feat: add tool-call fallback and resilient batch exports

artemis/agents/operator/operator.py

@@ -508,6 +508,9 @@ class OperatorNode:
         plan_gate_bounced = False
         no_progress_attempts = 0
         no_progress_limit = settings.ARTEMIS_OPERATOR_MAX_NO_PROGRESS_ATTEMPTS
+        fallback_limit = settings.ARTEMIS_TOOL_FALLBACK_MAX_NO_PROGRESS_ATTEMPTS
+        active_llm = base_llm
+        using_tool_fallback = False
 
         for iteration in range(max_iterations):
             control = getattr(self.ctx, "task_control", None)
@@ -532,7 +535,7 @@ class OperatorNode:
                     )
                 )
 
-            bound_llm = base_llm.bind_tools(tools=traced_tools)
+            bound_llm = active_llm.bind_tools(tools=traced_tools)
             interrupted, response = await self._await_or_pause(
                 invoke_llm_with_timeout_message(acomplete(bound_llm, current_messages))
             )
@@ -607,11 +610,25 @@ class OperatorNode:
 
             if not response.tool_calls:
                 no_progress_attempts += 1
-                if no_progress_attempts >= no_progress_limit:
-                    raise OperatorNoProgressError(
-                        f"模型连续 {no_progress_attempts} 次未提供可执行的工具调用，"
-                        "已停止任务以避免空转；未执行新的设备动作。"
+                if no_progress_attempts >= (
+                    fallback_limit if using_tool_fallback else no_progress_limit
+                ):
+                    if using_tool_fallback:
+                        raise OperatorNoProgressError(
+                            f"首选模型连续 {no_progress_limit} 次、兜底模型连续 "
+                            f"{no_progress_attempts} 次未提供可执行的工具调用；"
+                            "已停止任务以避免空转。"
+                        )
+                    active_llm = get_llm(ctx=self.ctx, name="operator", use_fallback=True)
+                    using_tool_fallback = True
+                    no_progress_attempts = 0
+                    logger.warning("Operator 首选模型未调用工具，当前步骤切换到兜底模型")
+                    current_messages.append(
+                        SystemMessage(
+                            content="本步骤已切换兜底模型。请使用工具完成操作或明确报告阻碍。"
+                        )
                     )
+                    continue
                 logger.warning("LLM returned no tool call; requesting a concrete tool call.")
                 current_messages.append(response)
                 current_messages.append(
@@ -931,11 +948,25 @@ class OperatorNode:
 
             if validation_errors:
                 no_progress_attempts += 1
-                if no_progress_attempts >= no_progress_limit:
-                    raise OperatorNoProgressError(
-                        f"模型连续 {no_progress_attempts} 次未提供可执行的工具调用，"
-                        "工具参数校验仍失败；已停止任务，未执行被拒绝的设备动作。"
+                if no_progress_attempts >= (
+                    fallback_limit if using_tool_fallback else no_progress_limit
+                ):
+                    if using_tool_fallback:
+                        raise OperatorNoProgressError(
+                            f"首选模型连续 {no_progress_limit} 次、兜底模型连续 "
+                            f"{no_progress_attempts} 次未提供可执行的工具调用，"
+                            "工具参数校验仍失败；已停止任务。"
+                        )
+                    active_llm = get_llm(ctx=self.ctx, name="operator", use_fallback=True)
+                    using_tool_fallback = True
+                    no_progress_attempts = 0
+                    logger.warning("Operator 首选模型工具调用校验失败，当前步骤切换到兜底模型")
+                    current_messages.append(
+                        SystemMessage(
+                            content="本步骤已切换兜底模型。请修正上面的工具参数后再调用。"
+                        )
                     )
+                    continue
                 logger.warning("Some action calls failed validation. Feeding back to LLM.")
                 continue
             no_progress_attempts = 0