fix: bound task failures and release stalled workers

artemis/agents/flash/runner.py

@@ -81,7 +81,7 @@ from artemis.graph.state import State
 from artemis.llm.router import ModelEndpoint
 from artemis.llm.structured import ParseFailure, parse_structured
 from artemis.mcp.action_executor import McpActionExecutor
-from artemis.mcp.observation import observe
+from artemis.mcp.observation import DeviceObservationError, require_observation
 from artemis.memory.transcript import PRO_UI_LIST_MARKER, TranscriptLedger, mark_ephemeral
 from artemis.services.llm import (
     RobustChatModelWrapper,
@@ -380,8 +380,16 @@ class FlashRunner:
             blocks.append(
                 {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img_b64}"}}
             )
-        if xml_list:
-            blocks.append({"type": "text", "text": f"{PRO_UI_LIST_MARKER}\n{xml_list}"})
+        blocks.append(
+            {
+                "type": "text",
+                "text": f"{PRO_UI_LIST_MARKER}\n"
+                + (
+                    xml_list
+                    or "当前没有可用的元素索引。仅根据当前截图定位坐标，禁止使用历史画面的元素编号。"
+                ),
+            }
+        )
         ephemeral: list[int] = []
         # (text, ephemeral) in tail order.
         per_turn: list[tuple[str, bool]] = [(text, True) for text in (notices or [])]
@@ -604,19 +612,14 @@ class FlashRunner:
         post_img_bytes = exec_result.screenshot_bytes
         if post_img_bytes or name not in action_names:
             return post_img_bytes
-        try:
-            obs, screenshot_bytes = await observe(self.ctx, self.controller, settle_ms=0)
-        except Exception as shot_err:
-            logger.warning(f"Failed to capture fallback screenshot in FlashRunner: {shot_err}")
-            return post_img_bytes
-        if not obs.ok:
-            logger.warning(f"Failed to capture fallback screenshot in FlashRunner: {obs.message}")
-            return post_img_bytes
+        if state is not None:
+            state.indexed_points = []
+            state.indexed_elements = []
+            state.latest_screenshot = None
+        obs, screenshot_bytes = await require_observation(self.ctx, self.controller, settle_ms=0)
         post_img_bytes = screenshot_bytes
-        if not exec_result.ui_elements_text:
-            exec_result.ui_elements_text = obs.elements_text
-        if not exec_result.screenshot_path and obs.screenshot_path:
-            exec_result.screenshot_path = obs.screenshot_path
+        exec_result.ui_elements_text = obs.elements_text
+        exec_result.screenshot_path = obs.screenshot_path
         if state is not None:
             if obs.hierarchy_ok:
                 state.indexed_points = [el["center"] for el in obs.elements]
@@ -831,7 +834,7 @@ class FlashRunner:
                 turn.actions.append((name, exec_result.status, exec_result.text_summary))
 
             # ⚡ Non-blocking dispatch of objective visual transition summarizer
-            if self.summarizer and name in action_names:
+            if self.summarizer and name in action_names and post_img_bytes:
                 action_sequence += 1
                 self.summarizer.dispatch(
                     step_number=action_sequence,
@@ -857,7 +860,7 @@ class FlashRunner:
 
             if post_img_bytes:
                 pre_screenshot_bytes = post_img_bytes
-            if exec_result.ui_elements_text:
+            if exec_result.ui_elements_text is not None:
                 xml_list = exec_result.ui_elements_text
 
             # Helper tools may return multimodal blocks (a step screenshot);
@@ -876,6 +879,8 @@ class FlashRunner:
                 )
             )
 
+        except DeviceObservationError:
+            raise
         except Exception as e:
             logger.error(f"Error executing tool {name}: {e}")
             if name in action_names:
@@ -1047,6 +1052,21 @@ class FlashRunner:
 
     @trace(type="agent", name="FlashRunner")
     async def run(self, state: State) -> dict:
+        try:
+            return await self._run(state)
+        except DeviceObservationError as exc:
+            state.indexed_points = []
+            state.indexed_elements = []
+            state.latest_screenshot = None
+            logger.error(f"Flash 因设备观测失败停止：{exc}")
+            return {
+                "status": "failed",
+                "explanation": str(exc),
+                "error": str(exc),
+                "code": "device_observation_failed",
+            }
+
+    async def _run(self, state: State) -> dict:
         limit = self.turn_limit
         logger.info(
             f"Starting Artemis Flash reactive loop for goal: {self.goal}"
@@ -1073,6 +1093,7 @@ class FlashRunner:
         no_tool_limit = settings.ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS
         fallback_llm = None
         using_tool_fallback = False
+        action_failure_attempts = 0
 
         while limit is None or turns < limit:
             control = getattr(self.ctx, "task_control", None)
@@ -1084,6 +1105,7 @@ class FlashRunner:
                 if resumed:
                     using_tool_fallback = False
                     no_tool_attempts = 0
+                    action_failure_attempts = 0
                     self._llm = llm
                     (
                         shot_path,
@@ -1157,6 +1179,7 @@ class FlashRunner:
                 # assembled before the instruction arrived.
                 using_tool_fallback = False
                 no_tool_attempts = 0
+                action_failure_attempts = 0
                 self._llm = llm
                 (
                     shot_path,
@@ -1241,6 +1264,7 @@ class FlashRunner:
                     fallback_llm = get_llm(self.ctx, name="operator", use_fallback=True)
                     using_tool_fallback = True
                     no_tool_attempts = 0
+                    action_failure_attempts = 0
                     logger.warning("Flash 首选模型未调用工具，当前步骤切换到兜底模型")
                     pending_notices.append("首选模型连续未调用工具；本步骤改用兜底模型。")
                 pending_notices.append(_NO_TOOL_CALL_NOTICE)
@@ -1279,6 +1303,37 @@ class FlashRunner:
                 using_tool_fallback = False
                 self._llm = llm
                 pending_notices.append("兜底模型已完成本步骤；后续步骤恢复首选模型。")
+            for _, status, _ in turn.actions:
+                action_failure_attempts = 0 if status == "success" else action_failure_attempts + 1
+            failure_limit = (
+                settings.ARTEMIS_TOOL_FALLBACK_MAX_NO_PROGRESS_ATTEMPTS
+                if using_tool_fallback
+                else settings.ARTEMIS_ACTION_FAILURE_ATTEMPTS
+            )
+            if action_failure_attempts >= failure_limit:
+                if using_tool_fallback:
+                    final_report = {
+                        "status": "failed",
+                        "explanation": f"兜底模型连续 {action_failure_attempts} 次设备动作失败，已停止任务。最后原因：{turn.result()}",
+                    }
+                    break
+                # A fresh, valid observation is required before spending any
+                # fallback model calls. Hardware recovery is not model retry.
+                (
+                    shot_path,
+                    current_pre_screenshot_bytes,
+                    current_xml_list,
+                ) = await capture_screenshot_and_parse_ui(self.ctx, state, self.controller)
+                state.latest_screenshot = shot_path
+                fallback_llm = get_llm(self.ctx, name="operator", use_fallback=True)
+                using_tool_fallback = True
+                self._llm = fallback_llm
+                action_failure_attempts = 0
+                no_tool_attempts = 0
+                pending_notices.append(
+                    f"首选模型连续 {settings.ARTEMIS_ACTION_FAILURE_ATTEMPTS} 次设备动作失败。"
+                    f"最后原因：{turn.result()}。已重新观察；本步骤由兜底模型重新定位，勿重复旧动作。"
+                )
             ledger.stage_turn(messages[turn_base:])
             previous_turn = turn
             if final_report_from_calls is not None: