@saber2pr/ai-agent 0.0.6 → 0.0.8

package/README.md

@@ -1,74 +1,141 @@
-# @saber2pr/ai-agent
+# 🚀 Saber2pr AI Agent
 
-A lightweight local AI assistant based on the **MCP (Model Context Protocol)**. It automatically loads your local tools (from Cursor or VSCode MCP configurations) and provides an intelligent orchestration layer via OpenAI-compatible APIs.
+A high-performance AI Agent toolkit designed for automated code auditing, repository mapping, and architectural analysis. It supports both a lightweight **Standard Edition** for direct API interaction and a powerful **LangChain Edition** for complex multi-step reasoning and private LLM integration.
 
-## ✨ Features
+## ✨ Core Features
 
-* **Native MCP Support**: Seamlessly connects to local MCP servers using Stdio transport.
-* **Auto-Discovery**: Automatically reads tool definitions from `~/.cursor/mcp.json` and `~/.vscode/mcp.json`.
-* **Persistent Configuration**: On the first run, it guides you through setting up your API endpoint, key, and model name, saving them to `~/.saber2pr-agent.json`.
-* **Namespace Management**: Prevents tool name conflicts by automatically prefixing functions (e.g., `serverName__toolName`).
-* **Interactive CLI**: Built-in REPL for multi-turn conversations and complex tool-chaining.
+* **Dual Mode Support**:
+* **Standard Mode**: Lightweight, fast, and uses direct OpenAI-compatible API calls.
+* **LangChain Mode**: Orchestrated via ReAct agents, supporting complex tool-chains and custom model extensions.
 
-## 📦 Installation
 
-Install globally via npm:
+* **MCP Integration**: Built on the Model Context Protocol to bridge local development environments with AI.
+* **Repository Intelligence**: Integrated `PromptEngine` for generating project maps and code skeletons without exhausting tokens.
+* **Automated Audit Workflow**: Specialized tools for locating code violations, providing line-specific fixes, and generating structured JSON reports.
+* **Private LLM Gateway**: Easily adapt to non-standard API protocols (e.g., Jarvis, internal enterprise gateways) by extending the `BaseChatModel`.
+
+---
+
+## 🛠️ Installation
 
 ```bash
-npm install -g @saber2pr/ai-agent
+# Clone the repository
+git clone https://github.com/saber2pr/ai-agent.git
+cd ai-agent
+
+# Install dependencies
+npm install
+
+# Build the project
+npm run build
+
 ```
 
-Or run directly using `npx`:
+---
+
+## 🚀 Usage Modes
+
+### 1. Standard Edition (Direct API)
+
+Best for quick scripts and simple chat interactions. It uses a straightforward message-loop logic.
+
+```javascript
+const McpAgent = require("./lib/agent").default;
+
+const agent = new McpAgent({
+  targetDir: "/path/to/project"
+});
+
+await agent.chat("Analyze the directory structure.");
 
-```bash
-npx @saber2pr/ai-agent
 ```
 
-## 🚀 Quick Start
+### 2. LangChain Edition (Advanced Agent)
 
-### 1. Launch the Agent
+Best for complex tasks like "Audit the whole project and fix bugs." It supports autonomous tool usage.
 
-Start the assistant by running the binary command:
+```javascript
+const McpAgent = require("./lib/agent-chain").default;
+const { MyPrivateLLM } = require("./your-custom-llm");
+
+const agent = new McpAgent({
+  apiModel: new MyPrivateLLM(), // Inject custom LLM
+  maxIterations: 15,
+  targetDir: "/path/to/project"
+});
+
+await agent.chat("Scan for hardcoded colors and submit a review report.");
 
-```bash
-sagent
 ```
 
-### 2. Initialize Configuration
+---
 
-If it's your first time running the agent, you will be prompted to provide:
+## 🔧 Extending with Private LLMs
 
-* **API Base URL**: e.g., `https://api.openai.com/v1` or your custom proxy.
-* **API Key**: Your model provider's API key.
-* **Model Name**: e.g., `gpt-4o`, `claude-3-5-sonnet`, or `deepseek-v3`.
+To use your own API protocol, extend the `BaseChatModel` from `@langchain/core`:
 
