zhitalk 0.1.1 → 0.1.3

package/CLAUDE.md

@@ -1,38 +1,57 @@
 # CLAUDE.md
 
-This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+This file provides guidance to Claude Code when working with code in this repository.
 
 ## Project Overview
 
-Zhitalk is a personal AI Agent (like OpenClaw) running in the terminal. It uses LangGraph for agent orchestration, Moonshot Kimi as the LLM, and Commander.js for CLI commands.
+Zhitalk is a terminal AI Agent using LangGraph for agent orchestration, OpenAI-compatible APIs (default Moonshot Kimi), and Commander.js. Node.js >=22 is required.
 
 ## Commands
 
 ```bash
-pnpm dev        # Run with ts-node (development)
-pnpm test       # Run tests
-pnpm test:watch # Run tests in watch mode
-pnpm test:coverage # Run tests with coverage
-pnpm build      # Compile TypeScript, copy skills, fix permissions
-pnpm start      # Run compiled output from dist/
-pnpm clean      # Remove dist directory
+pnpm dev        # ts-node/esm development
+pnpm test       # Jest with experimental-vm-modules
+pnpm test:watch # Watch mode
+pnpm test:coverage
+pnpm build      # Compile to dist/
+pnpm start      # Run dist/
+pnpm clean      # Remove dist/
 ```
 
-Run a single test with `pnpm test -- <pattern>`.
+Run a single test: `pnpm test -- <pattern>`.
 
 ## Architecture
 
-See [`src/agent/CLAUDE.md`](src/agent/CLAUDE.md) for details on the agent graph, tools, skills, memory, and session management.
+See [`src/agent/CLAUDE.md`](src/agent/CLAUDE.md) for the agent graph, tools, skills, memory, hooks, sub-agents, MCP, and session management.
+
+Entry point `src/index.ts` parses CLI args, runs `src/install.ts` for first-time setup, then launches `interactiveChat()` from `src/agent/cli.ts`.
+
+### CLI Modes
+
+- **Interactive** (default): `zhitalk` launches a REPL with streaming output, ESC cancellation, and slash commands (`/new`, `/rewind`, `/sessions`, `/compact`).
+- **Single-shot**: `zhitalk run <query> [options]` executes one turn and exits:
+  - `-a, --auto-approve` — auto-approve all tool calls (non-interactive)
+  - `-j, --json` — output JSON with `response`, `usage`, `thread_id`
+  - `-t, --thread-id <id>` — reuse a specific session
+
+### Agent APIs
+
+- `runAgentStream(userMessage, onToken, onToolConfirmation, threadId)` — interactive streaming with per-tool confirmation callbacks.
+- `runAgent(userMessage, threadId?, { autoApprove?, signal? })` — non-streaming, terminal-free API. Used by the `run` CLI command and by tests.
 
 ## TypeScript Config
 
-Module type is `commonjs`, output goes to `dist/`. The build script also copies `src/agent/skills/` to `dist/agent/skills/` and makes the CLI executable.
+ES modules (`"type": "module"`), output goes to `dist/`. The build script fixes CLI executable permissions.
 
 ## Testing
 
-Jest with ts-jest preset. Test files match `**/*.test.ts`. Run a single test with `pnpm test -- <pattern>`.
+Jest with ts-jest preset and `--experimental-vm-modules`. Test files match `**/*.test.ts`.
 
-## behavior rules
+- `FakeChatOpenAI` (`src/agent/testing/fake_model.ts`) replays scripted message sequences for testing the agent loop without calling real LLM APIs.
+- `createAgentGraph(nativeTools, customModel?)` accepts an optional model for testing.
+- Existing tests favor integration style: real temp files, real SQLite, mocked `fetch` for network tools.
+
+## Behavior Rules
 
 **Tradeoff:** These guidelines bias toward caution over speed. For trivial tasks, use judgment.
 
@@ -43,7 +62,7 @@ Jest with ts-jest preset. Test files match `**/*.test.ts`. Run a single test wit
 Before implementing:
 
 - State your assumptions explicitly. If uncertain, ask.
-- If multiple interpretations exist, present them - don't pick silently.
+- If multiple interpretations exist, present them — don't pick silently.
 - If a simpler approach exists, say so. Push back when warranted.
 - If something is unclear, stop. Name what's confusing. Ask.
 
@@ -57,8 +76,6 @@ Before implementing:
 - No error handling for impossible scenarios.
 - If you write 200 lines and it could be 50, rewrite it.
 
-Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.
-
 ### 3. Surgical Changes
 
 **Touch only what you must. Clean up only your own mess.**
@@ -68,12 +85,7 @@ When editing existing code:
 - Don't "improve" adjacent code, comments, or formatting.
 - Don't refactor things that aren't broken.
 - Match existing style, even if you'd do it differently.
-- If you notice unrelated dead code, mention it - don't delete it.
-
-When your changes create orphans:
-
-- Remove imports/variables/functions that YOUR changes made unused.
-- Don't remove pre-existing dead code unless asked.
+- When your changes create orphans, remove imports/variables/functions that YOUR changes made unused.
 
 The test: Every changed line should trace directly to the user's request.
 

package/README.md

@@ -42,7 +42,17 @@ zhitalk
 ### Command
 
 ```bash
+# 开发
 pnpm dev
 pnpm test
 pnpm build
+
+# 交互模式（默认）
+zhitalk
+
+# 单次命令模式（执行一次即退出）
+zhitalk run "帮我查一下北京天气"
+zhitalk run "帮我查一下北京天气" --auto-approve   # 自动批准 tools
+zhitalk run "帮我查一下北京天气" --json            # JSON 输出
+zhitalk run "继续" --thread-id xxx               # 指定会话 ID
 ```

package/dist/agent/agent.d.ts

@@ -1,4 +1,43 @@
-import { type UsageMetadata } from '@langchain/core/messages';
+import { type UsageMetadata, ToolMessage } from '@langchain/core/messages';
+import { CompiledStateGraph } from '@langchain/langgraph';
+import type { BaseMessage } from '@langchain/core/messages';
+import { tools } from './tools.js';
+export declare function createAgentGraph(toolList: typeof tools, customModel?: any): CompiledStateGraph<{
+    messages: BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[];
+    contextSummary: string | null;
+    compressionCount: number;
+    lastCompressedIndex: number;
+    todoList: import("./todo.js").TodoItem[] | null;
+}, {
+    messages?: BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[] | import("@langchain/langgraph").OverwriteValue<BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[]> | undefined;
+    contextSummary?: string | import("@langchain/langgraph").OverwriteValue<string | null> | null | undefined;
+    compressionCount?: number | import("@langchain/langgraph").OverwriteValue<number> | undefined;
+    lastCompressedIndex?: number | import("@langchain/langgraph").OverwriteValue<number> | undefined;
+    todoList?: import("./todo.js").TodoItem[] | import("@langchain/langgraph").OverwriteValue<import("./todo.js").TodoItem[] | null> | null | undefined;
+}, "tools" | "__start__" | "model_request", {
+    messages: import("@langchain/langgraph").BaseChannel<BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[], BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[] | import("@langchain/langgraph").OverwriteValue<BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[]>, unknown>;
+    contextSummary: import("@langchain/langgraph").BaseChannel<string | null, string | import("@langchain/langgraph").OverwriteValue<string | null> | null, unknown>;
+    compressionCount: import("@langchain/langgraph").BaseChannel<number, number | import("@langchain/langgraph").OverwriteValue<number>, unknown>;
+    lastCompressedIndex: import("@langchain/langgraph").BaseChannel<number, number | import("@langchain/langgraph").OverwriteValue<number>, unknown>;
+    todoList: import("@langchain/langgraph").BaseChannel<import("./todo.js").TodoItem[] | null, import("./todo.js").TodoItem[] | import("@langchain/langgraph").OverwriteValue<import("./todo.js").TodoItem[] | null> | null, unknown>;
+}, {
+    messages: import("@langchain/langgraph").BaseChannel<BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[], BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[] | import("@langchain/langgraph").OverwriteValue<BaseMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>, import("@langchain/core/messages").MessageType>[]>, unknown>;
+    contextSummary: import("@langchain/langgraph").BaseChannel<string | null, string | import("@langchain/langgraph").OverwriteValue<string | null> | null, unknown>;
+    compressionCount: import("@langchain/langgraph").BaseChannel<number, number | import("@langchain/langgraph").OverwriteValue<number>, unknown>;
+    lastCompressedIndex: import("@langchain/langgraph").BaseChannel<number, number | import("@langchain/langgraph").OverwriteValue<number>, unknown>;
+    todoList: import("@langchain/langgraph").BaseChannel<import("./todo.js").TodoItem[] | null, import("./todo.js").TodoItem[] | import("@langchain/langgraph").OverwriteValue<import("./todo.js").TodoItem[] | null> | null, unknown>;
+}, import("@langchain/langgraph").StateDefinition, {
+    model_request: {
+        messages: any[];
+    };
+    tools: {
+        messages: never[];
+        todoList?: undefined;
+    } | {
+        messages: ToolMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>>[];
+        todoList: import("./todo.js").TodoItem[] | null;
+    };
+}, unknown, unknown, []>;
 export declare function initAgent(): Promise<void>;
 /**
  * 以流式方式运行 agent，将 token 逐个回调给调用方
@@ -12,6 +51,22 @@ export declare function runAgentStream(userMessage: string, onToken: (token: str
     response: string;
     usageMetadata?: UsageMetadata;
 }>;
+/**
+ * 以非流式方式运行 agent，无需终端回调，适合脚本和测试
+ * @param {string} userMessage - 用户输入
+ * @param {string} threadId - 会话 ID
+ * @param {Object} options - 选项
+ * @param {boolean} options.autoApprove - 是否自动批准所有 tool 调用
+ * @param {AbortSignal} options.signal - 取消信号
+ * @returns {Promise<{ response: string; usageMetadata?: UsageMetadata }>}
+ */
+export declare function runAgent(userMessage: string, threadId?: string, options?: {
+    autoApprove?: boolean;
+    signal?: AbortSignal;
+}): Promise<{
+    response: string;
+    usageMetadata?: UsageMetadata;
+}>;
 /**
  * 启动一个 subagent 执行独立任务，完成后返回结果
  * @param {string} prompt - 给 subagent 的任务提示

