fix: prevent false timeout failures and release 1.2.2

.env.example

@@ -6,6 +6,8 @@ OPENAI_BASE_URL=https://api-n-cd.playones.com/v1
 OPENAI_MODEL=mimo-v2.6-flash
 # 单次模型 HTTP 请求超时（秒）
 OPENAI_TIMEOUT_SECONDS=300
+# 可选的后台步骤截图摘要单次超时，独立于上面的主模型请求超时。
+ARTEMIS_STEP_SUMMARIZER_TIMEOUT_SECONDS=25
 # 失败后最多重试 5 次，由 Artemis 统一管理，SDK 不叠加重试。
 OPENAI_MAX_RETRIES=5
 OPENAI_RETRY_INTERVAL_SECONDS=10
@@ -19,6 +21,8 @@ ARTEMIS_OUTPUTTER_ENABLED=true
 ARTEMIS_STEP_SUMMARIZER_ENABLED=true
 # Pro 连续无工具调用或参数校验失败的最大次数；耗尽后明确失败，避免空转。
 ARTEMIS_OPERATOR_MAX_NO_PROGRESS_ATTEMPTS=3
+# Flash 连续无工具调用的最大次数；无效回复不会重复送回模型。
+ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS=3
 
 # 视频分段、重试与单次分析超时（秒）
 ARTEMIS_VIDEO_CHUNK_SECONDS=60

apps/admin_console/services/automation_service.py

@@ -130,6 +130,11 @@ async def tick_batches():
                 batch.update(status="paused", reason="模型请求已暂停，请先恢复模型连接再继续批次")
                 store.put("batch", batch["id"], batch)
             for job in active:
+                # The queue records started_at at submission, then replaces it
+                # when the worker actually starts. Waiting for a device must
+                # not consume this case's execution budget.
+                if job["status"] != "running":
+                    continue
                 limit = int(
                     (job.get("case", {}).get("task_config") or {}).get("timeout_seconds")
                     or batch.get("timeout_seconds", 1800)
@@ -139,8 +144,7 @@ async def tick_batches():
                 if task_state.get("non_execution_started_at"):
                     excluded += time.time() - float(task_state["non_execution_started_at"])
                 if (
-                    job["status"] != "finalizing"
-                    and not job.get("timed_out")
+                    not job.get("timed_out")
                     and time.time() - job.get("started_at", time.time()) - excluded > limit
                 ):
                     job.update(
@@ -148,8 +152,8 @@ async def tick_batches():
                     )
                     store.save_job(job)
                     task_queue_service.stop_tasks(session_id=job["session_id"])
-                    batch.update(status="paused", reason="任务超时，已暂停剩余任务；请确认设备状态")
-                    store.put("batch", batch["id"], batch)
+                    # on_terminal records the failed case. The next pending
+                    # case then runs normally unless stop_on_failure is set.
             continue
         if batch["status"] != "running":
             continue

apps/admin_console/services/report_export.py

@@ -95,21 +95,46 @@ def failure_reason(directory: Path, status: str, recorded: str | None) -> str:
     state = read_json(directory / "status.json", {})
     if isinstance(state, dict) and state.get("error"):
         return redact(state["error"])
+    log_lines: list[str] = []
     for name in ("stdout.log", "stderr.log"):
         path = directory / name
         try:
             with path.open("rb") as stream:
                 stream.seek(0, 2)
                 stream.seek(max(0, stream.tell() - 131072))
-                lines = stream.read().decode("utf-8", errors="replace").splitlines()
+                log_lines.extend(stream.read().decode("utf-8", errors="replace").splitlines())
         except OSError:
             continue
-        for line in reversed(lines):
-            match = re.search(r"Error running automation:\s*(.+)", line)
-            if match:
-                return redact(match.group(1))
-            if re.search(r"(?i)(?:timeout|timed out|超时|deadline exceeded)", line):
-                return redact("模型请求或任务执行超时：" + line.strip())
+    for line in reversed(log_lines):
+        match = re.search(r"Error running automation:\s*(.+)", line)
+        if match:
+            return redact(match.group(1))
+    for line in reversed(log_lines):
+        stripped = line.strip()
+        if stripped.startswith(("⚠", "ℹ", "WARNING", "INFO", "DEBUG")):
+            continue
+        if any(
+            name in line
+            for name in (
+                "VisualStepSummarizer:",
+                "StepMemoryService:",
+                "Could not clean unowned screen wake locks",
+            )
+        ):
+            continue
+        exception_line = re.fullmatch(
+            r"(?:[A-Za-z_]\w*\.)*[A-Za-z_]\w*(?:Error|Exception|Timeout):\s*.*",
+            stripped,
+        )
+        if exception_line and not re.search(r"(?i)(?:timeout|timed out|超时)", stripped):
+            return redact(stripped)
+        if re.search(
+            r"(?i)(?:\b(?:TimeoutError|APITimeoutError|ReadTimeout|ConnectTimeout)\s*:|"
+            r"(?:模型请求|任务执行|model request|task execution).{0,100}"
+            r"(?:超时|timed out|deadline exceeded))",
+            stripped,
+        ):
+            return redact("模型请求或任务执行超时：" + stripped)
     return "任务未完成，运行日志中没有可提取的具体异常；请查看执行步骤。"
 