-Your settings will be stored in `~/.saber2pr-agent.json` for future use.
+```javascript
+const { BaseChatModel } = require("@langchain/core/language_models/chat_models");
+
+class MyPrivateLLM extends BaseChatModel {
+  async _generate(messages) {
+    const lastMessage = messages[messages.length - 1];
+    const response = await fetch("https://your-api.com/v1/chat", {
+      method: 'POST',
+      body: JSON.stringify({ query: lastMessage.content }),
+      headers: { 'Authorization': `Bearer ${token}` }
+    });
+    const data = await response.json();
+    return {
+      generations: [{ text: data.text, message: { content: data.text, role: "assistant" } }]
+    };
+  }
+  _llmType() { return "private_llm"; }
+}
+
+```
 
-### 3. Connect Local Tools
+---
 
-The agent automatically scans the following paths for MCP configurations:
+## 📦 Built-in Toolset
 
-* `~/.cursor/mcp.json`
-* `~/.vscode/mcp.json`
+| Tool              | Description                                                              |
+| ----------------- | ------------------------------------------------------------------------ |
+| `generate_review` | Finalizes the process by submitting a structured violation report.       |
+| `get_repo_map`    | Generates a high-level map of the project files and exports.             |
+| `read_full_code`  | Reads file content with line numbers for precise auditing.               |
+| `read_skeleton`   | Extracts class/function signatures without full logic (Token efficient). |
 
-Ensure your MCP servers are configured in these files, and `sagent` will gain the ability to call them immediately.
+---
+
+## 📋 Audit Rule Configuration
+
+You can pass structured rules via the `extraSystemPrompt`:
+
+```javascript
+const agent = new McpAgent({
+  extraSystemPrompt: {
+    role: "Code Auditor",
+    rules: [
+      { id: "THEME-001", name: "Theme Check", description: "No hardcoded hex colors." }
+    ]
+  }
+});
+
+```
 
-## 🛠️ Usage
+---
 
-| Command                  | Description                                     |
-| ------------------------ | ----------------------------------------------- |
-| `~/.saber2pr-agent.json` | Manually edit this file to update API settings. |
-| `exit`                   | Type during a chat to quit the program.         |
-| `sagent`                 | Enter interactive chat mode.                    |
+## ⚙️ Configuration
 
-## 🏗️ Tech Stack
+The agent stores API keys and base URLs in `~/.saber2pr-agent.json`.
 
-Built with:
+* `baseURL`: The API endpoint.
+* `apiKey`: Your authentication key.
+* `model`: The model name (e.g., `gpt-4o`, `claude-3-5-sonnet`).
 
-* [@modelcontextprotocol/sdk](https://www.google.com/search?q=https://github.com/modelcontextprotocol/typescript-sdk) - Official MCP SDK.
-* [openai](https://www.google.com/search?q=https://github.com/openai/openai-node) - Client for API interactions.
-* [TypeScript](https://www.google.com/search?q=https://www.typescriptlang.org/) - Ensuring type safety and robustness.
+---
 
-## 📄 License
+## 📜 License
 
-[ISC](https://www.google.com/search?q=./LICENSE) © saber2pr
+ISC

package/lib/agent-chain.d.ts

@@ -0,0 +1,48 @@
+import { BaseChatModel } from "@langchain/core/language_models/chat_models";
+interface ApiConfig {
+    baseURL: string;
+    apiKey: string;
+    model: string;
+}
+export interface CustomTool {
+    name: string;
+    description: string;
+    parameters: any;
+    handler: (args: any) => Promise<any>;
+}
+export interface AgentOptions {
+    targetDir?: string;
+    tools?: CustomTool[];
+    extraSystemPrompt?: any;
+    maxTokens?: number;
+    apiConfig?: ApiConfig;
+    apiModel?: BaseChatModel;
+    maxIterations?: number;
+}
+export default class McpChainAgent {
+    private allTools;
+    private messages;
+    private engine;
+    private encoder;
+    private extraTools;
+    private maxTokens;
+    private executor?;
+    private apiConfig;
+    private maxIterations;
+    private apiModel?;
+    constructor(options?: AgentOptions);
+    /**
+     * 工具处理器包装逻辑：增加日志打印和 Token 监控
+     */
+    private wrapHandler;
+    private registerBuiltinTools;
+    private injectCustomTools;
+    private calculateTokens;
+    private pruneMessages;
+    init(): Promise<void>;
+    chat(input: string): Promise<string>;
+    private showLoading;
+    start(): Promise<void>;
+    private ensureApiConfig;
+}
+export {};