package/dist/agent/agent.js

@@ -14,8 +14,6 @@ import { checkExecPermission } from './permission/exec.js';
 import { checkNetworkPermission } from './permission/network.js';
 import { runPreToolUseHooks, runPostToolUseHooks } from './hooks.js';
 import { createModel } from './model.js';
-// ── 模型 ──────────────────────────────────────────────────
-const model = createModel({ streaming: true });
 // ── State Schema ──────────────────────────────────────────
 const StateAnnotation = Annotation.Root({
     messages: Annotation({
@@ -100,8 +98,8 @@ function ensureCompleteToolCalls(messages) {
 // ── 记忆 ──────────────────────────────────────────────────
 const checkpointer = SqliteSaver.fromConnString(DB_PATH);
 // ── Agent Graph 工厂 ──────────────────────────────────────
-function createAgentGraph(toolList) {
-    const modelWithTheseTools = model.bindTools(toolList);
+export function createAgentGraph(toolList, customModel) {
+    const modelWithTheseTools = (customModel || createModel({ streaming: true })).bindTools(toolList);
     async function modelRequest(state, config) {
         let modelMessages = state.messages ?? [];
         // compact context
@@ -269,6 +267,10 @@ function createAgentGraph(toolList) {
         if (confirmCalls.length === 0) {
             return { messages: [...allowOutputs, ...blockMessages], todoList: todoResult.todoList };
         }
+        if (config?.configurable?.auto_approve) {
+            const confirmOutputs = await executeCalls(confirmCalls);
+            return { messages: [...allowOutputs, ...blockMessages, ...confirmOutputs], todoList: todoResult.todoList };
+        }
         const result = interrupt({ toolCalls: confirmCalls });
         if (result !== 'approved') {
             const deniedMessages = confirmCalls.map((call) => new ToolMessage({
@@ -316,8 +318,8 @@ function getSubAgent() {
     return subAgent;
 }
 // ── 核心运行逻辑 ──────────────────────────────────────────
-async function _runAgent(compiledAgent, userMessage, onToken, onToolConfirmation, threadId, signal) {
-    const config = { configurable: { thread_id: threadId } };
+async function _runAgent(compiledAgent, userMessage, onToken, onToolConfirmation, threadId, signal, options) {
+    const config = { configurable: { thread_id: threadId, auto_approve: options?.autoApprove } };
     let fullResponse = '';
     let usageMetadata;
     let input = { messages: [new HumanMessage(userMessage)] };
@@ -379,6 +381,18 @@ async function _runAgent(compiledAgent, userMessage, onToken, onToolConfirmation
 export async function runAgentStream(userMessage, onToken, onToolConfirmation, threadId = 'default-session', signal) {
     return _runAgent(getAgent(), userMessage, onToken, onToolConfirmation, threadId, signal);
 }
+/**
+ * 以非流式方式运行 agent，无需终端回调，适合脚本和测试
+ * @param {string} userMessage - 用户输入
+ * @param {string} threadId - 会话 ID
+ * @param {Object} options - 选项
+ * @param {boolean} options.autoApprove - 是否自动批准所有 tool 调用
+ * @param {AbortSignal} options.signal - 取消信号
+ * @returns {Promise<{ response: string; usageMetadata?: UsageMetadata }>}
+ */
+export async function runAgent(userMessage, threadId = 'default-session', options) {
+    return _runAgent(getAgent(), userMessage, () => { }, async () => true, threadId, options?.signal, options);
+}
 /**
  * 启动一个 subagent 执行独立任务，完成后返回结果
  * @param {string} prompt - 给 subagent 的任务提示

