@saber2pr/ai-agent 0.0.26 → 0.0.28

package/README.md

@@ -1,205 +1,101 @@
-# 🚀 Saber2pr AI Agent
+# 🤖 AI-Agent
 
-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.
+A professional code architect assistant built with **Model Context Protocol (MCP)** and **LangGraph**. It bridges Large Language Models (LLMs) with local file systems and remote services through standardized workflows.
 
-## ✨ Core Features
+## 🌟 Core Modes
 
-* **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.
+This project provides two core interaction modes to adapt to different LLM capabilities:
 
+### 1. 🚀 Standard Agent (`sagent`)
 
-* **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`.
+**Best for**: Modern models that support native `tools` parameters (e.g., GPT-4o, Claude 3.5, Qwen-Max).
 
----
-
-## 🛠️ Installation
-
-```bash
-sudo npm i -g @saber2pr/ai-agent
-
-# call openapi
-sagent
-
-# call third api
-sagent-chain
-
-# Clone the repository
-git clone https://github.com/saber2pr/ai-agent.git
-cd ai-agent
+* **Native Function Calling**: Leverages the model's built-in tool-calling capabilities for precise parameter parsing and low latency.
+* **Intuitive Interaction**: A classic Chat loop suitable for instant Q&A, quick code lookups, or single-file refactoring tasks.
 
-# Install dependencies
-npm install
+### 2. 🏗️ Graph Agent (`sagent-graph`)
 
-# Build the project
-npm run build
+**Best for**: Long-running automated audits or scenarios where the API does not support native `tools`.
 
-```
+* **State Machine Architecture**: Built on **LangGraph** to implement an automated "Think-Act-Observe-Summarize" loop.
+* **Audit Tracking**: Built-in progress manager that automatically records analyzed files to prevent redundant analysis in complex projects.
+* **Auto-Settlement**: Automatically summarizes and prints **Total Tokens** and execution duration upon task completion or manual exit.
 
 ---
 
-## 🚀 Usage Modes
+## 🛠️ Built-in Toolset
 
-### 1. Standard Edition (Direct API)
+The Agent comes pre-installed with a suite of tools specifically customized for **code architecture analysis**:
 
-Best for quick scripts and simple chat interactions. It uses a straightforward message-loop logic.
+| Category         | Tool Name         | Description                                                               |
+| ---------------- | ----------------- | ------------------------------------------------------------------------- |
+|                  | `list_directory`  | Lists files in a specified directory.                                     |
+|                  | `analyze_deps`    | Analyzes import dependencies of a specific file.                          |
+|                  | `read_skeleton`   | Reads only code skeletons (class names/signatures) to save tokens.        |
+|                  | `edit_file`       | Precise Diff-based editing to avoid rewriting large files.                |
+|                  | `get_method_body` | Precisely extracts the implementation of a specific method.               |
+|                  | `get_file_info`   | Gets file metadata like size, permissions, and timestamps.                |
+| **Architecture** | `get_repo_map`    | **Core**: Extracts export definitions and builds a module dependency map. |
+| **General**      | `search_files`    | Performs a global keyword search across the project.                      |
+| **Navigation**   | `directory_tree`  | Recursively gets the project structure (The entry point for analysis).    |
+| **Operations**   | `read_text_file`  | Reads file content (supports pagination via head/tail).                   |
 
-```javascript
-import McpAgent from "@saber2pr/ai-agent";
+---
 
-const agent = new McpAgent({
-  targetDir: "/path/to/project"
-});
+## 💻 CLI Guide
 
-await agent.chat("Analyze the directory structure.");
+### Installation & Build
 
+```bash
+yarn
+yarn build
 ```
 
-### 2. LangChain Edition (Advanced Agent)
-
-Best for complex tasks like "Audit the whole project and fix bugs." It supports autonomous tool usage.
-
-```javascript
-import { McpChainAgent } from "@saber2pr/ai-agent";
-import { MyPrivateLLM } from "./your-custom-llm";
+### Usage
 
-const agent = new McpChainAgent({
-  apiModel: new MyPrivateLLM(), // Inject custom LLM
-  maxIterations: 15,
-  targetDir: "/path/to/project"
-});
-
-await agent.chat("Scan for hardcoded colors and submit a review report.");
+```sh
+sudo npm i -g @saber2pr/ai-agent
 