 

apps/showcase_ui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "artemis-showcase-ui",
-  "version": "1.2.0",
+  "version": "1.2.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "artemis-showcase-ui",
-      "version": "1.2.0",
+      "version": "1.2.2",
       "dependencies": {
         "@angular/cdk": "^22.1.5",
         "@angular/common": "^22.1.4",

apps/showcase_ui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "artemis-showcase-ui",
-  "version": "1.2.0",
+  "version": "1.2.2",
   "scripts": {
     "ng": "ng",
     "start": "ng serve --proxy-config proxy.conf.json",

artemis/agents/flash/runner.py

@@ -167,6 +167,7 @@ class FlashRunner:
                 ctx,
                 model_name=self.step_summarizer_cfg.model,
                 retry_limit=self.memory_runtime_cfg.retry_limit,
+                timeout_seconds=self.step_summarizer_cfg.timeout_seconds,
                 max_concurrency=self.memory_runtime_cfg.max_concurrency,
                 flush_timeout_s=self.memory_runtime_cfg.flush_timeout_s,
             )
@@ -1063,6 +1064,8 @@ class FlashRunner:
         current_xml_list = xml_list
         previous_turn: _TurnRecord | None = None
         pending_notices: list[str] = []
+        no_tool_attempts = 0
+        no_tool_limit = settings.ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS
 
         while limit is None or turns < limit:
             control = getattr(self.ctx, "task_control", None)
@@ -1198,11 +1201,24 @@ class FlashRunner:
                 if is_final:
                     final_report = {"status": "failed", "explanation": raw_text}
                     break
+                no_tool_attempts += 1
+                if no_tool_attempts >= no_tool_limit:
+                    final_report = {
+                        "status": "failed",
+                        "explanation": (
+                            f"模型连续 {no_tool_attempts} 次未调用工具，"
+                            "已停止任务以避免空转和重复计费；未执行新的设备动作。"
+                        ),
+                    }
+                    break
                 pending_notices.append(_NO_TOOL_CALL_NOTICE)
-                ledger.stage_turn(messages[turn_base:])
+                # Keep the observation and any injected user instruction, but
+                # do not feed the model its own unproductive prose on retry.
+                ledger.stage_turn(messages[turn_base:-1])
                 previous_turn = _TurnRecord()
                 continue
 
+            no_tool_attempts = 0
             # Process tool calls
             turn = _TurnRecord()
             (

artemis/agents/flash/summarizer.py

@@ -151,6 +151,7 @@ class VisualStepSummarizer(StepMemoryService):
         ctx: ArtemisContext,
         model_name: str | None = None,
         retry_limit: int = 3,
+        timeout_seconds: float = 25.0,
         *,
         max_concurrency: int = 1,
         flush_timeout_s: float = 30.0,
@@ -161,6 +162,7 @@ class VisualStepSummarizer(StepMemoryService):
             retry_limit=retry_limit,
             flush_timeout_s=flush_timeout_s,
         )
+        self._timeout_seconds = timeout_seconds
 
