@axiom-lattice/core 1.0.20 → 1.0.21

package/src/createPlanExecuteAgent.ts

@@ -1,475 +0,0 @@
-import {
-  Annotation,
-  BaseCheckpointSaver,
-  END,
-  GraphInterrupt,
-  LangGraphRunnableConfig,
-  START,
-  StateGraph,
-  messagesStateReducer,
-} from "@langchain/langgraph";
-import {
-  BaseMessage,
-  HumanMessage,
-  mergeMessageRuns,
-} from "@langchain/core/messages";
-import { z } from "zod";
-import { ModelLattice } from "./model_lattice/ModelLattice";
-import { Logger } from "@logger/Logger";
-import { StateBaseType } from "./types";
-import { returnToolResponse } from "./util/returnToolResponse";
-import { returnAIResponse } from "./util/returnAIResponse";
-import { getLastHumanMessageData } from "./util/getLastHumanMessageData";
-import { tool } from "@langchain/core/tools";
-import { MemoryManager } from "./util/PGMemory";
-import { getModelLattice } from "./model_lattice/ModelLatticeManager";
-import { createReactAgent } from "@langchain/langgraph/prebuilt";
-
-/**
- * Plan-and-Execute Agent 状态定义
- */
-export const PlanExecuteState = Annotation.Root({
-  // 输入
-  input: Annotation<string>({
-    reducer: (x, y) => y ?? x ?? "",
-  }),
-  // 计划步骤列表
-  plan: Annotation<string[]>({
-    reducer: (x, y) => y ?? x ?? [],
-  }),
-  // 已执行的步骤 [步骤名称, 执行结果]
-  pastSteps: Annotation<[string, string][]>({
-    reducer: (x, y) => x.concat(y),
-    default: () => [],
-  }),
-  // 最终响应
-  response: Annotation<string | undefined>({
-    reducer: (x, y) => y ?? x,
-  }),
-  // 错误信息
-  error: Annotation<string | undefined>({
-    reducer: (x, y) => y ?? x,
-  }),
-  // 继承基础状态
-  "x-tenant-id": Annotation<number>(),
-  messages: Annotation<BaseMessage[]>({
-    reducer: messagesStateReducer,
-    default: () => [],
-  }),
-});
-
-/**
- * 计划工具的结构化输出模式
- */
-const planSchema = z.object({
-  steps: z.array(z.string()).describe("需要执行的步骤列表，应按照执行顺序排序"),
-});
-
-/**
- * 响应工具的结构化输出模式
- */
-const responseSchema = z.object({
-  response: z
-    .string()
-    .describe("给用户的最终响应，如果不需要执行步骤，则返回这个字段"),
-});
-
-/**
- * Plan-and-Execute Agent 配置接口
- */
-export interface PlanExecuteAgentConfig {
-  name: string;
-  description: string;
-  tools?: any[]; // 执行器可使用的工具
-  checkpointer?: false | BaseCheckpointSaver<number>;
-  llmManager?: ModelLattice;
-  logger?: Logger;
-  maxSteps?: number; // 最大步骤数限制
-  prompt?: string;
-}
-
-/**
- * 创建 Plan-and-Execute Agent
- *
- * 这是一个规划-执行风格的代理，首先制定多步骤计划，然后逐一执行。
- * 在完成每个任务后，会重新评估计划并根据需要进行修改。
- *
- * 相比传统的ReAct风格代理的优势：
- * 1. 明确的长期规划能力
- * 2. 可以在执行步骤使用较小的模型，只在规划步骤使用大模型
- * 3. 更好的任务分解和执行追踪
- */
-export function createPlanExecuteAgent(config: PlanExecuteAgentConfig) {
-  const {
-    name,
-    description,
-    tools: executorTools = [],
-    checkpointer,
-    llmManager,
-    logger,
-    maxSteps = 10,
-    prompt,
-  } = config;
-
-  const agentLogger =
-    logger || new Logger({ name: `PlanExecuteAgent:${name}` });
-  const llm = llmManager || getModelLattice("default")?.client;
-
-  if (!llm) {
-    throw new Error("LLM not found");
-  }
-
-  agentLogger.info(`初始化Plan-Execute代理: ${name}`, {
-    description,
-    maxSteps,
-    toolsCount: executorTools.length,
-  });
-
-  const agentExecutor = createReactAgent({
-    tools: executorTools,
-    llm: llm,
-  });
-
-  /**
-   * 规划步骤：创建初始计划
-   */
-  async function planStep(
-    state: typeof PlanExecuteState.State,
-    config: LangGraphRunnableConfig
-  ): Promise<Partial<typeof PlanExecuteState.State>> {
-    const input =
-      state.input || getLastHumanMessageData(state.messages)?.content;
-
-    const plannerPrompt = `${
-      prompt || ""
-    }\n针对给定的目标，制定一个简单的分步计划。
-这个计划应该包含独立的任务，如果正确执行将会得到正确的答案。不要添加多余的步骤。
-最后一步的结果应该是最终答案。确保每一步都有所需的所有信息 - 不要跳过步骤。
-
-可以使用的工具如下：
-${executorTools.map((tool) => tool.name).join("\n")}
-
-目标: ${input}`;
-
-    try {
-      const planResult = (await llm.invokeWithStructuredOutput(
-        [new HumanMessage(plannerPrompt)],
-        planSchema
-      )) as { steps: string[] };
-
-      const { messages } = await returnToolResponse(config, {
-        title: "我的计划",
-        content: planResult.steps.map((step: string) => step).join("\n\n"),
-        status: "success",
-      });
-
-      agentLogger.info("计划制定完成", { plan: planResult.steps });
-
-      return {
-        input: input as string,
-        plan: planResult.steps,
-        error: undefined,
-        messages,
-      };
-    } catch (error) {
-      const errorMsg = `计划制定失败: ${
-        error instanceof Error ? error.message : "未知错误"
-      }`;
-      agentLogger.error(errorMsg, { error });
-      return {
-        error: errorMsg,
-      };
-    }
-  }
-
-  /**
-   * 执行步骤：执行计划中的第一个任务
-   */
-  async function executeStep(
-    state: typeof PlanExecuteState.State,
-    config: LangGraphRunnableConfig
-  ): Promise<Partial<typeof PlanExecuteState.State>> {
-    if (!state.plan || state.plan.length === 0) {
-      agentLogger.warn("没有可执行的任务");
-      return { error: "没有可执行的任务" };
-    }
-
-    const currentTask = state.plan[0];
-    agentLogger.info("开始执行任务", { task: currentTask });
-
-    try {
-      // 这里可以集成真实的工具执行器
-      // 目前使用LLM来模拟执行
-      const executorPrompt = `执行以下任务并提供结果：
-
-任务: ${currentTask}
-
-上下文信息:
-- 原始目标: ${state.input}
-- 已完成的步骤: ${state.pastSteps
-        .map(([step, result]) => `${step}: ${result}`)
-        .join("\n")}
-
-请提供执行此任务的详细结果。`;
-
-      const executionResult = await agentExecutor.invoke({
-        messages: [new HumanMessage(executorPrompt)],
-      });
-
-      const resultContent =
-        executionResult.messages[executionResult.messages.length - 1].content;
-
-      // 生成工具响应消息
-      // const { messages } = await returnToolResponse(config, {
-      //   title: `执行任务: ${currentTask}`,
-      //   content: resultContent as string,
-      //   status: "success",
-      // });
-
-      return {
-        pastSteps: [[currentTask, resultContent]] as [string, string][],
-        plan: state.plan.slice(1), // 移除已执行的任务
-        //messages,
-        error: undefined,
-      };
-    } catch (error) {
-      if (error instanceof GraphInterrupt) {
-        throw error;
-      }
-      const errorMsg = `任务执行失败: ${
-        error instanceof Error ? error.message : "未知错误"
-      }`;
-      agentLogger.error(errorMsg, { task: currentTask, error });
-
-      // const { messages } = await returnToolResponse(config, {
-      //   title: `执行任务失败: ${currentTask}`,
-      //   content: errorMsg,
-      //   status: "error",
-      // });
-
-      return {
-        error: errorMsg,
-        //messages,
-      };
-    }
-  }
-
-  const responseTool = tool(
-    ({ response }: { response: string }) => {
-      return response;
-    },
-    {
-      name: "response",
-      description: "Respond to the user.",
-      schema: responseSchema,
-      returnDirect: true,
-    }
-  );
-
-  const planTool = tool(
-    ({ steps }: { steps: string[] }) => {
-      return steps.join("\n\n");
-    },
-    {
-      name: "plan",
-      description: "This tool is used to plan the steps to follow.",
-      schema: planSchema,
-      returnDirect: true,
-    }
-  );
-
-  const replanAgent = createReactAgent({
-    tools: [responseTool, planTool],
-    llm: llm,
-  });
-  /**
-   * 重新规划步骤：根据执行结果调整计划
-   */
-  async function replanStep(
-    state: typeof PlanExecuteState.State,
-    config: LangGraphRunnableConfig
-  ): Promise<Partial<typeof PlanExecuteState.State>> {
-    agentLogger.info("开始重新规划");
-
-    const replannerPrompt = `${
-      prompt || ""
-    }\n针对给定的目标，根据已执行的步骤来更新计划。
-这个计划应该包含独立的任务，如果正确执行将会得到正确的答案。不要添加多余的步骤。
-最后一步的结果应该是最终答案。确保每一步都有所需的所有信息 - 不要跳过步骤。
-
-你的目标是：
-${state.input}
-
-你的原始计划是：
-${state.plan.join("\n")}
-
-你目前已经完成了以下步骤：
-${state.pastSteps.map(([step, result]) => `${step}: ${result}`).join("\n")}
-
-Update your plan accordingly. If no more steps are needed and you can return to the user, then respond with that and use the 'response' function.
-Otherwise, fill out the plan.
-Only add steps to the plan that still NEED to be done. Do not return previously done steps as part of the plan.`;
-
-    try {
-      // 首先尝试获取最终响应
-      const responseResult = await replanAgent.invoke({
-        messages: [new HumanMessage(replannerPrompt)],
-      });
-
-      const toolCall =
-        responseResult.messages[responseResult.messages.length - 1];
-
-      if (toolCall?.name == "response") {
-        agentLogger.info("任务完成，提供最终响应");
-
-        const { messages } = await returnAIResponse(
-          config,
-          toolCall.content as string
-        );
-
-        return {
-          response: toolCall.content as string,
-          messages,
-          error: undefined,
-        };
-      } else if (toolCall?.name == "plan") {
-        let messages: any[] = [];
-        if (toolCall.content?.length > 0) {
-          messages = await returnToolResponse(config, {
-            title: "新的计划",
-            content: toolCall.content as string,
-            status: "success",
-          }).messages;
-        }
-        agentLogger.info("计划更新完成", { newPlan: toolCall.content });
-        return {
-          messages,
-          plan: (toolCall.content as string).split("\n\n"),
-          error: undefined,
-        };
-      } else {
-        return {
-          error: "重新规划失败",
-        };
-      }
-    } catch (error) {
-      if (error instanceof GraphInterrupt) {
-        throw error;
-      }
-      const errorMsg = `重新规划失败: ${
-        error instanceof Error ? error.message : "未知错误"
-      }`;
-      agentLogger.error(errorMsg, { error });
-      return {
-        error: errorMsg,
-      };
-    }
-  }
-
-  /**
-   * 判断是否应该结束执行
-   */
-  function shouldEnd(state: typeof PlanExecuteState.State): "end" | "continue" {
-    if (state.error) {
-      agentLogger.warn("因错误结束执行", { error: state.error });
-      return "end";
-    }
-
-    if (state.response) {
-      agentLogger.info("任务完成，结束执行");
-      return "end";
-    }
-
-    if (state.pastSteps.length >= maxSteps) {
-      agentLogger.warn("达到最大步骤数限制，结束执行", {
-        maxSteps,
-        currentSteps: state.pastSteps.length,
-      });
-      return "end";
-    }
-
-    if (!state.plan || state.plan.length === 0) {
-      agentLogger.info("计划为空，结束执行");
-      return "end";
-    }
-
-    return "continue";
-  }
-
-  // 构建状态图
-  const workflow = new StateGraph(PlanExecuteState)
-    .addNode("planner", planStep)
-    .addNode("executor", executeStep)
-    .addNode("replanner", replanStep)
-    .addEdge(START, "planner")
-    .addEdge("planner", "executor")
-    .addEdge("executor", "replanner")
-    .addConditionalEdges("replanner", shouldEnd, {
-      end: END,
-      continue: "executor",
-    });
-
-  // 编译图
-  //const memory = MemoryManager.getInstance();
-  const compiledGraph = workflow.compile({
-    checkpointer: checkpointer || MemoryManager.getInstance(),
-    name: `PlanExecuteAgent:${name}`,
-  });
-
-  agentLogger.info(`Plan-Execute代理编译完成: ${name}`);
-
-  return {
-    /**
-     * 执行代理
-     */
-    invoke: async (
-      input: { input: string; "x-tenant-id": number },
-      config?: LangGraphRunnableConfig
-    ) => {
-      agentLogger.info(`开始执行Plan-Execute代理: ${name}`, {
-        input: input.input,
-      });
-
-      try {
-        const result = await compiledGraph.invoke(input, config);
-        agentLogger.info(`代理执行完成: ${name}`);
-        return result;
-      } catch (error) {
-        agentLogger.error(
-          `代理执行失败: ${name}`,
-          error instanceof Error ? error : { error }
-        );
-        throw error;
-      }
-    },
-
-    /**
-     * 流式执行代理
-     */
-    stream: async (
-      input: { input: string; "x-tenant-id": number },
-      config?: LangGraphRunnableConfig
-    ) => {
-      agentLogger.info(`开始流式执行Plan-Execute代理: ${name}`, {
-        input: input.input,
-      });
-      return compiledGraph.stream(input, config);
-    },
-
-    /**
-     * 获取代理信息
-     */
-    getInfo: () => ({
-      name,
-      description,
-      type: "PlanExecuteAgent",
-      toolsCount: executorTools.length,
-      maxSteps,
-    }),
-
-    /**
-     * 获取编译后的图
-     */
-    getCompiledGraph: () => compiledGraph,
-  };
-}

package/src/deep_agent/graph.ts

@@ -1,106 +0,0 @@
-/**
- * Main createDeepAgent function for Deep Agents
- *
- * Main entry point for creating deep agents with TypeScript types for all parameters:
- * tools, instructions, model, subagents, and stateSchema. Combines built-in tools with
- * provided tools, creates task tool using createTaskTool(), and returns createReactAgent
- * with proper configuration. Ensures exact parameter matching and behavior with Python version.
- */
-
-// import "@langchain/anthropic/zod";
-import { createTaskTool } from "./subAgent";
-import { writeTodos, readFile, writeFile, editFile, ls } from "./tools";
-import type { CreateDeepAgentParams } from "./types";
-import type { StructuredTool } from "@langchain/core/tools";
-import { z } from "zod";
-import { DeepAgentState } from "./state";
-import { getModelLattice } from "../model_lattice/ModelLatticeManager";
-import { createReactAgent } from "@langchain/langgraph/prebuilt";
-import { getCheckpointSaver } from "@memory_lattice/MemoryLatticeManager";
-
-/**
- * Base prompt that provides instructions about available tools
- * Ported from Python implementation to ensure consistent behavior
- */
-const BASE_PROMPT = `You have access to a number of standard tools
-
-## \`write_todos\`
-
-You have access to the \`write_todos\` tools to help you manage and plan tasks. Use these tools VERY frequently to ensure that you are tracking your tasks and giving the user visibility into your progress.
-These tools are also EXTREMELY helpful for planning tasks, and for breaking down larger complex tasks into smaller steps. If you do not use this tool when planning, you may forget to do important tasks - and that is unacceptable.
-
-It is critical that you mark todos as completed as soon as you are done with a task. Do not batch up multiple tasks before marking them as completed.
-## \`task\`
-
-- When doing web search, prefer to use the \`task\` tool in order to reduce context usage.`;
-
-/**
- * Built-in tools that are always available in Deep Agents
- */
-const BUILTIN_TOOLS: StructuredTool[] = [
-  writeTodos,
-  readFile,
-  writeFile,
-  editFile,
-  ls,
-];
-
-/**
- * Create a Deep Agent with TypeScript types for all parameters.
- * Combines built-in tools with provided tools, creates task tool using createTaskTool(),
- * and returns createReactAgent with proper configuration.
- * Ensures exact parameter matching and behavior with Python version.
- */
-export function createDeepAgent<
-  StateSchema extends z.ZodObject<any, any, any, any, any>
->(params: CreateDeepAgentParams<StateSchema> = {}) {
-  const {
-    tools = [],
-    instructions,
-    model = getModelLattice("default")?.client,
-    subagents = [],
-  } = params;
-
-  if (!model) {
-    throw new Error("Model not found");
-  }
-
-  const stateSchema = params.stateSchema
-    ? DeepAgentState.extend(params.stateSchema.shape)
-    : DeepAgentState;
-
-  // Combine built-in tools with provided tools
-  const allTools: StructuredTool[] = [...BUILTIN_TOOLS, ...tools];
-  // Create task tool using createTaskTool() if subagents are provided
-  if (subagents.length > 0) {
-    // Create tools map for task tool creation
-    const toolsMap: Record<string, StructuredTool> = {};
-    for (const tool of allTools) {
-      if (tool.name) {
-        toolsMap[tool.name] = tool;
-      }
-    }
-
-    const taskTool = createTaskTool({
-      subagents,
-      tools: toolsMap,
-      model,
-      stateSchema,
-    });
-    allTools.push(taskTool);
-  }
-
-  // Combine instructions with base prompt like Python implementation
-  const finalInstructions = instructions
-    ? instructions + BASE_PROMPT
-    : BASE_PROMPT;
-
-  // Return createReactAgent with proper configuration
-  return createReactAgent({
-    llm: model,
-    tools: allTools,
-    stateSchema: stateSchema,
-    prompt: finalInstructions,
-    checkpointer: getCheckpointSaver("default"),
-  });
-}

package/src/deep_agent/index.ts

@@ -1,25 +0,0 @@
-/**
- * Deep Agents TypeScript Implementation
- *
- * A TypeScript port of the Python Deep Agents library for building controllable AI agents with LangGraph.
- * This implementation maintains 1:1 compatibility with the Python version.
- */
-
-export { createDeepAgent } from "./graph";
-export { writeTodos, readFile, writeFile, editFile, ls } from "./tools";
-export { DeepAgentState, fileReducer } from "./state";
-export {
-  WRITE_TODOS_DESCRIPTION,
-  TASK_DESCRIPTION_PREFIX,
-  TASK_DESCRIPTION_SUFFIX,
-  EDIT_DESCRIPTION,
-  TOOL_DESCRIPTION,
-} from "./prompts";
-export type {
-  SubAgent,
-  Todo,
-  DeepAgentStateType,
-  CreateDeepAgentParams,
-  CreateTaskToolParams,
-  TodoStatus,
-} from "./types";