 ```
 
----
-
-## 🔧 Extending with Private LLMs
-
-### Using AgentChainModel (Recommended)
-
-For LangChain mode, extend `AgentChainModel` which provides a simplified interface for integrating custom LLMs:
-
-```javascript
-import { AgentChainModel } from "@saber2pr/ai-agent";
-
-class MyPrivateLLM extends AgentChainModel {
-  constructor(fields) { 
-    super(fields || {}); 
-  }
-
-  async generateAgentChainResponse(messages) {
-    const lastMessage = messages[messages.length - 1];
-    const queryText = lastMessage.content;
-    
-    const response = await fetch("https://your-api-gateway.com/api/completions", {
-      method: 'POST',
-      body: JSON.stringify({ query: queryText, stream: false }),
-      headers: {
-        'Content-Type': 'application/json',
-        Authorization: `Bearer YOUR_API_KEY`,
-      }
-    });
-
-    if (!response.ok) {
-      const errorText = await response.text();
-      throw new Error(`HTTP Error: ${response.status}, ${errorText}`);
-    }
-
-    const data = await response.json();
-    let text = data.text || "";
-    
-    // Handle special response formats if needed
-    if (text.includes("Action:") && text.includes("Final Answer:")) {
-      text = text.split("Final Answer:")[0].trim();
-    }
-    
-    return text;
-  }
-}
+* **Start Standard Chat**:
+```bash
+sagent
 ```
 
-**Key Points:**
-- `AgentChainModel` abstracts away LangChain's internal message handling
-- You only need to implement `generateAgentChainResponse(messages)` which receives an array of messages
-- The method should return a plain string response
-- The base class handles conversion to LangChain's expected format
-
----
-
-## 📦 Built-in Toolset
-
-The toolkit provides a comprehensive set of built-in tools organized into two categories: **Filesystem Tools** and **Code Analysis Tools**. All tools operate within the `targetDir` scope for security.
-
-### Filesystem Tools
-
-| Tool | Description | Parameters |
-|------|-------------|------------|
-| `read_text_file` | Read complete file contents as text. Supports `head` and `tail` parameters for partial reading. Handles various text encodings. | `path` (required), `head?`, `tail?` |
-| `read_multiple_files` | Read multiple files simultaneously for efficient batch analysis. Individual failures won't stop the operation. | `paths` (array, required) |
-| `write_file` | Create a new file or completely overwrite an existing file. Use with caution as it overwrites without warning. | `path` (required), `content` (required) |
-| `edit_file` | Make line-based edits to a text file. Replaces exact line sequences with new content. Returns git-style diff. Supports `dryRun` mode for preview. | `path` (required), `edits` (array, required), `dryRun?` |
-| `get_directory_tree` | Get a recursive tree view of files and directories as JSON. Supports `excludePatterns` for filtering (minimatch patterns). Essential for understanding project structure. | `path` (required), `excludePatterns?` (array) |
-| `list_directory` | List all files and directories in a specified path. Results distinguish files and directories with `[FILE]` and `[DIR]` prefixes. | `path` (required) |
-| `list_directory_with_sizes` | List directory contents with file sizes. Supports sorting by name or size. Includes summary statistics. | `path` (required), `sortBy?` ("name" \| "size") |
-| `search_files` | Search for files matching a glob pattern. Supports exclude patterns for filtering. | `path` (required), `pattern` (required), `excludePatterns?` (array) |
-| `move_file` | Move or rename files and directories. Can move between directories and rename in a single operation. | `source` (required), `destination` (required) |
-| `create_directory` | Create a new directory or ensure it exists. Can create multiple nested directories recursively. | `path` (required) |
-| `get_file_info` | Get detailed metadata about a file: size, last modified time, type, etc. | `path` (required) |
 
-### Code Analysis Tools
-
-| Tool | Description | Parameters |
-|------|-------------|------------|
-| `get_repo_map` | Generate a high-level map of project files and exports. Extracts export definitions to understand module relationships. Use this first to understand project structure. | None |
-| `read_skeleton` | Extract structural definitions (interfaces, classes, method signatures) without implementation details. Token-efficient for code analysis. | `filePath` (required) |
-| `analyze_deps` | Analyze dependency relationships for a specific file. Supports TypeScript path alias resolution via tsconfig. | `filePath` (required) |
-| `get_method_body` | Get the complete implementation code for a specific method or function within a file. | `filePath` (required), `methodName` (required) |
-
-### Tool Usage Tips
-
-1. **Start with `get_directory_tree`**: Always begin by understanding the project structure before reading files.
-2. **Use `read_skeleton` before `read_text_file`**: Extract signatures first to save tokens, then read full content only when needed.
-3. **Leverage `excludePatterns`**: Use minimatch patterns to exclude `node_modules`, `.git`, build artifacts, etc.
-4. **Batch operations**: Use `read_multiple_files` when analyzing multiple files to improve efficiency.
-5. **Preview changes**: Use `edit_file` with `dryRun: true` to preview changes before applying them.
+* **Start Automated Audit**:
+```bash
+sagent-graph
+```
 
 ---
 
-## 📋 Audit Rule Configuration
-
-You can pass structured rules via the `extraSystemPrompt`:
+## ⚙️ Configuration
 
-```javascript
-import McpAgent from "@saber2pr/ai-agent";
+On the first run, the program will prompt you to configure `~/.saber2pr-agent.json`. You can define your API keys and dynamically connect **MCP Servers** here:
 
-const agent = new McpAgent({
-  extraSystemPrompt: {
-    role: "Code Auditor",
-    rules: [
-      { id: "THEME-001", name: "Theme Check", description: "No hardcoded hex colors." }
-    ]
-  }
-});
+```json
+{
+  "baseURL": "https://api.example.com/v1",
+  "apiKey": "sk-your-key",
+  "model": "gpt-4o"
+}
 
 ```
 
 ---
 
-## ⚙️ Configuration
+## 🔄 Professional Audit Workflow
 
-The agent stores API keys and base URLs in `~/.saber2pr-agent.json`.
+Regardless of the mode, the Agent follows a standardized logic chain:
 
-* `baseURL`: The API endpoint.
-* `apiKey`: Your authentication key.
-* `model`: The model name (e.g., `gpt-4o`, `claude-3-5-sonnet`).
+1. **Phase 1: Panoramic Perception (Where)**: Uses `directory_tree` to identify project layout (e.g., Monorepo vs. traditional src structure).
+2. **Phase 2: Logic Mapping (What)**: Uses `get_repo_map` to establish logical relationships—understanding "who calls whom."
+3. **Phase 3: Source Diving (How)**: Locates critical implementations and performs detailed auditing via `read_text_file` or specific extraction tools.
 
