(session_id="mcp-hung")},
+    )
+    events = []
+    monkeypatch.setattr(task_runner, "_kill_process_tree", lambda *_: events.append("kill"))
+
+    def fail_report(*args, **kwargs):
+        events.append("report")
+        raise OSError("disk full")
+
+    monkeypatch.setattr(task_runner, "save_failure_result", fail_report)
+    task_runner._watch_execution(proc, "mcp-hung", 0)
+    assert events == ["kill", "report"]
+
+
+def test_mcp_finalized_worker_is_reclaimed_without_overwriting_device_failure(monkeypatch):
+    from mcp_server.tools import task_runner
+
+    outcome = {"task_status": "failed", "error": "设备断连", "code": "device_observation_failed"}
+    monkeypatch.setattr(task_runner, "read_worker_outcome", lambda _: outcome)
+    monkeypatch.setattr(task_runner.trace_store, "read_status", lambda _: {"status": "running"})
+    updated = Mock()
+    monkeypatch.setattr(task_runner.trace_store, "update_trace_status", updated)
+    kill = Mock()
+    monkeypatch.setattr(task_runner, "_kill_process_tree", kill)
+    clock = Mock(side_effect=[0.0, 0.1])
+    monkeypatch.setattr(task_runner.time, "monotonic", clock)
+    task_runner._watch_execution(
+        Mock(pid=123, poll=Mock(return_value=None)), "test-mcp", 1800, cleanup_limit=0.01
+    )
+    kill.assert_called_once_with(123)
+    updated.assert_called_once_with("test-mcp", "failed", error="设备断连")
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("finalized", [False, True])
+async def test_real_hung_process_stops_and_next_worker_can_complete(
+    tmp_path, monkeypatch, finalized
+):
+    import asyncio
+    import sys
+    from apps.admin_console.services.task_queue_service import TaskQueueService
+
+    monkeypatch.setattr("artemis.runtime.task_result.TRACES_PATH", tmp_path)
+    monkeypatch.setattr(
+        "apps.admin_console.services.automation_store.store", Mock(job=Mock(return_value=None))
+    )
+    monkeypatch.setattr(
+        "artemis.runtime.task_limits.task_timeout",
+        lambda item, *args: 0 if item["session_id"] == "hung" and not finalized else 30,
+    )
+    monkeypatch.setattr("artemis.runtime.task_control.read_state", lambda *args: {})
+    monkeypatch.setattr(TaskQueueService, "_sync_control_states", lambda: None)
+    item = {"session_id": "hung", "runtime_overrides": {"worker_cleanup_timeout_seconds": 0.01}}
+    if finalized:
+        save_failure_result("hung", "设备截图失效", code="device_observation_failed")
+    hung = await asyncio.create_subprocess_exec(
+        sys.executable, "-c", "import time; time.sleep(3600)"
+    )
+    try:
+        code = await asyncio.wait_for(
+            TaskQueueService._await_worker_with_budget(hung, item, "hung"), 10
+        )
+        assert code != 0
+        assert hung.returncode is not None
+        result = json.loads((tmp_path / "hung" / "result.json").read_text(encoding="utf-8"))
+        assert result["code"] == ("device_observation_failed" if finalized else "task_timeout")
+        assert bool(item.get("timed_out")) is not finalized
+        next_worker = await asyncio.create_subprocess_exec(sys.executable, "-c", "pass")
+        assert (
+            await asyncio.wait_for(
+                TaskQueueService._await_worker_with_budget(
+                    next_worker, {"session_id": "next"}, "next"
+                ),
+                10,
+            )
+            == 0
+        )
+    finally:
+        if hung.returncode is None:
+            hung.kill()
+            await hung.wait()

tests/unit/tools/test_utils.py

@@ -47,6 +47,17 @@ else:
     del sys.modules["artemis.graph.state"]
 
 
+import base64
+from io import BytesIO
+from PIL import Image
+
+
+def _valid_screen_base64():
+    buffer = BytesIO()
+    Image.new("RGB", (64, 64), "white").save(buffer, format="PNG")
+    return base64.b64encode(buffer.getvalue()).decode("ascii")
+
+
 @pytest.fixture
 def mock_context():
     """Create a mock ArtemisContext for testing."""
@@ -70,6 +81,7 @@ def mock_context():
 
     # Mock the ADB client for Android
     mock_response = Mock()
+    mock_response.base64 = _valid_screen_base64()
     mock_response.json.return_value = {"elements": []}
     ctx.ui_adb_client.get_screen_data = Mock(return_value=mock_response)
     ctx.ui_adb_client.get_hierarchy = Mock(
@@ -288,6 +300,7 @@ class TestFocusElementIfNeeded:
         focused_element["attributes"]["focused"] = "true"
 
         mock_response = Mock()
+        mock_response.base64 = _valid_screen_base64()
         mock_response.json.return_value = {"elements": [focused_element]}
         mock_context.ui_adb_client.get_screen_data = Mock(return_value=mock_response)
         mock_find_element.return_value = focused_element["attributes"]
@@ -366,6 +379,7 @@ class TestFocusElementIfNeeded:
         }
 
         mock_response = Mock()
+        mock_response.base64 = _valid_screen_base64()
         mock_response.json.return_value = {"elements": [element_from_text]}
         mock_context.ui_adb_client.get_screen_data = Mock(return_value=mock_response)
         mock_find_id.return_value = element_from_id
@@ -401,6 +415,7 @@ class TestFocusElementIfNeeded:
         }
 
         mock_response = Mock()
+        mock_response.base64 = _valid_screen_base64()
         mock_response.json.return_value = {"elements": [element_with_bounds]}
         mock_context.ui_adb_client.get_screen_data = Mock(return_value=mock_response)
         mock_find_text.return_value = element_with_bounds["attributes"]
@@ -428,6 +443,7 @@ class TestFocusElementIfNeeded:
         """Test failure when no locator can find an element."""
 
         mock_response = Mock()
+        mock_response.base64 = _valid_screen_base64()
         mock_response.json.return_value = {"elements": []}
         mock_context.ui_adb_client.get_screen_data = Mock(return_value=mock_response)
         with (