package/lib/agent-chain.js

@@ -0,0 +1,315 @@
+"use strict";
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+const fs_1 = __importDefault(require("fs"));
+const path_1 = __importDefault(require("path"));
+const os_1 = __importDefault(require("os"));
+const readline = __importStar(require("readline"));
+const ts_context_mcp_1 = require("@saber2pr/ts-context-mcp");
+const js_tiktoken_1 = require("js-tiktoken");
+const openai_1 = require("@langchain/openai");
+const tools_1 = require("@langchain/core/tools");
+const agents_1 = require("langchain/agents");
+const prompts_1 = require("@langchain/core/prompts");
+const CONFIG_FILE = path_1.default.join(os_1.default.homedir(), ".saber2pr-agent.json");
+class McpChainAgent {
+    constructor(options) {
+        this.allTools = [];
+        this.messages = [];
+        this.encoder = (0, js_tiktoken_1.getEncoding)("cl100k_base");
+        this.extraTools = [];
+        this.engine = new ts_context_mcp_1.PromptEngine((options === null || options === void 0 ? void 0 : options.targetDir) || process.cwd());
+        this.extraTools = (options === null || options === void 0 ? void 0 : options.tools) || [];
+        this.maxTokens = (options === null || options === void 0 ? void 0 : options.maxTokens) || 100000;
+        this.apiConfig = options === null || options === void 0 ? void 0 : options.apiConfig;
+        this.maxIterations = (options === null || options === void 0 ? void 0 : options.maxIterations) || 20;
+        this.apiModel = options === null || options === void 0 ? void 0 : options.apiModel;
+        const baseSystemPrompt = `你是一个专业的 AI 代码架构师，具备深度的源码分析与工程化处理能力。
+    
+### 核心操作规范：
+1. **全局扫描（强制首选）**：在开始任何分析任务前，你【必须】首先调用 'get_repo_map'。这是理解项目结构、技术栈及模块关系的唯一来源。
+2. **循序渐进的分析路径**：
+   - 优先使用 'read_skeleton' 提取接口和函数签名。
+   - 仅在需要分析具体逻辑或准备修复代码时，才使用 'read_full_code'。
+3. **真实性原则**：所有的代码分析必须基于工具返回的真实内容，严禁虚假猜测。`;
+        this.messages.push({
+            role: "system",
+            content: (options === null || options === void 0 ? void 0 : options.extraSystemPrompt)
+                ? `${baseSystemPrompt}\n\n[额外指令]:\n${JSON.stringify(options.extraSystemPrompt)}`
+                : baseSystemPrompt,
+        });
+        this.registerBuiltinTools();
+        this.injectCustomTools();
+    }
+    /**
+     * 工具处理器包装逻辑：增加日志打印和 Token 监控
+     */
+    wrapHandler(name, handler) {
+        return async (args) => {
+            // 1. 打印工具执行日志
+            console.log(`\n   [工具调用]: ${name}`);
+            if (args === null || args === void 0 ? void 0 : args.filePath) {
+                console.log(`   [目标文件]: ${args.filePath}`);
+            }
+            // 2. 执行逻辑
+            const result = await handler(args);
+            const content = typeof result === "string" ? result : JSON.stringify(result);
+            // 3. 统计 Token 消耗
+            const tokens = this.encoder.encode(content).length;
+            console.log(`   [输出长度]: ${tokens} tokens`);
+            return content;
+        };
+    }
+    registerBuiltinTools() {
+        const builtinTools = [
+            {
+                type: "function",
+                function: { name: "get_repo_map", description: "获取项目全局结构图和导出清单", parameters: { type: "object" } },
+                _handler: this.wrapHandler("get_repo_map", async () => {
+                    this.engine.refresh();
+                    return this.engine.getRepoMap();
+                }),
+            },
+            {
+                type: "function",
+                function: {
+                    name: "read_skeleton",
+                    description: "读取代码骨架（接口、类定义等），非常节省 Token",
+                    parameters: { type: "object", properties: { filePath: { type: "string" } } },
+                },
+                _handler: this.wrapHandler("read_skeleton", async ({ filePath }) => this.engine.getSkeleton(filePath)),
+            },
+            {
+                type: "function",
+                function: {
+                    name: "read_full_code",
+                    description: "读取完整源码。注意：仅在需要具体行号或精细逻辑时使用",
+                    parameters: { type: "object", properties: { filePath: { type: "string" } } },
+                },
+                _handler: this.wrapHandler("read_full_code", async ({ filePath }) => {
+                    // --- 新增：Token 守卫 ---
+                    const currentTokens = this.calculateTokens();
+                    if (currentTokens > this.maxTokens) {
+                        return `[SYSTEM WARNING]: 当前上下文已达到 ${currentTokens} tokens (上限 ${this.maxTokens})。为了保证系统稳定，已拦截 read_full_code。请立即根据已知信息进行总结或停止阅读更多代码。`;
+                    }
+                    try {
+                        if (typeof filePath !== 'string' || !filePath) {
+                            return "Error: filePath 不能为空";
+                        }
+                        // 拼合绝对路径
+                        const fullPath = path_1.default.resolve(this.engine.getRootDir(), filePath);
+                        // 安全检查：防止 AI 尝试读取项目外的敏感文件
+                        if (!fullPath.startsWith(this.engine.getRootDir())) {
+                            return "Error: 权限拒绝，禁止访问项目目录外的文件。";
+                        }
+                        if (!fs_1.default.existsSync(fullPath)) {
+                            return `Error: 文件不存在: ${filePath}`;
+                        }
+                        const content = fs_1.default.readFileSync(fullPath, "utf-8");
+                        // 加上行号，AI 就能在 generate_review 里给出准确的 line 参数
+                        return content.split('\n')
+                            .map((line, i) => `${i + 1} | ${line}`)
+                            .join('\n');
+                    }
+                    catch (err) {
+                        return `Error: 读取文件失败: ${err.message}`;
+                    }
+                }),
+            }
+        ];
+        this.allTools.push(...builtinTools);
+    }
+    injectCustomTools() {
+        for (const tool of this.extraTools) {
