"""MoviePilot 自定义工具筛选中间件。""" from dataclasses import replace import json from collections.abc import Awaitable, Callable from typing import Annotated, Any, NotRequired from langchain.agents.middleware.types import ( AgentState, ContextT, ModelRequest, ModelResponse, ResponseT, ) from langchain.agents.middleware.types import ( PrivateStateAttr, # noqa ) from langchain.agents.middleware.tool_selection import ( DEFAULT_SYSTEM_PROMPT, LLMToolSelectorMiddleware, ) from langchain_core.language_models.chat_models import BaseChatModel from langchain_core.messages import AIMessage, HumanMessage from langchain_core.runnables import RunnableConfig from langchain_core.tools import BaseTool from langgraph.runtime import Runtime from typing_extensions import TypedDict # noqa from app.agent.llm import LLMHelper from app.agent.tools.tags import ToolTag from app.log import logger MIN_SELECTED_TOOL_COUNT = 4 RECENT_SELECTION_CONTEXT_MESSAGE_LIMIT = 6 RECENT_SELECTION_CONTEXT_MAX_CHARS = 6000 RECENT_SELECTION_CONTEXT_TRUNCATION_PREFIX = "..." TOOL_GROUP_EXCLUDED_TAGS = frozenset( { ToolTag.AgentTool.value, ToolTag.Read.value, ToolTag.Write.value, ToolTag.Admin.value, ToolTag.Message.value, ToolTag.UserInteraction.value, ToolTag.TerminalResponse.value, } ) MOVIEPILOT_TOOL_SELECTION_HINT = """ MoviePilot tool-chain hints: - Tools with the same capability tag belong to the same functional group. - For multi-step MoviePilot tasks, keep same-tag tools together when relevant. - Prefer selecting likely next-step tools in the same capability group instead of selecting only the first tool. """ class ToolSelectionState(AgentState): """工具筛选中间件私有状态。""" selected_tool_names: NotRequired[Annotated[list[str] | None, PrivateStateAttr]] """当前这条用户请求首轮筛选得到的工具名列表。""" class ToolSelectionStateUpdate(TypedDict): """工具筛选中间件状态更新项。""" selected_tool_names: list[str] | None class ToolSelectorMiddleware(LLMToolSelectorMiddleware): """ 为 DeepSeek 兼容端点提供更稳妥的工具筛选实现。 LangChain 默认会通过 `with_structured_output()` 走 OpenAI 的 `response_format=json_schema` 路径，但 DeepSeek 官方 OpenAI 兼容端点公开文档 仅保证 `json_object` 模式可用。对于 `deepseek-reasoner`，这会在工具筛选阶段 提前触发 400，导致 Agent 还没真正开始执行工具就失败。 因此这里仅在识别到 DeepSeek 模型/端点时，退回到显式 JSON 输出模式： 1. 使用 `response_format={"type": "json_object"}`； 2. 在提示词中明确约束返回 JSON 结构； 3. 手动解析 `{"tools": [...]}`，其余模型继续沿用 LangChain 默认实现。 另外，LangChain 原生工具筛选挂在 `wrap_model_call` 上，会在同一条用户请求 的每次“模型回合”前都重新筛选一次工具。对于会多轮调用工具的复杂任务， 这会重复消耗一次额外的 LLM 调用。这里改成： - `abefore_agent()`：在本轮 Agent 执行开始时筛选一次； - `awrap_model_call()`：从 `request.state` 读取首轮筛选结果并复用。 """ state_schema = ToolSelectionState def __init__( self, model: BaseChatModel | str | None = None, system_prompt: str = DEFAULT_SYSTEM_PROMPT, selection_tools: list[Any] | None = None, max_tools: int | None = None, always_include: list[str] | None = None, ) -> None: super().__init__( model=model, system_prompt=self._append_tool_selection_hint(system_prompt), max_tools=max_tools, always_include=always_include, ) self.selection_tools = selection_tools or [] @classmethod def _render_recent_conversation_context( cls, messages: list[Any], ) -> tuple[str, int]: """渲染最近对话上下文，供工具筛选模型理解多轮追问。""" rendered_messages = [] for message in messages: if isinstance(message, HumanMessage): role = "User" elif isinstance(message, AIMessage): role = "Assistant" else: continue content = LLMHelper.extract_text_content(message.content).strip() if not content: continue rendered_messages.append(f"{role}: {content}") recent_messages = rendered_messages[-RECENT_SELECTION_CONTEXT_MESSAGE_LIMIT:] context = "\n\n".join(recent_messages) if len(context) > RECENT_SELECTION_CONTEXT_MAX_CHARS: context = ( f"{RECENT_SELECTION_CONTEXT_TRUNCATION_PREFIX}" f"{context[-RECENT_SELECTION_CONTEXT_MAX_CHARS:]}" ) return context, len(recent_messages) @classmethod def _build_contextual_user_message( cls, messages: list[Any], last_user_message: HumanMessage, ) -> HumanMessage: """根据最近对话构造工具筛选专用用户消息。""" context, message_count = cls._render_recent_conversation_context(messages) if message_count <= 1: return last_user_message return HumanMessage( content=( "Recent conversation context for tool selection:\n" f"{context}\n\n" "Select tools for the latest user instruction. Use prior assistant " "messages and earlier user requests when the latest user message " "depends on previous context." ) ) def _prepare_selection_request( self, request: ModelRequest[ContextT], ) -> Any | None: """准备带最近对话上下文的工具筛选请求。""" selection_request = super()._prepare_selection_request(request) if selection_request is None: return None contextual_user_message = self._build_contextual_user_message( messages=request.messages, last_user_message=selection_request.last_user_message, ) if contextual_user_message is selection_request.last_user_message: return selection_request return replace(selection_request, last_user_message=contextual_user_message) @staticmethod def _append_tool_selection_hint(system_prompt: str) -> str: """追加 MoviePilot 工具组选择提示，避免复杂链路只选中首个工具。""" if "MoviePilot tool-chain hints:" in system_prompt: return system_prompt return f"{system_prompt.rstrip()}{MOVIEPILOT_TOOL_SELECTION_HINT}" def _get_tool_selection_limit(self, valid_tool_names: list[str]) -> int: """计算补齐筛选结果时允许使用的工具数量上限。""" if self.max_tools: return min(self.max_tools, len(valid_tool_names)) return len(valid_tool_names) @staticmethod def _normalize_tool_tags(tool: BaseTool) -> list[str]: """读取工具的业务标签，过滤掉无法表达工具组的通用标签。""" tags = getattr(tool, "tags", None) or [] if isinstance(tags, str): tags = [tags] normalized_tags = [] for tag in tags: tag_value = getattr(tag, "value", tag) if not tag_value: continue tag_name = str(tag_value) if tag_name in TOOL_GROUP_EXCLUDED_TAGS or tag_name in normalized_tags: continue normalized_tags.append(tag_name) return normalized_tags @classmethod def _build_tool_groups( cls, available_tools: list[BaseTool], valid_tool_names: list[str], ) -> list[tuple[str, list[str]]]: """根据工具标签构造能力组，保留当前工具列表中的稳定顺序。""" valid_tool_set = set(valid_tool_names) tool_groups: dict[str, list[str]] = {} for tool in available_tools: tool_name = getattr(tool, "name", None) if not tool_name or tool_name not in valid_tool_set: continue for tag in cls._normalize_tool_tags(tool): group_tool_names = tool_groups.setdefault(tag, []) if tool_name not in group_tool_names: group_tool_names.append(tool_name) return [ (tag, tool_names) for tag, tool_names in tool_groups.items() if len(tool_names) > 1 ] @classmethod def _get_matched_tool_groups( cls, selected_names: list[str], available_tools: list[BaseTool], valid_tool_names: list[str], ) -> list[tuple[str, list[str]]]: """返回已选工具命中的标签能力组。""" groups_by_tag = { tag: tool_names for tag, tool_names in cls._build_tool_groups( available_tools=available_tools, valid_tool_names=valid_tool_names, ) } tools_by_name = { tool.name: tool for tool in available_tools if getattr(tool, "name", None) } matched_groups: list[tuple[str, list[str]]] = [] seen_tags = set() for tool_name in selected_names: tool = tools_by_name.get(tool_name) if not tool: continue for tag in cls._normalize_tool_tags(tool): if tag in seen_tags or tag not in groups_by_tag: continue matched_groups.append((tag, groups_by_tag[tag])) seen_tags.add(tag) return matched_groups def _complete_low_count_selection( self, selected_tool_names: list[str], valid_tool_names: list[str], available_tools: list[BaseTool], ) -> list[str]: """ 当模型只选出极少工具时，按工具标签补齐同组工具。 工具标签是工具自身声明的能力归属。这里只补齐已经命中的标签组， 不会把所有工具组都展开。 """ limit = self._get_tool_selection_limit(valid_tool_names) selected_names = [ tool_name for tool_name in selected_tool_names if tool_name in valid_tool_names ] selected_set = set(selected_names) valid_tool_set = set(valid_tool_names) completed_names = list(selected_names) matched_groups = self._get_matched_tool_groups( selected_names=selected_names, available_tools=available_tools, valid_tool_names=valid_tool_names, ) if not matched_groups: return completed_names[:limit] matched_group_tool_names = { tool_name for _, group_tool_names in matched_groups for tool_name in group_tool_names } target_count = min( max(MIN_SELECTED_TOOL_COUNT, len(matched_group_tool_names)), limit, ) if len(selected_names) >= target_count: return selected_names[:limit] for _, group_tool_names in matched_groups: for tool_name in group_tool_names: if tool_name in selected_set or tool_name not in valid_tool_set: continue completed_names.append(tool_name) selected_set.add(tool_name) if len(completed_names) >= target_count: return completed_names[:limit] return completed_names[:limit] def _process_selection_response( self, response: dict[str, Any], available_tools: list[BaseTool], valid_tool_names: list[str], request: ModelRequest[ContextT], ) -> ModelRequest[ContextT]: """ 处理工具筛选响应，并保留空结果回退所有工具的 MoviePilot 策略。 """ if response.get("tools") == []: logger.warning("工具筛选结果为空，将恢复使用所有工具。") always_included_tools: list[BaseTool] = [ tool for tool in request.tools if not isinstance(tool, dict) and tool.name in self.always_include ] provider_tools = [tool for tool in request.tools if isinstance(tool, dict)] return request.override( tools=[*available_tools, *always_included_tools, *provider_tools] ) response["tools"] = self._complete_low_count_selection( selected_tool_names=[ tool_name for tool_name in response.get("tools", []) if isinstance(tool_name, str) ], valid_tool_names=valid_tool_names, available_tools=available_tools, ) return super()._process_selection_response( response, available_tools, valid_tool_names, request, ) @staticmethod def _is_deepseek_compatible_model(model: BaseChatModel) -> bool: """ 判断当前模型是否应当走 DeepSeek JSON 兼容分支。 除了官方 `langchain_deepseek`，用户也可能通过 OpenAI-compatible 配置把 DeepSeek 端点接到 `ChatOpenAI`。因此这里同时检查模块名、模型名 和 Base URL，避免只靠单一条件漏判。 """ module_name = type(model).__module__.lower() model_name = ( str(getattr(model, "model_name", "") or getattr(model, "model", "")) .strip() .lower() ) base_url = ( str(getattr(model, "openai_api_base", "") or getattr(model, "api_base", "")) .strip() .lower() ) return ( "deepseek" in module_name or model_name.startswith("deepseek-") or "api.deepseek.com" in base_url ) @staticmethod def _parse_json_object(text: str) -> dict[str, Any]: """ 解析模型返回的 JSON。 DeepSeek 在 JSON 模式下通常会返回纯 JSON，但这里仍做一层兜底， 兼容模型偶发输出围栏或前后说明文本的情况。 """ stripped_text = text.strip() if not stripped_text: raise ValueError("工具筛选返回了空响应") try: payload = json.loads(stripped_text) if isinstance(payload, dict): return payload except json.JSONDecodeError: pass start = stripped_text.find("{") end = stripped_text.rfind("}") if start == -1 or end == -1 or end <= start: raise ValueError(f"工具筛选返回的内容不是合法 JSON: {stripped_text}") payload = json.loads(stripped_text[start: end + 1]) if not isinstance(payload, dict): raise ValueError("工具筛选 JSON 顶层必须是对象") return payload @classmethod def _render_tool_list(cls, available_tools: list[Any]) -> str: """把工具名和描述渲染成稳定的文本列表。""" lines = [] for tool in available_tools: tags = cls._normalize_tool_tags(tool) tag_text = f" [group tags: {', '.join(tags)}]" if tags else "" lines.append(f"- {tool.name}{tag_text}: {tool.description}") return "\n".join(lines) @classmethod def _render_tool_groups(cls, available_tools: list[BaseTool]) -> str: """把当前可用工具按标签渲染成能力组提示。""" valid_tool_names = [ tool.name for tool in available_tools if getattr(tool, "name", None) ] groups = cls._build_tool_groups( available_tools=available_tools, valid_tool_names=valid_tool_names, ) if not groups: return "" rendered_groups = "\n".join( f"- {tag}: {', '.join(tool_names)}" for tag, tool_names in groups ) return f"Capability groups from tool tags:\n{rendered_groups}\n\n" def _build_deepseek_selection_prompt(self, selection_request: Any) -> str: """ 为 DeepSeek 生成显式 JSON 输出提示。 DeepSeek 官方文档要求在 JSON 输出模式下，提示词中必须明确包含 JSON 约束，否则兼容端点可能返回空内容或无意义输出。 """ limit_instruction = "" if self.max_tools: limit_instruction = f"- Select up to {self.max_tools} tools. IF NO TOOLS ARE RELEVANT, DO NOT RETURN AN EMPTY ARRAY. SELECT THE MOST APPLICABLE ONES TO ENSURE THE REQUEST IS HANDLED." return ( f"{selection_request.system_message}\n\n" "Return the answer in JSON only.\n" 'Use exactly this shape: {"tools": ["tool_name_1", "tool_name_2"]}\n' "Rules:\n" "- The `tools` field must be a JSON array of strings.\n" "- Only use tool names from the allowed list below.\n" "- Order tools by relevance, with the most relevant first.\n" "- Tools sharing the same capability tag are in the same group; include same-group tools together when relevant.\n" f"{limit_instruction}\n" "- Do not add explanations, markdown, or extra keys.\n\n" f"{self._render_tool_groups(selection_request.available_tools)}" "Allowed tools:\n" f"{self._render_tool_list(selection_request.available_tools)}" ) def _normalize_selection_response(self, response: Any) -> dict[str, list[str]]: """ 解析并标准化 DeepSeek JSON 模式的工具筛选结果。 """ content = getattr(response, "content", response) text = LLMHelper.extract_text_content(content) logger.debug(f"工具筛选原始响应: {text}") payload = self._parse_json_object(text) tools = payload.get("tools") if not isinstance(tools, list): raise ValueError(f"工具筛选 JSON 缺少 `tools` 数组: {payload}") normalized_tools = [ tool_name for tool_name in tools if isinstance(tool_name, str) ] logger.debug(f"工具筛选标准化结果: {normalized_tools}") return {"tools": normalized_tools} async def _aselect_tools_with_deepseek( self, selection_request: Any ) -> dict[str, list[str]]: """ 使用 DeepSeek 兼容的 JSON 输出模式执行异步工具筛选。 """ logger.debug("工具筛选走 DeepSeek JSON 兼容分支") structured_model = selection_request.model.bind( response_format={"type": "json_object"} ) response = await structured_model.ainvoke( [ { "role": "system", "content": self._build_deepseek_selection_prompt(selection_request), }, selection_request.last_user_message, ] ) return self._normalize_selection_response(response) @staticmethod def _extract_selected_tool_names(request: ModelRequest) -> list[str]: """从已筛选后的请求中提取最终工具名，保留原有顺序。""" return [tool.name for tool in request.tools if not isinstance(tool, dict)] @staticmethod def _apply_selected_tools( request: ModelRequest[ContextT], selected_tool_names: list[str], ) -> ModelRequest[ContextT]: """ 将已筛选出的工具集应用到当前模型请求。 这里只复用首次筛选出的客户端工具名；provider-specific 的 dict 工具仍然 原样保留，避免破坏 LangChain/provider 自身的工具绑定约定。 """ if not selected_tool_names: return request current_tools_by_name = { tool.name: tool for tool in request.tools if not isinstance(tool, dict) } selected_tools = [ current_tools_by_name[tool_name] for tool_name in selected_tool_names if tool_name in current_tools_by_name ] provider_tools = [tool for tool in request.tools if isinstance(tool, dict)] return request.override(tools=[*selected_tools, *provider_tools]) async def _aselect_request_once( self, request: ModelRequest[ContextT] ) -> ModelRequest[ContextT]: """ 执行一次真实工具筛选，并返回筛选后的请求对象。 这里单独抽成 helper，便于首次筛选后缓存结果，也便于测试覆盖 “首轮筛选，后续复用”的行为。 """ selection_request = self._prepare_selection_request(request) if selection_request is None: return request if not self._is_deepseek_compatible_model(selection_request.model): captured_request: ModelRequest[ContextT] = request async def _capture_handler( updated_request: ModelRequest[ContextT], ) -> ModelRequest[ContextT]: nonlocal captured_request captured_request = updated_request return updated_request await super().awrap_model_call(request, _capture_handler) return captured_request response = await self._aselect_tools_with_deepseek(selection_request) return self._process_selection_response( response, selection_request.available_tools, selection_request.valid_tool_names, request, ) async def abefore_agent( # noqa self, state: ToolSelectionState, runtime: Runtime, # noqa config: RunnableConfig, ) -> ToolSelectionStateUpdate | None: # ty: ignore[invalid-method-override] """ 在本轮 Agent 执行开始前完成一次真实工具筛选。 这样后续多轮 `model -> tools -> model` 循环都只复用这一次结果， 不会为每次模型回合重复追加一笔 selector LLM 开销。 """ if "selected_tool_names" in state: return None if not self.selection_tools or self.model is None: return ToolSelectionStateUpdate(selected_tool_names=None) selection_request = ModelRequest( model=self.model, tools=list(self.selection_tools), messages=state["messages"], state=state, runtime=runtime, ) modified_request = await self._aselect_request_once(selection_request) selected_tool_names = self._extract_selected_tool_names(modified_request) return ToolSelectionStateUpdate(selected_tool_names=selected_tool_names or None) async def awrap_model_call( self, request: ModelRequest[ContextT], handler: Callable[ [ModelRequest[ContextT]], Awaitable[ModelResponse[ResponseT]] ], ) -> ModelResponse[ResponseT]: """ 从 state 中读取首次筛选结果，并应用到每次模型回合。 """ selected_tool_names = request.state.get("selected_tool_names") # noqa # 正常路径下，`abefore_agent()` 已经提前写入状态；这里只保留一层兜底， # 兼容直接单测或未来某些绕过 before_agent 的调用场景。 if ( selected_tool_names is None and self.selection_tools and self.model is not None ): request = await self._aselect_request_once(request) selected_tool_names = self._extract_selected_tool_names(request) or None request.state["selected_tool_names"] = selected_tool_names # noqa if selected_tool_names: request = self._apply_selected_tools(request, selected_tool_names) return await handler(request)