 ---
 
-## 📜 License
+## 📄 License
 
 ISC

package/lib/core/agent-graph.d.ts

@@ -1,5 +1,5 @@
-import { BaseMessage } from "@langchain/core/messages";
-import { AgentOptions } from "../types/type";
+import { BaseMessage } from '@langchain/core/messages';
+import { GraphAgentOptions } from '../types/type';
 export declare const CONFIG_FILE: string;
 interface TokenUsage {
     total: number;
@@ -23,7 +23,17 @@ export default class McpGraphAgent {
     private verbose;
     private alwaysSystem;
     private recursionLimit;
-    constructor(options?: AgentOptions);
+    private apiConfig;
+    private maxTargetCount;
+    private maxTokens;
+    private mcpClients;
+    constructor(options?: GraphAgentOptions);
+    private printLoadedTools;
+    private loadMcpConfigs;
+    private initMcpTools;
+    private prepareTools;
+    private ensureInitialized;
+    private closeMcpClients;
     private showLoading;
     private startLoading;
     private stopLoading;
@@ -33,7 +43,7 @@ export default class McpGraphAgent {
     start(): Promise<void>;
     private renderOutput;
     callModel(state: typeof AgentState.State): Promise<{
-        messages: unknown[];
+        messages: any[];
         tokenUsage: {
             total: number;
         };
@@ -58,7 +68,7 @@ export default class McpGraphAgent {
         mode: import("@langchain/langgraph").BinaryOperatorAggregate<"chat" | "auto", "chat" | "auto">;
         tokenUsage: import("@langchain/langgraph").BinaryOperatorAggregate<TokenUsage, TokenUsage>;
         totalDuration: import("@langchain/langgraph").BinaryOperatorAggregate<number, number>;
-    }>, "tools" | "agent" | "__start__" | "progress", {
+    }>, "tools" | "__start__" | "agent" | "progress", {
         messages: import("@langchain/langgraph").BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
         auditedFiles: import("@langchain/langgraph").BinaryOperatorAggregate<string[], string[]>;
         targetCount: import("@langchain/langgraph").BinaryOperatorAggregate<number, number>;