+            this.allTools.push({
+                type: "function",
+                function: { name: tool.name, description: tool.description, parameters: tool.parameters },
+                _handler: this.wrapHandler(tool.name, tool.handler),
+            });
+        }
+    }
+    calculateTokens() {
+        return this.messages.reduce((acc, msg) => acc + this.encoder.encode(String(msg.content || "")).length, 0);
+    }
+    pruneMessages() {
+        const current = this.calculateTokens();
+        if (current > this.maxTokens) {
+            console.log(`\n⚠️ 上下文达到限制 (${current} tokens)，正在裁剪旧消息...`);
+            // 保留 system prompt (index 0)，移除后续消息
+            while (this.calculateTokens() > this.maxTokens * 0.8 && this.messages.length > 2) {
+                this.messages.splice(1, 1);
+            }
+            console.log(`✅ 裁剪完成，当前: ${this.calculateTokens()} tokens`);
+        }
+    }
+    async init() {
+        if (this.executor)
+            return;
+        let model;
+        if (this.apiModel) {
+            console.log("ℹ️ 使用自定义 API Model 实例");
+            model = this.apiModel;
+        }
+        else {
+            // 降级方案：使用配置创建默认的 ChatOpenAI
+            const apiConfig = await this.ensureApiConfig();
+            console.log(`ℹ️ 使用默认 ChatOpenAI (${apiConfig.model})`);
+            model = new openai_1.ChatOpenAI({
+                configuration: { baseURL: apiConfig.baseURL, apiKey: apiConfig.apiKey },
+                modelName: apiConfig.model,
+                temperature: 0,
+                streaming: false
+            });
+        }
+        const langchainTools = this.allTools.map(t => new tools_1.DynamicTool({
+            name: t.function.name,
+            description: t.function.description || "",
+            func: t._handler
+        }));
+        const prompt = prompts_1.PromptTemplate.fromTemplate(`
+{system_prompt}
+
+TOOLS:
+------
+You can use the following tools:
+{tools}
+
+To use a tool, please use the following format:
+Thought: Do I need to use a tool? Yes
+Action: the action to take, should be one of [{tool_names}]
+Action Input: the input to the action (JSON format)
+Observation: the result of the action
+... (repeat N times)
+Thought: I now know the final answer
+Final Answer: the final answer to the original input question
+
+Begin!
+Question: {input}
+Thought: {agent_scratchpad}`);
+        const agent = await (0, agents_1.createReactAgent)({ llm: model, tools: langchainTools, prompt });
+        this.executor = new agents_1.AgentExecutor({
+            agent,
+            tools: langchainTools,
+            verbose: false, // 我们已经有了 wrapHandler 日志，关闭原生 verbose 以保持整洁
+            handleParsingErrors: true,
+            maxIterations: this.maxIterations
+        });
+    }
+    async chat(input) {
+        var _a;
+        if (!this.executor)
+            await this.init();
+        this.messages.push({ role: "user", content: input });
+        this.pruneMessages();
+        console.log(`\n📊 状态: Context ${this.calculateTokens()} / Limit ${this.maxTokens} tokens`);
+        const stopLoading = this.showLoading("🤖 Agent 正在思考并执行工具...");
+        try {
+            const response = await this.executor.invoke({
+                input: input,
+                system_prompt: this.messages[0].content,
+            });
+            let output = response.output;
+            // 清洗 ReAct 冗余标签
+            if (output.includes("Final Answer:")) {
+                output = ((_a = output.split("Final Answer:").pop()) === null || _a === void 0 ? void 0 : _a.trim()) || output;
+            }
+            this.messages.push({ role: "assistant", content: output });
+            return output;
+        }
+        finally {
+            stopLoading();
+        }
+    }
+    showLoading(text) {
+        const chars = ['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
+        let i = 0;
+        const timer = setInterval(() => {
+            process.stdout.write(`\r${chars[i]} ${text}`);
+            i = (i + 1) % chars.length;
+        }, 80);
+        return () => {
+            clearInterval(timer);
+            process.stdout.write('\r\x1b[K');
+        };
+    }
+    async start() {
+        await this.init();
+        const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
+        console.log(`\n🚀 AI 助手启动 (LangChain 核心)`);
+        console.log(`📂 目标目录: ${this.engine.getRootDir()}`);
+        const chatLoop = () => {
+            rl.question("\n👤 你: ", async (input) => {
+                if (!input.trim())
+                    return chatLoop();
+                if (input.toLowerCase() === "exit")
+                    process.exit(0);
+                try {
+                    const result = await this.chat(input);
+                    console.log(`\n🤖 Agent: ${result}`);
+                }
+                catch (err) {
+                    console.error("\n❌ 系统错误:", err.message);
+                }
+                chatLoop();
+            });
+        };
+        chatLoop();
+    }
+    async ensureApiConfig() {
+        if (this.apiConfig)
+            return this.apiConfig;
+        if (fs_1.default.existsSync(CONFIG_FILE)) {
+            return JSON.parse(fs_1.default.readFileSync(CONFIG_FILE, "utf-8"));
+        }
+        const rl = readline.createInterface({
+            input: process.stdin,
+            output: process.stdout,
+        });
+        const question = (q) => new Promise((res) => rl.question(q, res));
+        console.log("\n🔑 配置 API 凭据:");
+        const config = {
+            baseURL: await question("? API Base URL (如 https://api.openai.com/v1): "),
+            apiKey: await question("? API Key: "),
+            model: await question("? Model Name (如 gpt-4o): "),
+        };
+        fs_1.default.writeFileSync(CONFIG_FILE, JSON.stringify(config, null, 2));
+        rl.close();
+        return config;
+    }
+}
+exports.default = McpChainAgent;

package/lib/agent.d.ts

@@ -1,3 +1,8 @@
+interface ApiConfig {
+    baseURL: string;
+    apiKey: string;
+    model: string;
+}
 export interface CustomTool {
     name: string;
     description: string;
@@ -10,6 +15,8 @@ export interface AgentOptions {
     tools?: any[];
     /** 注入到 System Prompt 中的额外指令/规则/上下文 */
     extraSystemPrompt?: any;
+    maxTokens?: number;
+    apiConfig?: ApiConfig;
 }
 export default class McpAgent {
     private openai;
@@ -20,6 +27,8 @@ export default class McpAgent {
     private engine;
     private encoder;
     private extraTools;
+    private maxTokens;
+    private apiConfig;
     constructor(options?: AgentOptions);
     /**
      * 计算当前消息列表的总 Token 消耗
@@ -36,6 +45,11 @@ export default class McpAgent {
     private loadMcpConfigs;
     init(): Promise<void>;
     private processChat;
+    /**
+   * 裁剪上下文消息列表
+   * 保留第一条 System 消息，并移除中间的旧消息直到低于阈值
+   */
+    private pruneMessages;
     /**
      * 简易 Loading 动画辅助函数
      */
@@ -48,3 +62,4 @@ export default class McpAgent {
     chat(input: string): Promise<string>;
     start(): Promise<void>;
 }
+export {};

package/lib/agent.js

@@ -56,11 +56,26 @@ class McpAgent {
         this.extraTools = [];
         this.engine = new ts_context_mcp_1.PromptEngine((options === null || options === void 0 ? void 0 : options.targetDir) || process.cwd());
         this.extraTools = (options === null || options === void 0 ? void 0 : options.tools) || []; // 接收外部传入的工具
-        let baseSystemPrompt = `你是一个专业的 AI 代码架构师。
-你可以访问本地文件系统并利用 AST (抽象语法树) 技术分析代码。
-你的核心目标是提供准确的代码结构、依赖关系和逻辑分析。
-请先使用 get_repo_map 查看项目整体代码结构。
-请优先使用 read_skeleton 查看结构，只有在必要时才使用 read_full_code 或 get_method_body。`;
+        this.maxTokens = (options === null || options === void 0 ? void 0 : options.maxTokens) || 100000; // 默认 100k
+        this.apiConfig = options === null || options === void 0 ? void 0 : options.apiConfig;
+        let baseSystemPrompt = `你是一个专业的 AI 代码架构师，具备深度的源码分析与工程化处理能力。
+
+    ### 核心操作规范：
+    1. **全局扫描（强制首选）**：在开始任何分析任务前，你【必须】首先调用 'get_repo_map'。这是你理解项目目录结构、技术栈及模块关系的唯一权威来源。
+    2. **循序渐进的分析路径**：
+       - 优先使用 'read_skeleton' 提取接口、类和函数签名，以最低的 Token 成本建立代码逻辑视图。
+       - 仅在需要深入分析具体业务逻辑、提取精准代码块或进行代码修改建议时，才使用 'read_full_code' 或 'get_method_body'。
+    3. **真实性原则**：
+       - 所有的代码分析、行号定位和逻辑推断必须基于工具返回的真实内容，严禁基于文件名进行虚假猜测。
+       - 如果工具返回结果为空或报错，应尝试调整路径或更换工具。
+    
+    ### 技术能力：
+    - 精通 TypeScript/JavaScript 及其 AST 结构，能准确识别各种复杂的声明与调用关系。
+    - 能够理解代码间的依赖链路，并结合项目上下文给出合理的架构建议。
+    
+    ### 执行准则：
+    - **任务导向**：直接通过工具链解决问题，减少不必要的中间对话。
+    - **自主决策**：根据任务需求自主选择最合适的工具组合，无需每一步都向用户请示。`;
         // 2. 拼接额外指令
         if (options === null || options === void 0 ? void 0 : options.extraSystemPrompt) {
             const extra = typeof options.extraSystemPrompt === 'string'
@@ -211,7 +226,13 @@ class McpAgent {
                         required: ["filePath", "methodName"],
                     },
                 },
-                _handler: async ({ filePath, methodName }) => this.engine.getMethodImplementation(filePath, methodName),
+                _handler: async ({ filePath, methodName }) => {
+                    // --- 新增：同样的 Token 守卫 ---
+                    if (this.calculateTokens() > this.maxTokens) {
+                        return `[SYSTEM WARNING]: Token 消耗已达上限，禁止获取详细方法体。请利用已获取的 Skeleton 信息进行分析。`;
+                    }
+                    return this.engine.getMethodImplementation(filePath, methodName);
+                },
             },
             {
                 type: "function",
@@ -228,6 +249,11 @@ class McpAgent {
                 },
                 // 核心实现：直接利用 fs 读取
                 _handler: async ({ filePath }) => {
+                    // --- 新增：Token 守卫 ---
+                    const currentTokens = this.calculateTokens();
+                    if (currentTokens > this.maxTokens) {
+                        return `[SYSTEM WARNING]: 当前上下文已达到 ${currentTokens} tokens (上限 ${this.maxTokens})。为了保证系统稳定，已拦截 read_full_code。请立即根据已知信息进行总结或停止阅读更多代码。`;
+                    }
                     try {
                         if (typeof filePath !== 'string' || !filePath) {
                             return "Error: filePath 不能为空";
@@ -257,6 +283,8 @@ class McpAgent {
     }
     // --- 初始化与环境准备 (API Config & MCP Servers) ---
     async ensureApiConfig() {
+        if (this.apiConfig)
+            return this.apiConfig;
         if (fs_1.default.existsSync(CONFIG_FILE)) {
             return JSON.parse(fs_1.default.readFileSync(CONFIG_FILE, "utf-8"));
         }
@@ -329,9 +357,18 @@ class McpAgent {
         var _a;
         this.messages.push({ role: 'user', content: userInput });
         while (true) {
+            // --- 新增：发送请求前先检查并裁剪 ---
+            this.pruneMessages();
             // 打印当前上下文的累计 Token
             const currentInputTokens = this.calculateTokens();
             console.log(`\n📊 当前上下文累计: ${currentInputTokens} tokens`);
+            // 如果接近上限（如 80%），在消息队列中插入一条隐含的系统指令
+            if (currentInputTokens > this.maxTokens * 0.8 && currentInputTokens <= this.maxTokens) {
+                this.messages.push({
+                    role: "system",
+                    content: "注意：上下文即将耗尽。请停止读取新文件，优先处理现有信息并尽快输出结果。"
+                });
+            }
             const stopLoading = this.showLoading("🤖 Agent 正在思考...");
             let response;
             try {
@@ -388,6 +425,24 @@ class McpAgent {
             }
         }
     }
+    /**
+   * 裁剪上下文消息列表
+   * 保留第一条 System 消息，并移除中间的旧消息直到低于阈值
+   */
+    pruneMessages() {
+        const currentTokens = this.calculateTokens();
+        if (currentTokens <= this.maxTokens)
+            return;
+        console.log(`\n⚠️ 上下文达到限制 (${currentTokens} tokens)，正在自动裁剪...`);
+        // 策略：保留索引 0 (System)，从索引 1 开始删除
+        // 每次删除一对 (通常是助理请求 + 工具回复，或者用户提问 + 助理回答)
+        while (this.calculateTokens() > this.maxTokens && this.messages.length > 2) {
+            // 始终保留系统提示词 (index 0) 和最后一条消息 (保持对话连贯)
+            // 删除索引为 1 的消息
+            this.messages.splice(1, 1);
+        }
+        console.log(`✅ 裁剪完成，当前上下文: ${this.calculateTokens()} tokens`);
+    }
     /**
      * 简易 Loading 动画辅助函数
      */

