"""Maisaka 推理引擎。""" from datetime import datetime from typing import TYPE_CHECKING, Any, Optional import asyncio import difflib import json import time import traceback from src.chat.heart_flow.heartFC_utils import CycleDetail from src.chat.message_receive.message import SessionMessage from src.chat.utils.utils import process_llm_response from src.common.data_models.message_component_data_model import EmojiComponent, ImageComponent, MessageSequence, TextComponent from src.common.logger import get_logger from src.config.config import global_config from src.core.tooling import ToolExecutionContext, ToolExecutionResult, ToolInvocation, ToolSpec from src.llm_models.exceptions import ReqAbortException from src.llm_models.payload_content.tool_option import ToolCall from src.services import database_service as database_api from .builtin_tool import build_builtin_tool_handlers as build_split_builtin_tool_handlers from .builtin_tool.context import BuiltinToolRuntimeContext from .context_messages import ( AssistantMessage, LLMContextMessage, SessionBackedMessage, ToolResultMessage, ) from .message_adapter import ( build_visible_text_from_sequence, clone_message_sequence, format_speaker_content, ) from .planner_message_utils import build_planner_user_prefix_from_session_message if TYPE_CHECKING: from .runtime import MaisakaHeartFlowChatting from .tool_provider import BuiltinToolHandler logger = get_logger("maisaka_reasoning_engine") class MaisakaReasoningEngine: """负责内部思考、推理与工具执行。""" def __init__(self, runtime: "MaisakaHeartFlowChatting") -> None: self._runtime = runtime self._last_reasoning_content: str = "" @property def last_reasoning_content(self) -> str: """返回最近一轮思考文本。""" return self._last_reasoning_content def build_builtin_tool_handlers(self) -> dict[str, "BuiltinToolHandler"]: """构造 Maisaka 内置工具处理器映射。 Returns: dict[str, BuiltinToolHandler]: 工具名到处理器的映射。 """ return build_split_builtin_tool_handlers(BuiltinToolRuntimeContext(self, self._runtime)) async def run_loop(self) -> None: """独立消费消息批次，并执行对应的内部思考轮次。""" try: while self._runtime._running: cached_messages = await self._runtime._internal_turn_queue.get() timeout_triggered = cached_messages is None if not timeout_triggered and not cached_messages: self._runtime._internal_turn_queue.task_done() continue self._runtime._agent_state = self._runtime._STATE_RUNNING if cached_messages: self._clear_pending_wait_tool_call_id() await self._ingest_messages(cached_messages) anchor_message = cached_messages[-1] else: anchor_message = self._get_timeout_anchor_message() if anchor_message is None: logger.warning( f"{self._runtime.log_prefix} 等待超时后缺少可复用的锚点消息，跳过本轮继续思考" ) self._runtime._internal_turn_queue.task_done() continue logger.info(f"{self._runtime.log_prefix} 等待超时后开始新一轮思考") self._clear_pending_wait_tool_call_id() self._trim_chat_history() try: for round_index in range(self._runtime._max_internal_rounds): cycle_detail = self._start_cycle() self._runtime._log_cycle_started(cycle_detail, round_index) planner_started_at = time.time() try: logger.info( f"{self._runtime.log_prefix} 规划器开始执行: " f"回合={round_index + 1} " f"历史消息数={len(self._runtime._chat_history)} " f"开始时间={planner_started_at:.3f}" ) interrupt_flag = asyncio.Event() self._runtime._planner_interrupt_flag = interrupt_flag self._runtime._chat_loop_service.set_interrupt_flag(interrupt_flag) try: response = await self._runtime._chat_loop_service.chat_loop_step(self._runtime._chat_history) finally: if self._runtime._planner_interrupt_flag is interrupt_flag: self._runtime._planner_interrupt_flag = None self._runtime._chat_loop_service.set_interrupt_flag(None) cycle_detail.time_records["planner"] = time.time() - planner_started_at logger.info( f"{self._runtime.log_prefix} 规划器执行完成: " f"回合={round_index + 