         # Initialize lightweight VLM: prioritize explicit model_name
         target_model = model_name or settings.OPENAI_MODEL
@@ -459,7 +461,9 @@ class VisualStepSummarizer(StepMemoryService):
                 HumanMessage(content=content_blocks),
             ]
 
-            response = await asyncio.wait_for(self._llm.ainvoke(messages), timeout=25.0)
+            response = await asyncio.wait_for(
+                self._llm.ainvoke(messages), timeout=self._timeout_seconds
+            )
             self._meter_lens_call(response)
             summary_raw = response.content if isinstance(response.content, str) else ""
             if isinstance(response.content, list):
@@ -505,8 +509,14 @@ class VisualStepSummarizer(StepMemoryService):
                         logger.debug(f"DataEngine step summary update skipped: {de_err}")
                 return True
 
+        except TimeoutError:
+            logger.warning(
+                f"VisualStepSummarizer: Step {step_number} summary request timed out "
+                f"(limit {self._timeout_seconds:.1f}s)"
+            )
         except Exception as e:
             logger.warning(
-                f"VisualStepSummarizer: Error generating summary for step {step_number}: {e}"
+                f"VisualStepSummarizer: Error generating summary for step {step_number}: "
+                f"{type(e).__name__}: {e}"
             )
         return False

artemis/config/agent.py

@@ -370,6 +370,11 @@ class StepSummarizerConfig(BaseModel):
         default_factory=lambda: settings.OPENAI_MODEL,
         description="Lightweight model used for background step state summarization.",
     )
+    timeout_seconds: float = Field(
+        default=25.0,
+        gt=0,
+        description="Timeout for one optional background visual-summary model request.",
+    )
     prune_history_xml: bool = Field(
         default=True,
         description="Whether to prune outdated heavy UI XML trees from historical turns.",

artemis/config/settings.py

@@ -114,6 +114,8 @@ class Settings(BaseSettings):
     ARTEMIS_STEP_SUMMARIZER_ENABLED: bool = True
     # Consecutive Pro responses without a valid tool call before failing the task.
     ARTEMIS_OPERATOR_MAX_NO_PROGRESS_ATTEMPTS: int = Field(default=3, ge=1, le=20)
+    # Consecutive Flash responses without any tool call before failing the task.
+    ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS: int = Field(default=3, ge=1, le=20)
 
     # Android ADB Connectivity
     ADB_HOST: str | None = Field(default=DEFAULT_ADB_HOST)

artemis/resources/config/artemis.jsonc

@@ -33,6 +33,7 @@
       "step_summarizer": {
         "enabled": "${env:ARTEMIS_STEP_SUMMARIZER_ENABLED:-true}",
         "model": "${env:OPENAI_MODEL:-mimo-v2.6-flash}",
+        "timeout_seconds": "${env:ARTEMIS_STEP_SUMMARIZER_TIMEOUT_SECONDS:-25}",
         "prune_history_xml": true
       }
     },

