feat: add tool-call fallback and resilient batch exports

artemis/agents/flash/runner.py

@@ -1071,6 +1071,8 @@ class FlashRunner:
         pending_notices: list[str] = []
         no_tool_attempts = 0
         no_tool_limit = settings.ARTEMIS_FLASH_MAX_NO_PROGRESS_ATTEMPTS
+        fallback_llm = None
+        using_tool_fallback = False
 
         while limit is None or turns < limit:
             control = getattr(self.ctx, "task_control", None)
@@ -1080,6 +1082,9 @@ class FlashRunner:
                     on_resume=getattr(self.ctx, "on_task_resume", None),
                 )
                 if resumed:
+                    using_tool_fallback = False
+                    no_tool_attempts = 0
+                    self._llm = llm
                     (
                         shot_path,
                         current_pre_screenshot_bytes,
@@ -1131,14 +1136,18 @@ class FlashRunner:
 
             if isinstance(control, TaskControl):
                 interrupted, response = await control.run_interruptibly(
-                    self._invoke_model(llm, current_tools, messages),
+                    self._invoke_model(
+                        fallback_llm if using_tool_fallback else llm, current_tools, messages
+                    ),
                     on_pause=getattr(self.ctx, "on_task_pause", None),
                     on_resume=getattr(self.ctx, "on_task_resume", None),
                 )
             else:
                 interrupted, response = (
                     False,
-                    await self._invoke_model(llm, current_tools, messages),
+                    await self._invoke_model(
+                        fallback_llm if using_tool_fallback else llm, current_tools, messages
+                    ),
                 )
 
             if interrupted or (
@@ -1146,6 +1155,9 @@ class FlashRunner:
             ):
                 # Do not dispatch a stale action or final answer from a prompt
                 # assembled before the instruction arrived.
+                using_tool_fallback = False
+                no_tool_attempts = 0
+                self._llm = llm
                 (
                     shot_path,
                     current_pre_screenshot_bytes,
@@ -1186,6 +1198,9 @@ class FlashRunner:
 
             tool_calls = self._resolve_tool_calls(response, raw_text)
             if isinstance(control, TaskControl) and has_pending_instruction(control.directory):
+                using_tool_fallback = False
+                no_tool_attempts = 0
+                self._llm = llm
                 (
                     shot_path,
                     current_pre_screenshot_bytes,
@@ -1207,15 +1222,27 @@ class FlashRunner:
                     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
+                active_limit = (
+                    settings.ARTEMIS_TOOL_FALLBACK_MAX_NO_PROGRESS_ATTEMPTS
+                    if using_tool_fallback
+                    else no_tool_limit
+                )
+                if no_tool_attempts >= active_limit:
+                    if using_tool_fallback:
+                        final_report = {
+                            "status": "failed",
+                            "explanation": (
+                                f"首选模型连续 {no_tool_limit} 次、兜底模型连续 "
+                                f"{no_tool_attempts} 次未调用工具，已停止任务；"
+                                "未执行新的设备动作。"
+                            ),
+                        }
+                        break
+                    fallback_llm = get_llm(self.ctx, name="operator", use_fallback=True)
+                    using_tool_fallback = True
+                    no_tool_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)
                 # Keep the observation and any injected user instruction, but
                 # do not feed the model its own unproductive prose on retry.
@@ -1224,6 +1251,10 @@ class FlashRunner:
                 continue
 
             no_tool_attempts = 0
+            if using_tool_fallback:
+                # A tool-producing turn completes the fallback's scope. The next
+                # observation goes back to the configured primary model.
+                self._llm = fallback_llm
             # Process tool calls
             turn = _TurnRecord()
             (
@@ -1244,6 +1275,10 @@ class FlashRunner:
                 injected=injected,
                 native_text=native_text,
             )
+            if using_tool_fallback and any(status == "success" for _, status, _ in turn.actions):
+                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。")
             ledger.stage_turn(messages[turn_base:])
             previous_turn = turn
             if final_report_from_calls is not None: