from __future__ import annotations import json from typing import Any from app.schemas.steward import ( StewardRuntimeDecisionRequest, StewardRuntimeDecisionResponse, ) from app.services.runtime_chat import RuntimeChatService STEWARD_RUNTIME_DECISION_FUNCTION_NAME = "submit_steward_runtime_decision" RUNTIME_NEXT_ACTIONS = { "plan_new_tasks", "submit_current_application", "continue_next_task", "fill_current_slot", "ask_user", "cancel_current_action", "no_op", } class StewardRuntimeDecisionAgent: """用小财管家运行时上下文判断用户当前输入应落到哪个等待动作。""" def __init__(self, runtime_chat_service: RuntimeChatService) -> None: self.runtime_chat_service = runtime_chat_service def decide(self, request: StewardRuntimeDecisionRequest) -> StewardRuntimeDecisionResponse: normalized_request = self._normalize_request(request) result = self.runtime_chat_service.complete_with_tool_call( self._build_messages(normalized_request), tools=[self._build_tool_schema()], tool_choice={ "type": "function", "function": {"name": STEWARD_RUNTIME_DECISION_FUNCTION_NAME}, }, max_tokens=1000, temperature=0.05, timeout_seconds=30, max_attempts=1, ) traces = result.calls_as_dicts() if result.tool_call is not None and result.tool_call.name == STEWARD_RUNTIME_DECISION_FUNCTION_NAME: response = self._build_response_from_model_payload(result.tool_call.arguments, normalized_request, traces) if response is not None: return response return self._build_rule_fallback(normalized_request, traces) @staticmethod def _normalize_request(request: StewardRuntimeDecisionRequest) -> StewardRuntimeDecisionRequest: return StewardRuntimeDecisionRequest( user_message=str(request.user_message or "").strip(), session_type=str(request.session_type or "steward").strip() or "steward", runtime_state=request.runtime_state if isinstance(request.runtime_state, dict) else {}, context_json=request.context_json if isinstance(request.context_json, dict) else {}, ) @staticmethod def _build_messages(request: StewardRuntimeDecisionRequest) -> list[dict[str, Any]]: payload = { "user_message": request.user_message, "session_type": request.session_type, "runtime_state": request.runtime_state, "context_json": request.context_json, } return [ { "role": "system", "content": ( "你是 X-Financial 小财管家的运行时决策智能体。" "你必须基于 runtime_state 判断用户当前输入对应哪个等待动作，不能把每次输入都当成全新任务。" "runtime_state 会包含 current_task、remaining_tasks、completed_tasks、pending_application、" "pending_steward_action、waiting_for、recent_structured_result 等上下文。" "如果用户是在确认当前申请核对表无误，应返回 submit_current_application；" "如果用户是在确认继续下一项，应返回 continue_next_task；" "如果用户补充了当前等待字段，应返回 fill_current_slot；" "如果当前结构化结果仍缺字段，应返回 ask_user；" "只有当前没有可匹配上下文，且用户输入明显是新财务事项时，才返回 plan_new_tasks。" "提交、入库、绑定、审批等高风险动作只返回结构化意图，实际执行由系统安全校验完成。" "rationale 和 response_text 必须面向用户，不暴露内部推理链。" ), }, {"role": "user", "content": json.dumps(payload, ensure_ascii=False)}, ] @staticmethod def _build_tool_schema() -> dict[str, Any]: return { "type": "function", "function": { "name": STEWARD_RUNTIME_DECISION_FUNCTION_NAME, "description": "提交小财管家基于运行时上下文的下一步动作决策。", "parameters": { "type": "object", "properties": { "next_action": { "type": "string", "enum": sorted(RUNTIME_NEXT_ACTIONS), }, "target_task_id": {"type": "string"}, "target_message_id": {"type": "string"}, "field_key": {"type": "string"}, "field_value": {"type": "string"}, "confirmation_required": {"type": "boolean"}, "question": {"type": "string"}, "response_text": {"type": "string"}, "rationale": {"type": "string"}, }, "required": [ "next_action", "target_task_id", "target_message_id", "field_key", "field_value", "confirmation_required", "question", "response_text", "rationale", ], }, }, } def _build_response_from_model_payload( self, payload: dict[str, Any], request: StewardRuntimeDecisionRequest, traces: list[dict[str, Any]], ) -> StewardRuntimeDecisionResponse | None: next_action = str(payload.get("next_action") or "").strip() if next_action not in RUNTIME_NEXT_ACTIONS: return None return StewardRuntimeDecisionResponse( decision_source="llm_function_call", next_action=next_action, # type: ignore[arg-type] target_task_id=self._clean_text(payload.get("target_task_id")), target_message_id=self._clean_text(payload.get("target_message_id")), field_key=self._clean_text(payload.get("field_key")), field_value=self._clean_text(payload.get("field_value")), confirmation_required=bool(payload.get("confirmation_required")), question=self._clean_text(payload.get("question")), response_text=self._clean_text(payload.get("response_text")), rationale=self._clean_text(payload.get("rationale")), model_call_traces=traces, ) def _build_rule_fallback( self, request: StewardRuntimeDecisionRequest, traces: list[dict[str, Any]], ) -> StewardRuntimeDecisionResponse: state = request.runtime_state pending_application = state.get("pending_application") if isinstance(state.get("pending_application"), dict) else {} pending_steward_action = state.get("pending_steward_action") if isinstance(state.get("pending_steward_action"), dict) else {} waiting_for = str(state.get("waiting_for") or "").strip() message = request.user_message.replace(" ", "") confirmation_text = message in {"确认", "确定", "无误", "确认提交", "可以提交", "提交", "没问题"} if confirmation_text and pending_application.get("ready_to_submit"): return StewardRuntimeDecisionResponse( decision_source="rule_fallback", next_action="submit_current_application", target_message_id=str(pending_application.get("message_id") or ""), target_task_id=str(pending_application.get("task_id") or ""), rationale="模型运行时决策暂不可用，我先按当前待提交申请单上下文处理你的确认。", model_call_traces=traces, ) if confirmation_text and pending_steward_action: return StewardRuntimeDecisionResponse( decision_source="rule_fallback", next_action="continue_next_task", target_message_id=str(pending_steward_action.get("message_id") or ""), target_task_id=str(pending_steward_action.get("target_task_id") or ""), rationale="模型运行时决策暂不可用，我先按当前待确认的下一项任务继续处理。", model_call_traces=traces, ) if waiting_for: return StewardRuntimeDecisionResponse( decision_source="rule_fallback", next_action="ask_user", question="我需要先确认当前等待事项，请补充或选择当前问题对应的信息。", rationale="模型运行时决策暂不可用，当前仍存在等待用户补充的信息。", model_call_traces=traces, ) return StewardRuntimeDecisionResponse( decision_source="rule_fallback", next_action="plan_new_tasks", rationale="模型运行时决策暂不可用，当前没有可安全匹配的等待动作，回到任务规划。", model_call_traces=traces, ) @staticmethod def _clean_text(value: Any) -> str: return str(value or "").strip()