artemis/runtime/awake_lease.py

@@ -113,22 +113,28 @@ class ScreenAwakeLease:
         }
 
     @contextmanager
-    def _mutex(self):
+    def _mutex(self, *, wait_timeout_seconds: float | None = None):
         self.mutex_path.parent.mkdir(parents=True, exist_ok=True)
         started_at = time.monotonic()
+        timeout = (
+            self._MUTEX_TIMEOUT_SECONDS
+            if wait_timeout_seconds is None
+            else max(0.0, wait_timeout_seconds)
+        )
         acquired = False
         while not acquired:
             try:
                 fd = os.open(self.mutex_path, os.O_CREAT | os.O_EXCL | os.O_WRONLY)
             except FileExistsError:
+                if time.monotonic() - started_at >= timeout:
+                    raise TimeoutError("Timed out acquiring the Artemis awake-lease mutex.")
                 if self._path_is_stale(self.mutex_path):
                     try:
                         self.mutex_path.unlink(missing_ok=True)
                     except OSError:
                         pass
-                    continue
-                if time.monotonic() - started_at >= self._MUTEX_TIMEOUT_SECONDS:
-                    raise TimeoutError("Timed out acquiring the Artemis awake-lease mutex.")
+                    else:
+                        continue
                 time.sleep(self._MUTEX_POLL_SECONDS)
             else:
                 try:
@@ -274,10 +280,15 @@ class ScreenAwakeLease:
     def cleanup_unowned_references(self) -> bool:
         """Drain legacy shell wake locks only when no live old host owns them."""
         try:
-            with self._mutex():
+            # Migration cleanup is optional. A busy owner must not delay every
+            # queued task by the normal 10-second lease-acquisition deadline.
+            with self._mutex(wait_timeout_seconds=1.0):
                 if self._live_leases():
                     return False
                 return self._drain_references()
+        except TimeoutError:
+            logger.debug(f"Skipped optional wake-lock cleanup on {self.device_id}: mutex busy")
+            return False
         except Exception as exc:
             logger.warning(f"Could not clean unowned screen wake locks on {self.device_id}: {exc}")
             return False

artemis/runtime/awake_service.py

@@ -95,9 +95,10 @@ def _configure_usb_stay_awake(device_id: str) -> str | None:
     if not _awake_enabled() or os.environ.get("ARTEMIS_CLOUD_MODE") == "1":
         return None
 
-    # Migration cleanup: previous Artemis versions used a persistent shell
-    # wake lock. Drain it only after every live legacy owner has exited.
-    ScreenAwakeLease(device_id).cleanup_unowned_references()
+    # Migration cleanup belongs to the long-lived host. A queued task worker
+    # must not wait on another process's optional legacy wake-lock cleanup.
+    if os.environ.get("ARTEMIS_TASK_WORKER") != "1":
+        ScreenAwakeLease(device_id).cleanup_unowned_references()
 
     commands = (
         (

config/artemis.jsonc

@@ -33,6 +33,7 @@
       "step_summarizer": {
         "enabled": "${env:ARTEMIS_STEP_SUMMARIZER_ENABLED:-true}",
         "model": "${env:OPENAI_MODEL:-mimo-v2.6-flash}",
+        "timeout_seconds": "${env:ARTEMIS_STEP_SUMMARIZER_TIMEOUT_SECONDS:-25}",
         "prune_history_xml": true
       }
     },

docs/releases/1.2.1.md

@@ -0,0 +1,14 @@
+# ARTEMIS-Eapil 1.2.1
+
+## 修复
+
+- 批量任务达到执行超时后，该用例记录为失败，不再自动暂停整批。当前用例结算后继续执行下一项；开启“失败即停止”的批次仍会跳过剩余项。排队等待不计入执行超时。
+- Flash 连续收到无工具调用的回复时，最多尝试 3 次并给出明确失败原因，避免持续空转和重复计费。重试会保留设备观察与用户指导，但不把模型的无效回复送回上下文。可通过 `ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS` 调整上限。
+
+模型请求中的相同缓存 token 数量通常是提示词前缀缓存命中，不能据此判断缓存造成循环。本版针对无工具调用的执行循环建立退出条件。
+
+## 安装与验证
+
+请下载附件 `artemis-eapil-1.2.1-source.zip`，解压后运行 `start.bat`；国内网络使用 `start-cn.bat`。附件包含精简 Git 仓库，可供后续启动脚本检查更新。Gitea 自动生成的源码 ZIP 不含 `.git`，请使用发布附件。升级后重启 ARTEMIS 服务、刷新网页并重新加载 AI 客户端的 MCP。
+
+验证：Python 默认测试 2182 项通过、6 项跳过；Ruff、质量基线和核心类型检查通过。

docs/releases/1.2.2.md

@@ -0,0 +1,13 @@
+# ARTEMIS-Eapil 1.2.2
+
+## 修复
+
+- 任务报告区分实际执行失败与可恢复的诊断警告。屏幕唤醒锁清理超时、步骤视觉摘要重试不再被误写为任务或模型请求超时；真实终止异常仍会显示。
+- 旧屏幕唤醒锁清理遇到设备锁繁忙时快速跳过，避免队列任务因清理动作等待或连续失败。长驻服务负责后续清理。
+- Flash 步骤视觉摘要使用独立超时，默认 25 秒，可通过 `ARTEMIS_STEP_SUMMARIZER_TIMEOUT_SECONDS` 设置，不再占用主模型请求的 300 秒超时额度。
+
+## 安装与验证
+
+下载附件 `artemis-eapil-1.2.2-source.zip`，解压后运行 `start.bat`；国内网络使用 `start-cn.bat`。附件含精简 Git 仓库，后续可由启动脚本检查更新。升级后重启 ARTEMIS 服务、刷新网页并重新加载 AI 客户端的 MCP。
+
+验证：Python 默认测试 2191 项通过、2 项跳过、8 项按标记排除；Ruff、质量基线和核心类型检查通过。两项真实模拟器 Flash 批量任务均完成；干净环境未复现原先的锁争用，锁繁忙与报告归因通过回归测试覆盖。

pyproject.toml

@@ -14,7 +14,7 @@
 
 [project]
 name = "artemis"
-version = "1.2.0"
+version = "1.2.2"
 description = "Autonomous AI Assistant for Mobile Automation"
 
 authors = [

tests/unit/admin_console/test_automation.py

@@ -4,7 +4,7 @@ from io import BytesIO
 import json
 import sqlite3
 import time
-from unittest.mock import AsyncMock
+from unittest.mock import AsyncMock, Mock
 from uuid import UUID
 import zipfile
 
@@ -347,6 +347,61 @@ async def test_batch_timeout_stays_visible_in_job_report_and_history(client, iso
     )
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("stop_on_failure", [False, True])
+async def test_timed_out_case_releases_next_batch_case(isolated, monkeypatch, stop_on_failure):
+    from apps.admin_console import server
+    from artemis.runtime.device_pool import DeviceStatus
+
+    batch = service.create_batch(payload(timeout_seconds=1, stop_on_failure=stop_on_failure))
+    service.batch_action(batch["id"], "resume")
+    first = isolated.jobs(batch["id"])[0]
+    isolated.save_job({**first, "status": "running", "started_at": time.time() - 2})
+    monkeypatch.setattr(
+        service, "state", type("State", (), {"is_paused": False, "active_runs": {}})()
+    )
+    stop = Mock(return_value=True)
+    enqueue = AsyncMock(return_value={"tasks": [{"device_serial": "test-device"}]})
+    monkeypatch.setattr(server.task_queue_service, "stop_tasks", stop)
+    monkeypatch.setattr(server.task_queue_service, "enqueue_tasks", enqueue)
+    monkeypatch.setattr(
+        "artemis.runtime.device_pool.try_list_devices_async",
+        AsyncMock(return_value=[DeviceStatus(serial="test-device", state="device")]),
+    )
+    monkeypatch.setattr(reports, "export_batch", Mock())
+
+    await service.tick_batches()
+    assert isolated.get("batch", batch["id"])["status"] == "running"
+    assert "超时" in isolated.job(first["session_id"])["error"]
+    stop.assert_called_once_with(session_id=first["session_id"])
+    await service.tick_batches()
+    stop.assert_called_once()
+
+    await service.on_terminal({"session_id": first["session_id"]}, "cancelled")
+    await service.tick_batches()
+    statuses = [job["status"] for job in isolated.jobs(batch["id"])]
+    assert statuses == (["failed", "skipped"] if stop_on_failure else ["failed", "queued"])
+    assert enqueue.await_count == (0 if stop_on_failure else 1)
+
+
+@pytest.mark.asyncio
+async def test_queued_case_does_not_consume_execution_timeout(isolated, monkeypatch):
+    from apps.admin_console import server
+
+    batch = service.create_batch(payload(timeout_seconds=1))
+    service.batch_action(batch["id"], "resume")
+    first = isolated.jobs(batch["id"])[0]
+    isolated.save_job({**first, "status": "queued", "started_at": time.time() - 30})
+    monkeypatch.setattr(service, "state", type("State", (), {"is_paused": False})())
+    stop = Mock()
+    monkeypatch.setattr(server.task_queue_service, "stop_tasks", stop)
+
+    await service.tick_batches()
+    stop.assert_not_called()
+    assert isolated.job(first["session_id"])["status"] == "queued"
+    assert isolated.get("batch", batch["id"])["status"] == "running"
+
+
 @pytest.mark.asyncio
 async def test_model_timeout_from_runner_log_reaches_failed_task_and_report(isolated):
     job = isolated.jobs(service.create_batch(payload())["id"])[0]
@@ -365,6 +420,33 @@ async def test_model_timeout_from_runner_log_reaches_failed_task_and_report(isol
     assert "超时" in reports.session_report(job["session_id"])["error"]
 
 
+def test_optional_timeout_warnings_are_not_reported_as_task_failure(tmp_path):
+    (tmp_path / "stdout.log").write_text(
+        "⚠ Could not clean unowned screen wake locks on emulator-5554: "
+        "Timed out acquiring the Artemis awake-lease mutex.\n"
+        "⚠ VisualStepSummarizer: Error generating summary for step 3: \n"
+        "ℹ StepMemoryService: Retrying Step 3 (attempt 4) after 1.0s\n",
+        encoding="utf-8",
+    )
+    assert reports.failure_reason(tmp_path, "failed", None) == (
+        "任务未完成，运行日志中没有可提取的具体异常；请查看执行步骤。"
+    )
+
+    (tmp_path / "stderr.log").write_text(
+        "TimeoutError: model request timed out after 300 seconds\n", encoding="utf-8"
+    )
+    reason = reports.failure_reason(tmp_path, "failed", None)
+    assert "model request timed out" in reason
+    assert "awake-lease" not in reason
+
+    (tmp_path / "stderr.log").write_text(
+        "DeviceBusyError: device queue reservation was cancelled\n", encoding="utf-8"
+    )
+    assert reports.failure_reason(tmp_path, "failed", None) == (
+        "DeviceBusyError: device queue reservation was cancelled"
+    )
+
+
 def test_reports_do_not_execute_formulas_or_invent_pass(isolated):
     batch = service.create_batch(payload())
     job = isolated.jobs(batch["id"])[0]

tests/unit/agents/test_flash_runner_ledger.py

@@ -636,6 +636,32 @@ async def test_run_no_tool_call_turn_is_nudged_not_terminated(mock_context):
         and any(EXECUTION_RESULT_MARKER in b.get("text", "") for b in m.content)
         for m in turn2_messages
     )
+    assert not any(
+        isinstance(m, AIMessage) and m.content == "thinking only" for m in turn2_messages
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("attempts", [1, 3])
+async def test_flash_stops_repeated_no_tool_responses(mock_context, monkeypatch, attempts):
+    from artemis.config.settings import settings
+
+    monkeypatch.setattr(settings, "ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS", attempts)
+    responses = [AIMessage(content="I will click now.", tool_calls=[]) for _ in range(attempts)]
+    with (
+        patch("artemis.controllers.unified_controller.get_driver"),
+        patch(
+            "artemis.agents.flash.runner.capture_screenshot_and_parse_ui",
+            AsyncMock(return_value=_INITIAL_OBSERVATION),
+        ),
+    ):
+        runner = _make_runner(mock_context, responses, max_turns=0)
+        result = await runner.run(State(initial_goal="Open Settings"))
+
+    assert result["status"] == "failed"
+    assert f"连续 {attempts} 次未调用工具" in result["explanation"]
+    assert runner._invoke_model.await_count == attempts
+    runner.executor.execute.assert_not_awaited()
 
 
 # ---------------------------------------------------------------------------

tests/unit/agents/test_flash_step_summarizer.py

@@ -14,6 +14,7 @@
 
 """Unit tests for VisualStepSummarizer and Context Compressor in Flash profile."""
 
+import asyncio
 from unittest.mock import ANY, AsyncMock, Mock, call
 from uuid import uuid4
 import pytest
@@ -106,6 +107,29 @@ async def test_summarizer_retries_independently(mock_context):
     assert attempt_count == 2
 
 
+@pytest.mark.asyncio
+async def test_visual_summary_uses_its_own_configured_timeout(mock_context):
+    summarizer = VisualStepSummarizer(mock_context, retry_limit=0, timeout_seconds=0.01)
+
+    async def slow_response(_messages):
+        await asyncio.sleep(1)
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=slow_response)
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [100, 100]},
+        pre_img_bytes=b"pre",
+        post_img_bytes=b"post",
+        exec_outcome="Clicked",
+    )
+
+    await summarizer.flush()
+    assert summarizer.has_failed(1)
+    assert summarizer._llm.ainvoke.await_count == 1
+
+
 @pytest.mark.asyncio
 async def test_summarizer_bounded_retry_marks_failed(mock_context):
     """Persistent failures exit after 1 + retry_limit attempts with an explicit failed state."""

tests/unit/runtime/test_awake_lease.py

@@ -13,6 +13,8 @@
 # limitations under the License.
 
 import json
+from pathlib import Path
+import time
 
 import pytest
 
@@ -126,3 +128,30 @@ def test_migration_cleanup_never_drains_a_live_legacy_owner(monkeypatch):
     cleanup._drain_references = lambda: pytest.fail("must not drain a live owner")
 
     assert cleanup.cleanup_unowned_references() is False
+
+
+def test_optional_cleanup_skips_a_busy_mutex_without_delaying_task_start(monkeypatch):
+    owner = ScreenAwakeLease("device-123")
+    owner.mutex_path.write_text(json.dumps(owner._payload()), encoding="utf-8")
+    cleanup = ScreenAwakeLease("device-123")
+    monkeypatch.setattr(cleanup, "_drain_references", lambda: pytest.fail("must not drain"))
+
+    started_at = time.monotonic()
+    assert cleanup.cleanup_unowned_references() is False
+    assert time.monotonic() - started_at < 2.0
+
+
+def test_unremovable_stale_mutex_still_obeys_deadline(monkeypatch):
+    lease = ScreenAwakeLease("device-123")
+    lease.mutex_path.write_text(json.dumps({"pid": 999999}), encoding="utf-8")
+    original_unlink = Path.unlink
+
+    def refuse_mutex_unlink(path, *args, **kwargs):
+        if path == lease.mutex_path:
+            raise PermissionError("file is in use")
+        return original_unlink(path, *args, **kwargs)
+
+    monkeypatch.setattr(Path, "unlink", refuse_mutex_unlink)
+    with pytest.raises(TimeoutError, match="awake-lease mutex"):
+        with lease._mutex(wait_timeout_seconds=0.1):
+            pytest.fail("stale mutex must not be acquired")

tests/unit/runtime/test_awake_service.py

@@ -1,3 +1,4 @@
+import os
 from unittest.mock import MagicMock, patch
 
 from artemis.runtime.awake_service import (
@@ -37,6 +38,22 @@ def test_usb_policy_is_primary_when_android_reports_it_active(mock_run, lease_ty
     assert not any("KEYCODE_UNKNOWN" in command for command in commands)
 
 
+@patch.dict(os.environ, {"ARTEMIS_TASK_WORKER": "1"})
+@patch("artemis.runtime.awake_service.ScreenAwakeLease")
+@patch("artemis.runtime.awake_service.subprocess.run")
+def test_queued_worker_does_not_run_optional_legacy_cleanup(mock_run, lease_type):
+    mock_run.side_effect = [
+        completed(),
+        completed(),
+        completed(),
+        completed("2\n"),
+        completed("  mStayOn=true\n"),
+    ]
+
+    assert _configure_usb_stay_awake("device-123") == AWAKE_STRATEGY_USB
+    lease_type.assert_not_called()
+
+
 @patch("artemis.runtime.awake_service.ScreenAwakeLease")
 @patch("artemis.runtime.awake_service.subprocess.run")
 def test_inactive_usb_policy_falls_back_to_effective_host_heartbeat(mock_run, _lease_type):

uv.lock

@@ -206,7 +206,7 @@ wheels = [
 
 [[package]]
 name = "artemis"
-version = "1.2.0"
+version = "1.2.2"
 source = { editable = "." }
 dependencies = [
     { name = "adbutils" },