package/lib/core/agent-graph.js

@@ -4,18 +4,21 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.CONFIG_FILE = void 0;
-const openai_1 = require("@langchain/openai");
+const events_1 = require("events");
+const fs_1 = __importDefault(require("fs"));
+const os_1 = __importDefault(require("os"));
+const path_1 = __importDefault(require("path"));
+const readline_1 = __importDefault(require("readline"));
 const messages_1 = require("@langchain/core/messages");
+const prompts_1 = require("@langchain/core/prompts");
 const langgraph_1 = require("@langchain/langgraph");
 const prebuilt_1 = require("@langchain/langgraph/prebuilt");
-const prompts_1 = require("@langchain/core/prompts");
-const readline_1 = __importDefault(require("readline"));
-const fs_1 = __importDefault(require("fs"));
-const path_1 = __importDefault(require("path"));
-const os_1 = __importDefault(require("os"));
-const events_1 = require("events");
+const openai_1 = require("@langchain/openai");
 const builtin_1 = require("../tools/builtin");
 const convertToLangChainTool_1 = require("../utils/convertToLangChainTool");
+const index_js_1 = require("@modelcontextprotocol/sdk/client/index.js");
+const stdio_js_1 = require("@modelcontextprotocol/sdk/client/stdio.js");
+const jsonSchemaToZod_1 = require("../utils/jsonSchemaToZod");
 exports.CONFIG_FILE = path_1.default.join(os_1.default.homedir(), ".saber2pr-agent.json");
 // ✅ 全局设置：修复 AbortSignal 监听器数量警告
 // LangChain 的 HTTP 客户端会创建多个 AbortSignal，需要增加默认限制
@@ -56,28 +59,139 @@ class McpGraphAgent {
         this.checkpointer = new langgraph_1.MemorySaver();
         this.langchainTools = [];
         this.stopLoadingFunc = null;
+        this.mcpClients = [];
         this.options = options;
         this.verbose = options.verbose || false;
         this.alwaysSystem = options.alwaysSystem || true;
         this.targetDir = options.targetDir || process.cwd();
-        this.recursionLimit = options.recursionLimit || 200;
+        this.recursionLimit = options.recursionLimit || 80;
+        this.apiConfig = options.apiConfig;
+        this.maxTargetCount = options.maxTargetCount || 4;
+        this.maxTokens = options.maxTokens || 8000;
         process.setMaxListeners(100);
         // ✅ 修复 AbortSignal 监听器数量警告
         // LangChain 的 HTTP 客户端会创建多个 AbortSignal，需要增加默认限制
         // 设置 EventEmitter 的默认 maxListeners，这会影响所有事件发射器（包括 AbortSignal）
         events_1.EventEmitter.defaultMaxListeners = 100;
-        const cleanup = () => {
+        const cleanup = async () => {
             this.stopLoading();
+            await this.closeMcpClients(); // 清理 MCP 连接
             process.stdout.write('\u001B[?25h');
             process.exit(0);
         };
         process.on("SIGINT", cleanup);
         process.on("SIGTERM", cleanup);
-        // ✅ 初始化内置工具
-        const builtinToolInfos = (0, builtin_1.createDefaultBuiltinTools)({ options });
-        this.langchainTools = [...builtinToolInfos, ...(options.tools || [])].map((t) => (0, convertToLangChainTool_1.convertToLangChainTool)(t));
+    }
+    printLoadedTools() {
+        console.log("\n🛠️  [Graph] 正在加载工具节点...");
+        this.langchainTools.forEach((tool) => {
+            // 工具名称
+            console.log(`\n🧰 工具名: ${tool.name}`);
+            // 提取参数结构 (LangChain Tool 的 schema 是 Zod 对象)
+            const schema = tool.schema;
+            if (schema && schema.shape) {
+                // 如果是 ZodObject，打印其内部 key
+                const keys = Object.keys(schema.shape);
+                console.log(`   参数结构: [ ${keys.join(", ")} ]`);
+            }
+            else if (schema && schema._def) {
+                // 兼容其他 Zod 类型
+                console.log(`   参数类型: ${schema._def.typeName}`);
+            }
+            else {
+                // 降级：如果已经是 JSON 对象
+                console.log(`   参数结构:`, JSON.stringify(schema, null, 2));
+            }
+        });
+        console.log(`\n✅ Graph 节点就绪，总计加载 ${this.langchainTools.length} 个工具。\n`);
+    }
+    loadMcpConfigs() {
+        const combined = { mcpServers: {} };
+        const paths = [
+            path_1.default.join(os_1.default.homedir(), ".cursor", "mcp.json"),
+            path_1.default.join(os_1.default.homedir(), ".vscode", "mcp.json"),
+        ];
+        paths.forEach((p) => {
+            if (fs_1.default.existsSync(p)) {
+                const content = JSON.parse(fs_1.default.readFileSync(p, "utf-8"));
+                Object.assign(combined.mcpServers, content.mcpServers);
+            }
+        });
+        return combined;
+    }
+    async initMcpTools() {
+        const mcpConfig = this.loadMcpConfigs();
+        const mcpServers = mcpConfig.mcpServers || {};
+        const mcpToolInfos = [];
+        for (const [name, config] of Object.entries(mcpServers)) {
+            try {
+                const transport = new stdio_js_1.StdioClientTransport({
+                    command: config.command,
+                    args: config.args,
+                    env: { ...process.env, ...(config.env || {}) },
+                });
+                const client = new index_js_1.Client({ name: "mcp-graph-client", version: "1.0.0" }, { capabilities: {} });
+                await client.connect(transport);
+                this.mcpClients.push(client);
+                const { tools } = await client.listTools();
+                tools.forEach((tool) => {
+                    mcpToolInfos.push({
+                        type: "function",
+                        function: {
+                            name: tool.name,
+                            description: tool.description,
+                            parameters: (0, jsonSchemaToZod_1.jsonSchemaToZod)(tool.inputSchema), // MCP 使用 JSON Schema
+                        },
+                        _handler: async (args) => {
+                            const result = await client.callTool({
+                                name: tool.name,
+                                arguments: args,
+                            });
+                            return result.content;
+                        },
+                    });
+                });
+                console.log(`\n✅ 已连接 MCP 服务 [${name}]: 加载了 ${tools.length} 个工具`);
+            }
+            catch (error) {
+                console.error(`\n❌ 连接 MCP 服务 [${name}] 失败:`, error.message);
+            }
+        }
+        return mcpToolInfos;
+    }
+    async prepareTools() {
+        const builtinToolInfos = (0, builtin_1.createDefaultBuiltinTools)({ options: this.options });
+        const mcpToolInfos = await this.initMcpTools();
+        // 合并内置、手动传入和 MCP 工具
+        const allToolInfos = [
+            ...builtinToolInfos,
+            ...(this.options.tools || []),
+            ...mcpToolInfos
+        ];
+        this.langchainTools = allToolInfos.map((t) => (0, convertToLangChainTool_1.convertToLangChainTool)(t));
         this.toolNode = new prebuilt_1.ToolNode(this.langchainTools);
     }
+    // ✅ 修改：初始化逻辑
+    async ensureInitialized() {
+        if (this.model && this.langchainTools.length > 0)
+            return;
+        // 1. 加载所有工具（含 MCP）
+        await this.prepareTools();
+        // 2. 初始化模型
+        await this.getModel();
+        // 3. 打印工具状态
+        this.printLoadedTools();
+    }
+    // ✅ 新增：关闭连接
+    async closeMcpClients() {
+        for (const client of this.mcpClients) {
+            try {
+                await client.close();
+            }
+            catch (e) { }
+        }
+        this.mcpClients = [];
+    }
     showLoading(text) {
         const chars = ['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
         let i = 0;
@@ -113,13 +227,15 @@ class McpGraphAgent {
                 configuration: { baseURL: config.baseURL },
                 modelName: config.model,
                 temperature: 0,
+                maxTokens: this.maxTokens,
             });
         }
-        // 绑定工具，使模型具备调用能力
         this.model = modelInstance.bindTools(this.langchainTools);
         return this.model;
     }
     async askForConfig() {
+        if (this.apiConfig)
+            return this.apiConfig;
         let config = {};
         if (fs_1.default.existsSync(exports.CONFIG_FILE)) {
             try {
@@ -139,17 +255,19 @@ class McpGraphAgent {
         return config;
     }
     async chat(query = "开始代码审计") {
+        await this.ensureInitialized();
         await this.getModel();
         const app = await this.createGraph();
         const stream = await app.stream({
             messages: [new messages_1.HumanMessage(query)],
             mode: "auto",
-            targetCount: 4,
+            targetCount: this.maxTargetCount,
         }, { configurable: { thread_id: "auto_worker" }, recursionLimit: this.recursionLimit, debug: this.verbose, });
         for await (const output of stream)
             this.renderOutput(output);
     }
     async start() {
+        await this.ensureInitialized();
         await this.getModel();
         const app = await this.createGraph();
         const rl = readline_1.default.createInterface({ input: process.stdin, output: process.stdout });
@@ -165,7 +283,7 @@ class McpGraphAgent {
                     rl.close();
                     return;
                 }
-                const stream = await app.stream({ messages: [new messages_1.HumanMessage(input)], mode: "chat" }, { configurable: { thread_id: "session" }, recursionLimit: 50 });
+                const stream = await app.stream({ messages: [new messages_1.HumanMessage(input)], mode: "chat" }, { configurable: { thread_id: "session" }, recursionLimit: this.recursionLimit, debug: this.verbose, });
                 for await (const output of stream)
                     this.renderOutput(output);
                 ask();
@@ -252,9 +370,8 @@ class McpGraphAgent {
 2. 严禁对 JSON 内容进行二次转义。
 
 # 指令
-1. 一旦完成对一个文件的修改（apply_fix），请立即调用 generate_review 总结该文件的变动。
-2. 避免陷入在同一个文件上的无限循环尝试。
-3. 不要重复调用相同的工具和参数，如果工具已经返回结果，请基于结果继续工作而不是再次调用。{recentToolCalls}
+1. 避免陷入在同一个文件上的无限循环尝试。
+2. 不要重复调用相同的工具和参数，如果工具已经返回结果，请基于结果继续工作而不是再次调用。{recentToolCalls}
 {extraPrompt}`;
         // 2. 核心逻辑：处理消息上下文
         let inputMessages;
@@ -352,6 +469,13 @@ class McpGraphAgent {
             const messages = state.messages;
             const lastMsg = messages[messages.length - 1];
             const content = lastMsg.content || "";
+            // 🛑 新增：全局 Token 熔断保护
+            // 如果已消耗 Token 超过了 options 中设置的 maxTokens (假设是总限额)
+            if (this.options.maxTokens && state.tokenUsage.total >= this.options.maxTokens) {
+                console.warn("⚠️ [警告] 已达到最大 Token 限制，强制结束任务。");
+                this.printFinalSummary(state);
+                return langgraph_1.END;
+            }
             // 1. 如果 AI 想要调用工具，去 tools 节点
             if (lastMsg.tool_calls && lastMsg.tool_calls.length > 0) {
                 return "tools";
@@ -360,7 +484,7 @@ class McpGraphAgent {
             // - 模式是 auto 且审计完成
             // - 或者 AI 明确输出了结束语
             // - 或者 AI 输出了普通内容且没有工具调用（针对问答模式）
-            const isAutoFinished = state.mode === "auto" && state.auditedFiles.length >= state.targetCount;
+            const isAutoFinished = state.mode === "auto" && state.auditedFiles.length > state.targetCount;
             const isFinalAnswer = content.includes("Final Answer");
             // ✅ 修复核心：如果 AI 只是在聊天（没有工具调用），直接结束，不要跳回 agent
             if (isAutoFinished || isFinalAnswer || state.mode === "chat") {

package/lib/core/agent.d.ts

@@ -19,6 +19,7 @@ export default class McpAgent {
     private ensureApiConfig;
     private loadMcpConfigs;
     init(): Promise<void>;
+    private getToolsForOpenAIAPI;
     private processChat;
     /**
    * 裁剪上下文消息列表

package/lib/core/agent.js

@@ -46,6 +46,7 @@ const index_js_1 = require("@modelcontextprotocol/sdk/client/index.js");
 const stdio_js_1 = require("@modelcontextprotocol/sdk/client/stdio.js");
 const config_1 = require("../config/config");
 const builtin_1 = require("../tools/builtin");
+const zod_to_json_schema_1 = require("zod-to-json-schema");
 const jsonSchemaToZod_1 = require("../utils/jsonSchemaToZod");
 class McpAgent {
     constructor(options) {
@@ -215,6 +216,25 @@ class McpAgent {
             }
         }
     }
+    getToolsForOpenAIAPI() {
+        return this.allTools.map((tool) => {
+            const { _handler, _client, _originalName, ...rest } = tool;
+            let parameters = rest.function.parameters;
+            // 💡 核心逻辑：判断是否为 Zod 实例并转换
+            // Zod 对象通常包含 _def 属性，或者你可以用 instanceof z.ZodType
+            if (parameters && typeof parameters === 'object' && ('_def' in parameters || parameters.safeParse)) {
+                // 使用 zod-to-json-schema 转换为标准 JSON Schema
+                parameters = (0, zod_to_json_schema_1.zodToJsonSchema)(parameters);
+            }
+            return {
+                ...rest,
+                function: {
+                    ...rest.function,
+                    parameters: parameters, // 确保这里是 Plain Object
+                },
+            };
+        });
+    }
     async processChat(userInput) {
         var _a;
         this.messages.push({ role: 'user', content: userInput });
@@ -237,7 +257,7 @@ class McpAgent {
                 response = await this.openai.chat.completions.create({
                     model: this.modelName,
                     messages: this.messages,
-                    tools: this.allTools.map(({ _handler, _client, _originalName, ...rest }) => rest),
+                    tools: this.getToolsForOpenAIAPI(),
                     tool_choice: 'auto'
                 });
             }
@@ -262,7 +282,9 @@ class McpAgent {
                 if (args.filePath) {
                     console.log(`   📂 正在查看文件: ${args.filePath}`);
                 }
-                console.log(`   🛠️  执行: ${call.function.name}`);
+                // 2. 打印操作信息和参数
+                console.log(`\n   🛠️  执行工具: \x1b[36m${call.function.name}\x1b[0m`);
+                console.log(`   📦 传入参数: \x1b[2m${JSON.stringify(args)}\x1b[0m`);
                 let result;
                 if (tool === null || tool === void 0 ? void 0 : tool._handler) {
                     result = await tool._handler(args);

package/lib/index.d.ts

@@ -1,7 +1,5 @@
 export * from './core/agent';
-export { default as McpChainAgent } from './core/agent-chain';
 export { default } from './core/agent';
 export { createTool } from './utils/createTool';
-export { AgentChainModel } from './model/AgentChainModel';
 export { AgentGraphModel, AgentGraphLLMResponse } from './model/AgentGraphModel';
 export { default as McpGraphAgent } from './core/agent-graph';

package/lib/index.js

@@ -17,16 +17,12 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
     return (mod && mod.__esModule) ? mod : { "default": mod };
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.McpGraphAgent = exports.AgentGraphModel = exports.AgentChainModel = exports.createTool = exports.default = exports.McpChainAgent = void 0;
+exports.McpGraphAgent = exports.AgentGraphModel = exports.createTool = exports.default = void 0;
 __exportStar(require("./core/agent"), exports);
-var agent_chain_1 = require("./core/agent-chain");
-Object.defineProperty(exports, "McpChainAgent", { enumerable: true, get: function () { return __importDefault(agent_chain_1).default; } });
 var agent_1 = require("./core/agent");
 Object.defineProperty(exports, "default", { enumerable: true, get: function () { return __importDefault(agent_1).default; } });
 var createTool_1 = require("./utils/createTool");
 Object.defineProperty(exports, "createTool", { enumerable: true, get: function () { return createTool_1.createTool; } });
-var AgentChainModel_1 = require("./model/AgentChainModel");
-Object.defineProperty(exports, "AgentChainModel", { enumerable: true, get: function () { return AgentChainModel_1.AgentChainModel; } });
 var AgentGraphModel_1 = require("./model/AgentGraphModel");
 Object.defineProperty(exports, "AgentGraphModel", { enumerable: true, get: function () { return AgentGraphModel_1.AgentGraphModel; } });
 var agent_graph_1 = require("./core/agent-graph");

package/lib/types/type.d.ts

@@ -1,5 +1,6 @@
 import { Client } from '@modelcontextprotocol/sdk/client/index';
 import { z } from 'zod';
+import { AgentGraphModel } from '../model/AgentGraphModel';
 export interface ApiConfig {
     baseURL: string;
     apiKey: string;
@@ -30,41 +31,12 @@ export interface AgentOptions {
     tools?: ToolInfo[];
     extraSystemPrompt?: any;
     maxTokens?: number;
-    /**
-     * only for chain agent
-     */
-    apiConfig?: ApiConfig;
-    /**
-     * only for chain agent
-     * extends BaseChatModel
-     */
-    apiModel?: any;
-    /**
-     * only for chain agent
-     */
-    maxIterations?: number;
-    /**
-     * only for chain agent
-     */
     verbose?: boolean;
-    /**
-     * only for graph agent
-     */
-    baseURL?: string;
-    /**
-     * only for graph agent
-     */
-    apiKey?: string;
-    /**
-     * only for graph agent
-     */
-    modelName?: string;
-    /**
-     * only for graph agent
-     */
+    apiConfig?: ApiConfig;
+}
+export interface GraphAgentOptions extends AgentOptions {
+    apiModel?: AgentGraphModel;
     alwaysSystem?: boolean;
-    /**
-     * only for graph agent
-     */
     recursionLimit?: number;
+    maxTargetCount?: number;
 }

package/lib/utils/createTool.d.ts

@@ -4,7 +4,7 @@ export interface CreateToolOptions {
     name: string;
     description: string;
     /**
-     * zod@3.23.8
+     * zod@3.25
      */
     parameters: z.ZodObject<any>;
     handler: (args: any) => Promise<string>;

package/lib/utils/jsonSchemaToZod.js

@@ -4,6 +4,14 @@ exports.jsonSchemaToZod = jsonSchemaToZod;
 const zod_1 = require("zod");
 function jsonSchemaToZod(parameters) {
     var _a;
+    // ✅ 跨版本/跨实例的特征检测
+    const isZod = parameters && (typeof parameters.safeParse === 'function' || // 所有 Zod 类型都有
+        !!parameters._def || // Zod 内部定义属性
+        !!parameters.shape // 专门针对 ZodObject 的特征
+    );
+    if (isZod) {
+        return parameters;
+    }
     if (!parameters || typeof parameters !== 'object') {
         return zod_1.z.object({}).passthrough();
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@saber2pr/ai-agent",
-  "version": "0.0.26",
+  "version": "0.0.28",
   "description": "AI Assistant CLI.",
   "author": "saber2pr",
   "license": "ISC",
@@ -9,7 +9,6 @@
   ],
   "bin": {
     "sagent": "./lib/cli.js",
-    "sagent-chain": "./lib/cli-chain.js",
     "sagent-graph": "./lib/cli-graph.js"
   },
   "publishConfig": {
@@ -20,6 +19,7 @@
   "scripts": {
     "test": "node lib/index",
     "start": "tsc --watch",
+    "build": "tsc",
     "prepublishOnly": "tsc"
   },
   "dependencies": {
@@ -34,9 +34,11 @@
     "langchain": "0.3.15",
     "minimatch": "^10.0.1",
     "openai": "^6.16.0",
-    "zod": "3.23.8"
+    "zod": "3.25",
+    "zod-to-json-schema": "3.23.2"
   },
   "resolutions": {
+    "zod": "3.25",
     "@langchain/core": "0.3.39"
   },
   "devDependencies": {

package/lib/cli-chain.d.ts

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

package/lib/cli-chain.js

@@ -1,9 +0,0 @@
-#!/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("./core/agent-chain"));
-const agent = new agent_chain_1.default();
-agent.start();

package/lib/core/agent-chain.d.ts

@@ -1,23 +0,0 @@
-import { AgentOptions } from '../types/type';
-export default class McpChainAgent {
-    private allTools;
-    private encoder;
-    private extraTools;
-    private maxTokens;
-    private executor?;
-    private apiConfig;
-    private maxIterations;
-    private apiModel?;
-    private memory?;
-    private systemPrompt;
-    private runningTokenCounter;
-    private verbose;
-    constructor(options?: AgentOptions);
-    private initTools;
-    private wrapHandler;
-    init(): Promise<void>;
-    chat(input: string): Promise<string>;
-    private showLoading;
-    start(): Promise<void>;
-    private ensureApiConfig;
-}

package/lib/core/agent-chain.js

@@ -1,246 +0,0 @@
-"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 js_tiktoken_1 = require("js-tiktoken");
-const agents_1 = require("langchain/agents");
-const memory_1 = require("langchain/memory");
-const readline = __importStar(require("readline"));
-const prompts_1 = require("@langchain/core/prompts");
-const tools_1 = require("@langchain/core/tools");
-const openai_1 = require("@langchain/openai");
-const config_1 = require("../config/config");
-const builtin_1 = require("../tools/builtin");
-class McpChainAgent {
-    constructor(options) {
-        this.allTools = [];
-        this.encoder = (0, js_tiktoken_1.getEncoding)("cl100k_base");
-        this.extraTools = [];
-        this.runningTokenCounter = 0;
-        this.extraTools = (options === null || options === void 0 ? void 0 : options.tools) || [];
-        this.maxTokens = (options === null || options === void 0 ? void 0 : options.maxTokens) || 8000;
-        this.apiConfig = options === null || options === void 0 ? void 0 : options.apiConfig;
-        this.maxIterations = (options === null || options === void 0 ? void 0 : options.maxIterations) || 50;
-        this.apiModel = options === null || options === void 0 ? void 0 : options.apiModel;
-        this.verbose = (options === null || options === void 0 ? void 0 : options.verbose) || false;
-        const baseSystemPrompt = `你是一个专业的代码架构师。
-你的目标是理解并分析用户项目，请务必遵循以下工作流：
-
-### 第一阶段：全景感知 (The "Where" Phase)
-1. **必须首先调用 'get_directory_tree'**：获取项目完整文件列表，包括样式文件 (.less, .css) 和资源文件。
-2. 结合目录结构，观察项目架构（如 Monorepo 结构或 src 布局）。
-
-### 第二阶段：逻辑映射 (The "What" Phase)
-1. **调用 'get_repo_map'**：针对代码文件提取导出定义，理解模块间的调用关系。
-2. 如果需要查看具体的样式定义，直接使用 'read_text_file' 读取 .less 或 .css 文件。
-
-### 核心原则：
-- 不要猜测文件是否存在，先看目录树。
-- 优先查看 Skeleton（骨架），只有需要修复逻辑时才读取完整 Text（全文）。
-- 始终以中文回答思考过程。`;
-        this.systemPrompt = (options === null || options === void 0 ? void 0 : options.extraSystemPrompt)
-            ? `${baseSystemPrompt}\n\n[额外指令]:\n${JSON.stringify(options.extraSystemPrompt)}`
-            : baseSystemPrompt;
-        this.initTools(options);
-    }
-    initTools(options) {
-        const allTools = [...(0, builtin_1.createDefaultBuiltinTools)({ options: { ...options, ...this } }), ...this.extraTools];
-        this.allTools = allTools.map((t) => ({
-            ...t,
-            _handler: this.wrapHandler(t.function.name, t._handler),
-        }));
-    }
-    wrapHandler(name, handler) {
-        return async (args) => {
-            console.log(`\n   [工具调用]: ${name}`);
-            const result = await handler(args);
-            const content = typeof result === "string" ? result : JSON.stringify(result);
-            this.runningTokenCounter += this.encoder.encode(content).length;
-            return content;
-        };
-    }
-    async init() {
-        if (this.executor)
-            return;
-        let model;
-        if (this.apiModel) {
-            model = this.apiModel;
-        }
-        else {
-            const apiConfig = await this.ensureApiConfig();
-            model = new openai_1.ChatOpenAI({
-                configuration: { baseURL: apiConfig.baseURL, apiKey: apiConfig.apiKey },
-                modelName: apiConfig.model,
-                temperature: 0,
-            });
-        }
-        // 1. 初始化 SummaryBufferMemory
-        // maxTokenLimit 决定了当对话历史超过多少 Token 时触发“自动总结”
-        this.memory = new memory_1.ConversationSummaryBufferMemory({
-            llm: model,
-            maxTokenLimit: this.maxTokens,
-            memoryKey: "chat_history",
-            returnMessages: true,
-            // 必须添加下面这两行显式声明：
-            inputKey: "input", // 对应 invoke 里的 input 字段
-            outputKey: "output", // 对应 Agent 输出的字段
-            chatHistory: new memory_1.ChatMessageHistory(), // 👈 显式指定 history 实例
-        });
-        const langchainTools = this.allTools.map(t => new tools_1.DynamicStructuredTool({
-            name: t.function.name,
-            description: t.function.description || "",
-            schema: (t.function.parameters),
-            func: async (args) => await t._handler(args),
-        }));
-        const prompt = prompts_1.PromptTemplate.fromTemplate(`{system_prompt}
-
-### 🛠 可用工具：
-{tools}
-
-### 🛠 工具名称：
-[{tool_names}]
-
-### 📝 历史记录：
-{chat_history}
-
-### ⚠️ 回复规范（严格遵守）：
-1. 首先输出 **Thought:**，用中文详细说明你的分析思路。
-2. 接着输出一个 **JSON Action 代码块**。
-
-示例格式：
-Thought: 正在查看目录。
-\`\`\`json
-{{
-  "action": "directory_tree",
-  "action_input": {{}}
-}}
-\`\`\`
-
-Begin!
-Question: {input}
-{agent_scratchpad}`); // 👈 强制以 Thought: 开头，解决断更问题
-        const agent = await (0, agents_1.createStructuredChatAgent)({ llm: model, tools: langchainTools, prompt });
-        this.executor = new agents_1.AgentExecutor({
-            agent,
-            tools: langchainTools,
-            memory: this.memory, // 挂载内存模块
-            verbose: this.verbose,
-            maxIterations: this.maxIterations,
-            handleParsingErrors: (e) => {
-                // 简化报错，不要再给 AI 错误的 JSON 示例
-                return `格式不正确。请确保你输出了一个正确的 Markdown JSON 代码块，例如：\n\`\`\`json\n{ "action": "...", "action_input": { ... } }\n\`\`\``;
-            },
-        });
-    }
-    async chat(input) {
-        if (!this.executor)
-            await this.init();
-        this.runningTokenCounter = this.encoder.encode(input).length;
-        const stopLoading = this.showLoading("🤖 Agent 正在思考并管理上下文...");
-        try {
-            // 执行请求，AgentExecutor 会自动：
-            // 1. 从 memory 加载历史 (chat_history)
-            // 2. 将这次对话的结果 saveContext 到 memory
-            const response = await this.executor.invoke({
-                input: input,
-                system_prompt: this.systemPrompt,
-            }, {
-                callbacks: [{
-                        handleAgentAction: (action) => {
-                            const rawLog = action.log || "";
-                            // 兼容 Thought: Thought: [内容]
-                            let thought = rawLog.split(/```json|\{/)[0]
-                                .replace(/Thought:/gi, "") // 全局替换掉所有 Thought: 标签
-                                .trim();
-                            if (thought) {
-                                console.log(`\n💭 [思考]: ${thought}`);
-                            }
-                        }
-                    }]
-            });
-            return typeof response.output === 'string' ? response.output : JSON.stringify(response.output);
-        }
-        finally {
-            stopLoading();
-        }
-    }
-    showLoading(text) {
-        const chars = ['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
-        let i = 0;
-        const timer = setInterval(() => {
-            process.stdout.write(`\r${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 (Summary+Window 模式) 已启动`);
-        const chatLoop = () => {
-            rl.question("\n👤 你: ", async (input) => {
-                if (!input.trim())
-                    return chatLoop();
-                if (input.toLowerCase() === "exit")
-                    process.exit(0);
-                try {
-                    const res = await this.chat(input);
-                    console.log(`\n🤖 Agent: ${res}`);
-                }
-                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_1.CONFIG_FILE))
-            return JSON.parse(fs_1.default.readFileSync(config_1.CONFIG_FILE, "utf-8"));
-        const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
-        const question = (q) => new Promise((res) => rl.question(q, res));
-        const config = { baseURL: await question("? API Base URL: "), apiKey: await question("? API Key: "), model: await question("? Model Name: ") };
-        fs_1.default.writeFileSync(config_1.CONFIG_FILE, JSON.stringify(config, null, 2));
-        rl.close();
-        return config;
-    }
-}
-exports.default = McpChainAgent;

package/lib/model/AgentChainModel.d.ts

@@ -1,18 +0,0 @@
-import { BaseChatModel } from '@langchain/core/language_models/chat_models';
-import { AIMessage, MessageFieldWithRole } from '@langchain/core/messages';
-interface AgentChainModelImpl {
-    generateResponse: (messages: MessageFieldWithRole[]) => Promise<string>;
-}
-export declare abstract class AgentChainModel extends BaseChatModel implements AgentChainModelImpl {
-    bind(args: any): any;
-    constructor(fields?: any);
-    generateResponse(messages: MessageFieldWithRole[]): Promise<string>;
-    _generate(messages: any): Promise<{
-        generations: {
-            text: string;
-            message: AIMessage;
-        }[];
-    }>;
-    _llmType(): string;
-}
-export {};

package/lib/model/AgentChainModel.js

@@ -1,23 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.AgentChainModel = void 0;
-const chat_models_1 = require("@langchain/core/language_models/chat_models");
-const messages_1 = require("@langchain/core/messages");
-class AgentChainModel extends chat_models_1.BaseChatModel {
-    bind(args) {
-        // 逻辑上调用基类（即使基类在类型上说没有，运行时通常是有的）
-        // 如果运行时也没有，这里就返回 this 本身
-        // @ts-ignore
-        return (super.bind ? super.bind(args) : this);
-    }
-    constructor(fields) { super(fields || {}); }
-    async generateResponse(messages) {
-        return '';
-    }
-    async _generate(messages) {
-        let text = await this.generateResponse(messages);
-        return { generations: [{ text, message: new messages_1.AIMessage(text) }] };
-    }
-    _llmType() { return "my_private_llm"; }
-}
-exports.AgentChainModel = AgentChainModel;