package/lib/cli-chain.d.ts

@@ -0,0 +1,2 @@
+#!/usr/bin/env node
+export {};

package/lib/cli-chain.js

@@ -0,0 +1,9 @@
+#!/usr/bin/env node
+"use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+const agent_chain_1 = __importDefault(require("./agent-chain"));
+const manager = new agent_chain_1.default();
+manager.start();

package/lib/index.d.ts

@@ -1,2 +1,3 @@
 export * from './agent';
+export { default as McpChainAgent } from './agent-chain';
 export { default } from './agent';

package/lib/index.js

@@ -17,7 +17,9 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
     return (mod && mod.__esModule) ? mod : { "default": mod };
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.default = void 0;
+exports.default = exports.McpChainAgent = void 0;
 __exportStar(require("./agent"), exports);
+var agent_chain_1 = require("./agent-chain");
+Object.defineProperty(exports, "McpChainAgent", { enumerable: true, get: function () { return __importDefault(agent_chain_1).default; } });
 var agent_1 = require("./agent");
 Object.defineProperty(exports, "default", { enumerable: true, get: function () { return __importDefault(agent_1).default; } });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@saber2pr/ai-agent",
-  "version": "0.0.6",
+  "version": "0.0.8",
   "description": "AI Assistant CLI.",
   "author": "saber2pr",
   "license": "ISC",
@@ -8,7 +8,8 @@
     "lib"
   ],
   "bin": {
-    "sagent": "./lib/cli.js"
+    "sagent": "./lib/cli.js",
+    "sagent-chain": "./lib/cli-chain.js"
   },
   "publishConfig": {
     "access": "public",
@@ -21,9 +22,12 @@
     "prepublishOnly": "tsc"
   },
   "dependencies": {
+    "@langchain/core": "^1.1.18",
+    "@langchain/openai": "^1.2.4",
     "@modelcontextprotocol/sdk": "^1.25.3",
     "@saber2pr/ts-context-mcp": "^0.0.6",
     "js-tiktoken": "^1.0.21",
+    "langchain": "~0.3",
     "openai": "^6.16.0"
   },
   "devDependencies": {