1} " f"耗时={cycle_detail.time_records['planner']:.3f} 秒" ) reasoning_content = response.content or "" if self._should_replace_reasoning(reasoning_content): response.content = "让我根据新情况重新思考：" response.raw_message.content = "让我根据新情况重新思考：" logger.info(f"{self._runtime.log_prefix} 当前思考与上一轮过于相似，已替换为重新思考提示") self._last_reasoning_content = reasoning_content self._runtime._render_context_usage_panel( selected_history_count=response.selected_history_count, prompt_tokens=response.prompt_tokens, ) self._runtime._chat_history.append(response.raw_message) if response.tool_calls: tool_started_at = time.time() should_pause = await self._handle_tool_calls( response.tool_calls, response.content or "", anchor_message, ) cycle_detail.time_records["tool_calls"] = time.time() - tool_started_at if should_pause: break continue if response.content: continue break except ReqAbortException: interrupted_at = time.time() logger.info( f"{self._runtime.log_prefix} 规划器打断成功: " f"回合={round_index + 1} " f"开始时间={planner_started_at:.3f} " f"打断时间={interrupted_at:.3f} " f"耗时={interrupted_at - planner_started_at:.3f} 秒" ) break finally: self._end_cycle(cycle_detail) finally: if self._runtime._agent_state == self._runtime._STATE_RUNNING: self._runtime._agent_state = self._runtime._STATE_STOP self._runtime._internal_turn_queue.task_done() except asyncio.CancelledError: self._runtime._log_internal_loop_cancelled() raise except Exception: logger.exception(f"{self._runtime.log_prefix} Maisaka 内部循环发生异常") logger.error(traceback.format_exc()) raise def _get_timeout_anchor_message(self) -> Optional[SessionMessage]: """在 wait 超时后复用最近一条真实用户消息作为锚点。""" if self._runtime.message_cache: return self._runtime.message_cache[-1] return None def _clear_pending_wait_tool_call_id(self) -> None: """清理等待状态残留的 wait 工具调用编号。""" self._runtime._pending_wait_tool_call_id = None def _build_wait_timeout_message(self) -> ToolResultMessage: """构造 wait 超时后的工具结果消息。""" tool_call_id = self._runtime._pending_wait_tool_call_id or "wait_timeout" self._runtime._pending_wait_tool_call_id = None return ToolResultMessage( content="等待已超时，期间没有收到新的用户输入。请基于现有上下文继续下一轮思考。", timestamp=datetime.now(), tool_call_id=tool_call_id, tool_name="wait", ) def _append_wait_interrupted_message_if_needed(self) -> None: """如果 wait 被新消息打断，则补一条对应的工具结果消息。""" tool_call_id = self._runtime._pending_wait_tool_call_id if not tool_call_id: return self._runtime._pending_wait_tool_call_id = None self._runtime._chat_history.append( ToolResultMessage( content="等待过程被新的用户输入打断，已继续处理最新消息。", timestamp=datetime.now(), tool_call_id=tool_call_id, tool_name="wait", ) ) async def _ingest_messages(self, messages: list[SessionMessage]) -> None: """处理传入消息列表，将其转换为历史消息并加入聊天历史缓存。""" for message in messages: # 构建用户消息序列 user_sequence, visible_text = await self._build_message_sequence(message) if not user_sequence.components: continue history_message = SessionBackedMessage.from_session_message( message, raw_message=user_sequence, visible_text=visible_text, source_kind="user", ) self._insert_chat_history_message(history_message) self._trim_chat_history() async def _build_message_sequence(self, message: SessionMessage) -> tuple[MessageSequence, str]: message_sequence = MessageSequence([]) planner_prefix = build_planner_user_prefix_from_session_message(message) appended_component = False source_sequence = message.raw_message planner_components = clone_message_sequence(source_sequence).components if global_config.maisaka.direct_image_input: await self._hydrate_visual_components(planner_components) if planner_components and isinstance(planner_components[0], TextComponent): planner_components[0].text = planner_prefix + planner_components[0].text else: planner_components.insert(0, TextComponent(planner_prefix)) for component in planner_components: message_sequence.components.append(component) appended_component = True legacy_visible_text = self._build_legacy_visible_text(message, source_sequence) if not appended_component: if not message.processed_plain_text: await message.process() content = (message.processed_plain_text or "").strip() if content: message_sequence.text(planner_prefix + content) legacy_visible_text = self._build_legacy_visible_text_from_text(message, content) return message_sequence, legacy_visible_text async def _hydrate_visual_components(self, planner_components: list[object]) -> None: """在 Maisaka 真正需要图片或表情时，按需回填二进制数据。""" load_tasks: list[asyncio.Task[None]] = [] for component in planner_components: if isinstance(component, ImageComponent) and not component.binary_data: load_tasks.append(asyncio.create_task(component.load_image_binary())) continue if isinstance(component, EmojiComponent) and not component.binary_data: load_tasks.append(asyncio.create_task(component.load_emoji_binary())) if not load_tasks: return results = await asyncio.gather(*load_tasks, return_exceptions=True) for result in results: if isinstance(result, Exception): logger.warning(f"{self._runtime.log_prefix} 回填图片或表情二进制数据失败，Maisaka 将退化为文本占位: {result}") def _build_legacy_visible_text(self, message: SessionMessage, source_sequence: MessageSequence) -> str: user_info = message.message_info.user_info speaker_name = user_info.user_cardname or user_info.user_nickname or user_info.user_id legacy_sequence = MessageSequence([]) visible_message_id = None if message.is_notify else message.message_id legacy_sequence.text(format_speaker_content(speaker_name, "", message.timestamp, visible_message_id)) for component in clone_message_sequence(source_sequence).components: legacy_sequence.components.append(component) return build_visible_text_from_sequence(legacy_sequence).strip() def _build_legacy_visible_text_from_text(self, message: SessionMessage, content: str) -> str: user_info = message.message_info.user_info speaker_name = user_info.user_cardname or user_info.user_nickname or user_info.user_id visible_message_id = None if message.is_notify else message.message_id return format_speaker_content(speaker_name, content, message.timestamp, visible_message_id).strip() def _insert_chat_history_message(self, message: LLMContextMessage) -> int: """将消息按处理顺序追加到聊天历史末尾。""" self._runtime._chat_history.append(message) return len(self._runtime._chat_history) - 1 def _start_cycle(self) -> CycleDetail: """开始一轮 Maisaka 思考循环。""" self._runtime._cycle_counter += 1 self._runtime._current_cycle_detail = CycleDetail(cycle_id=self._runtime._cycle_counter) self._runtime._current_cycle_detail.thinking_id = f"maisaka_tid{round(time.time(), 2)}" return self._runtime._current_cycle_detail def _end_cycle(self, cycle_detail: CycleDetail, only_long_execution: bool = True) -> CycleDetail: """结束并记录一轮 Maisaka 思考循环。""" cycle_detail.end_time = time.time() self._runtime.history_loop.append(cycle_detail) timer_strings = [ f"{name}: {duration:.2f}s" for name, duration in cycle_detail.time_records.items() if not only_long_execution or duration >= 0.1 ] self._runtime._log_cycle_completed(cycle_detail, timer_strings) return cycle_detail def _trim_chat_history(self) -> None: """裁剪聊天历史，保证用户消息数量不超过配置限制。""" conversation_message_count = sum(1 for message in self._runtime._chat_history if message.count_in_context) if conversation_message_count <= self._runtime._max_context_size: return trimmed_history = list(self._runtime._chat_history) removed_count = 0 while conversation_message_count > self._runtime._max_context_size and trimmed_history: removed_message = trimmed_history.pop(0) removed_count += 1 if removed_message.count_in_context: conversation_message_count -= 1 trimmed_history, pruned_orphan_count = self._drop_leading_orphan_tool_results(trimmed_history) removed_count += pruned_orphan_count self._runtime._chat_history = trimmed_history self._runtime._log_history_trimmed(removed_count, conversation_message_count) @staticmethod def _drop_leading_orphan_tool_results( chat_history: list[LLMContextMessage], ) -> tuple[list[LLMContextMessage], int]: """清理历史前缀中缺少对应 assistant tool_call 的工具结果消息。""" if not chat_history: return chat_history, 0 available_tool_call_ids = { tool_call.call_id for message in chat_history if isinstance(message, AssistantMessage) for tool_call in message.tool_calls if tool_call.call_id } first_valid_index = 0 while first_valid_index < len(chat_history): message = chat_history[first_valid_index] if not isinstance(message, ToolResultMessage): break if message.tool_call_id in available_tool_call_ids: break first_valid_index += 1 if first_valid_index == 0: return chat_history, 0 return chat_history[first_valid_index:], first_valid_index @staticmethod def _calculate_similarity(text1: str, text2: str) -> float: """计算两个文本之间的相似度。 Args: text1: 第一个文本 text2: 第二个文本 Returns: float: 相似度值，范围 0-1，1 表示完全相同 """ return difflib.SequenceMatcher(None, text1, text2).ratio() def _should_replace_reasoning(self, current_content: str) -> bool: """判断是否需要替换推理内容。 当当前推理内容与上一次相似度大于90%时，返回True。 Args: current_content: 当前的推理内容 Returns: bool: 是否需要替换 """ if not self._last_reasoning_content or not current_content: logger.info( f"{self._runtime.log_prefix} 跳过思考相似度判定: " f"上一轮为空={not bool(self._last_reasoning_content)} " f"当前为空={not bool(current_content)} 相似度=0.00" ) return False similarity = self._calculate_similarity(current_content, self._last_reasoning_content) logger.info(f"{self._runtime.log_prefix} 思考内容相似度: {similarity:.2f}") return similarity > 0.9 @staticmethod def _post_process_reply_text(reply_text: str) -> list[str]: """沿用旧回复链的文本后处理，执行分段与错别字注入。""" processed_segments: list[str] = [] for segment in process_llm_response(reply_text): normalized_segment = segment.strip() if normalized_segment: processed_segments.append(normalized_segment) if processed_segments: return processed_segments return [reply_text.strip()] def _build_tool_invocation(self, tool_call: ToolCall, latest_thought: str) -> ToolInvocation: """将模型输出的工具调用转换为统一调用对象。 Args: tool_call: 模型返回的工具调用。 latest_thought: 当前轮的最新思考文本。 Returns: ToolInvocation: 统一工具调用对象。 """ return ToolInvocation( tool_name=tool_call.func_name, arguments=dict(tool_call.args or {}), call_id=tool_call.call_id, session_id=self._runtime.session_id, stream_id=self._runtime.session_id, reasoning=latest_thought, ) def _build_tool_execution_context( self, latest_thought: str, anchor_message: SessionMessage, ) -> ToolExecutionContext: """构造统一工具执行上下文。 Args: latest_thought: 当前轮的最新思考文本。 anchor_message: 当前轮的锚点消息。 Returns: ToolExecutionContext: 统一工具执行上下文。 """ return ToolExecutionContext( session_id=self._runtime.session_id, stream_id=self._runtime.session_id, reasoning=latest_thought, metadata={"anchor_message": anchor_message}, ) @staticmethod def _normalize_tool_record_value(value: Any) -> Any: """将工具记录中的任意值规范化为可序列化结构。 Args: value: 原始值。 Returns: Any: 适合写入 JSON 的规范化结果。 """ if value is None or isinstance(value, (str, int, float, bool)): return value if isinstance(value, datetime): return value.isoformat() if isinstance(value, dict): normalized_dict: dict[str, Any] = {} for key, item in value.items(): normalized_dict[str(key)] = MaisakaReasoningEngine._normalize_tool_record_value(item) return normalized_dict if isinstance(value, (list, tuple, set)): return [MaisakaReasoningEngine._normalize_tool_record_value(item) for item in value] if isinstance(value, bytes): return f"" if hasattr(value, "model_dump"): try: return MaisakaReasoningEngine._normalize_tool_record_value(value.model_dump()) except Exception: return str(value) if hasattr(value, "__dict__"): try: return MaisakaReasoningEngine._normalize_tool_record_value(dict(value.__dict__)) except Exception: return str(value) return str(value) @staticmethod def _truncate_tool_record_text(text: str, max_length: int = 180) -> str: """截断工具记录中的展示文本。 Args: text: 原始文本。 max_length: 最长保留字符数。 Returns: str: 截断后的文本。 """ normalized_text = text.strip() if len(normalized_text) <= max_length: return normalized_text return f"{normalized_text[: max_length - 1]}…" def _build_tool_record_payload( self, invocation: ToolInvocation, result: ToolExecutionResult, tool_spec: Optional[ToolSpec], ) -> dict[str, Any]: """构造统一工具落库数据。 Args: invocation: 工具调用对象。 result: 工具执行结果。 tool_spec: 对应的工具声明。 Returns: dict[str, Any]: 可直接写入数据库的工具记录数据。 """ payload: dict[str, Any] = { "call_id": invocation.call_id, "session_id": invocation.session_id, "stream_id": invocation.stream_id, "arguments": self._normalize_tool_record_value(invocation.arguments), "success": result.success, "content": result.content, "error_message": result.error_message, "history_content": result.get_history_content(), "structured_content": self._normalize_tool_record_value(result.structured_content), "metadata": self._normalize_tool_record_value(result.metadata), } if tool_spec is not None: payload["provider_name"] = tool_spec.provider_name payload["provider_type"] = tool_spec.provider_type payload["brief_description"] = tool_spec.brief_description payload["detailed_description"] = tool_spec.detailed_description payload["title"] = tool_spec.title return payload def _build_tool_display_prompt( self, invocation: ToolInvocation, result: ToolExecutionResult, tool_spec: Optional[ToolSpec], ) -> str: """构造展示给历史回放与 UI 的工具摘要。 Args: invocation: 工具调用对象。 result: 工具执行结果。 tool_spec: 对应的工具声明。 Returns: str: 用于展示的工具摘要文本。 """ custom_display_prompt = result.metadata.get("record_display_prompt") if isinstance(custom_display_prompt, str) and custom_display_prompt.strip(): return custom_display_prompt.strip() structured_content = ( result.structured_content if isinstance(result.structured_content, dict) else {} ) history_content = self._truncate_tool_record_text(result.get_history_content(), max_length=200) normalized_args = self._normalize_tool_record_value(invocation.arguments) if invocation.tool_name == "reply": target_user_name = str(structured_content.get("target_user_name") or "对方").strip() or "对方" reply_text = str(structured_content.get("reply_text") or "").strip() if result.success and reply_text: return f"你对{target_user_name}进行了回复：{reply_text}" target_message_id = str(invocation.arguments.get("msg_id") or "").strip() error_text = self._truncate_tool_record_text(result.error_message or history_content, max_length=120) return f"你尝试回复消息 {target_message_id or 'unknown'}，但失败了：{error_text}" if invocation.tool_name == "send_emoji": description = str(structured_content.get("description") or "").strip() emotion_list = structured_content.get("emotion") if isinstance(emotion_list, list): emotion_text = "、".join(str(item).strip() for item in emotion_list if str(item).strip()) else: emotion_text = "" if result.success and description: if emotion_text: return f"你发送了表情包：{description}（情绪：{emotion_text}）" return f"你发送了表情包：{description}" return f"你尝试发送表情包，但失败了：{self._truncate_tool_record_text(result.error_message or history_content, 120)}" if invocation.tool_name == "wait": wait_seconds = invocation.arguments.get("seconds", 30) return f"你让当前对话先等待 {wait_seconds} 秒。" if invocation.tool_name == "no_reply": return "你暂停了当前对话循环，等待新的外部消息。" if invocation.tool_name == "query_jargon": words = invocation.arguments.get("words", []) if isinstance(words, list): words_text = "、".join(str(item).strip() for item in words if str(item).strip()) else: words_text = "" if words_text: return f"你查询了这些黑话或词条：{words_text}" return "你查询了一次黑话或词条信息。" if invocation.tool_name == "query_person_info": person_name = str(invocation.arguments.get("person_name") or "").strip() if person_name: return f"你查询了人物信息：{person_name}" return "你查询了一次人物信息。" brief_description = "" if tool_spec is not None: brief_description = tool_spec.brief_description.strip() if normalized_args: arguments_text = self._truncate_tool_record_text( json.dumps(normalized_args, ensure_ascii=False), max_length=160, ) else: arguments_text = "{}" if result.success: if brief_description: return f"{brief_description} 参数={arguments_text}；结果：{history_content or '执行成功'}" return f"你调用了工具 {invocation.tool_name}，参数={arguments_text}；结果：{history_content or '执行成功'}" error_text = self._truncate_tool_record_text(result.error_message or history_content, max_length=160) return f"你调用了工具 {invocation.tool_name}，参数={arguments_text}；执行失败：{error_text}" async def _store_tool_execution_record( self, invocation: ToolInvocation, result: ToolExecutionResult, tool_spec: Optional[ToolSpec], ) -> None: """将工具执行结果落库到统一工具记录表。 Args: invocation: 工具调用对象。 result: 工具执行结果。 tool_spec: 对应的工具声明。 """ if self._runtime.chat_stream is None: logger.debug( f"{self._runtime.log_prefix} 当前没有 chat_stream，跳过工具记录存储: " f"工具={invocation.tool_name}" ) return builtin_prompt = "" if tool_spec is not None: builtin_prompt = tool_spec.build_llm_description() try: await database_api.store_tool_info( chat_stream=self._runtime.chat_stream, builtin_prompt=builtin_prompt, display_prompt=self._build_tool_display_prompt(invocation, result, tool_spec), tool_id=invocation.call_id, tool_data=self._build_tool_record_payload(invocation, result, tool_spec), tool_name=invocation.tool_name, tool_reasoning=invocation.reasoning, ) except Exception: logger.exception( f"{self._runtime.log_prefix} 写入工具记录失败: 工具={invocation.tool_name} 调用编号={invocation.call_id}" ) def _append_tool_execution_result(self, tool_call: ToolCall, result: ToolExecutionResult) -> None: """将统一工具执行结果写回 Maisaka 历史。 Args: tool_call: 原始工具调用对象。 result: 统一工具执行结果。 """ history_content = result.get_history_content() if not history_content: history_content = "工具执行成功。" if result.success else f"工具 {tool_call.func_name} 执行失败。" self._runtime._chat_history.append( ToolResultMessage( content=history_content, timestamp=datetime.now(), tool_call_id=tool_call.call_id, tool_name=tool_call.func_name, success=result.success, ) ) async def _handle_tool_calls( self, tool_calls: list[ToolCall], latest_thought: str, anchor_message: SessionMessage, ) -> bool: """执行一批统一工具调用。 Args: tool_calls: 模型返回的工具调用列表。 latest_thought: 当前轮的最新思考文本。 anchor_message: 当前轮的锚点消息。 Returns: bool: 是否需要暂停当前思考循环。 """ if self._runtime._tool_registry is None: for tool_call in tool_calls: invocation = self._build_tool_invocation(tool_call, latest_thought) result = ToolExecutionResult( tool_name=tool_call.func_name, success=False, error_message="统一工具注册表尚未初始化。", ) await self._store_tool_execution_record(invocation, result, None) self._append_tool_execution_result(tool_call, result) return False execution_context = self._build_tool_execution_context(latest_thought, anchor_message) tool_spec_map = { tool_spec.name: tool_spec for tool_spec in await self._runtime._tool_registry.list_tools() } for tool_call in tool_calls: invocation = self._build_tool_invocation(tool_call, latest_thought) result = await self._runtime._tool_registry.invoke(invocation, execution_context) await self._store_tool_execution_record( invocation, result, tool_spec_map.get(invocation.tool_name), ) self._append_tool_execution_result(tool_call, result) if not result.success and tool_call.func_name == "reply": logger.warning(f"{self._runtime.log_prefix} 回复工具未生成可见消息，将继续下一轮循环") if bool(result.metadata.get("pause_execution", False)): return True return False