package/dist/agent/context.js

@@ -18,8 +18,10 @@ const MODEL_CONTEXT_LIMITS = {
     'deepseek-r1': 65536,
     // MiniMax
     'minimax-text-01': 1048576,
-    'minimax-m1': 1048576,
-    'minimax-m3': 1048576,
+    'minimax-m2': 204800,
+    'minimax-m2.7': 204800,
+    'minimax-m2.5': 204800,
+    'minimax-m3': 1000000,
     // GLM (Zhipu)
     'glm-4': 131072,
     'glm-4-plus': 131072,

package/dist/agent/prompt.js

@@ -3,13 +3,10 @@ import path from 'path';
 import { discoverSkills, getSkillsListText } from './skills.js';
 import { WORKSPACE_DIR } from './config.js';
 import { listRecentMemories } from './db.js';
-// ── Skills ────────────────────────────────────────────────
-discoverSkills();
-const skillsText = getSkillsListText();
 const basePrompt = 'You are zhitalk, 中文名字叫智语, a personal AI Agent like OpenClaw, including tools, skills, memory, hook, sub-agent, MCP server, etc. Run in the terminal.';
-function readProfile() {
+function readProfile(profilePath) {
     try {
-        const filePath = path.join(WORKSPACE_DIR, '.data', 'profile.md');
+        const filePath = profilePath || path.join(WORKSPACE_DIR, '.data', 'profile.md');
         const content = fs.readFileSync(filePath, 'utf-8').trim();
         return `<profile_info>${content}</profile_info>`;
     }
@@ -17,7 +14,8 @@ function readProfile() {
         return '<profile_info></profile_info>';
     }
 }
-const profilePrompt = `## User Profile
+function getProfilePrompt() {
+    return `## User Profile
 
 When you learn information about the user that fits the following categories, store it as profile information for future personalization.
 
@@ -32,22 +30,28 @@ When you learn information about the user that fits the following categories, st
 
 Here is the user's current profile information:
 ${readProfile()}`;
+}
 const memoryPrompt = `## Memory Management Rules
 
 - When deleting a memory, first use memory_retrieve to find its id, then call memory_delete with that id. If no matching memory is found, politely inform the user.
 - When updating a memory, first delete the old memory, then create a new one.
 - When information fits the categories in <profile_template>, do not record it as a memory. It will be stored in the profile file instead.`;
-const dateTimePrompt = `## Current DateTime
+function getDateTimePrompt() {
+    return `## Current DateTime
 
 ${new Date().toLocaleString('zh-CN', { timeZone: 'Asia/Shanghai' })}`;
-const skillList = skillsText
-    ? `You have access to the following skills. When a user's request matches a skill's description, you MUST call the \`load_skill\` tool to load that skill's full instructions, then follow them.
+}
+function getSkillPrompt() {
+    const skillsText = getSkillsListText();
+    const skillList = skillsText
+        ? `You have access to the following skills. When a user's request matches a skill's description, you MUST call the \`load_skill\` tool to load that skill's full instructions, then follow them.
 
 ${skillsText}`
-    : 'There are no skills available.';
-const skillPrompt = `## Skills
+        : 'There are no skills available.';
+    return `## Skills
 
 ${skillList}\n\nNote: If you want to add a new skill, it must be placed in the ${path.join(WORKSPACE_DIR, 'skills')} directory.`;
+}
 const todoPrompt = `## Task Planning
 
 When a user's request is a complex, multi-step task (such as data analysis, writing a document, planning a task, or any task clearly requiring more than 3 steps), you MUST:
@@ -62,6 +66,7 @@ export function invalidateMemoryCache() {
     cachedMemories = null;
 }
 export function buildSystemPrompt() {
+    discoverSkills();
     if (!cachedMemories) {
         cachedMemories = listRecentMemories(10);
     }
@@ -70,11 +75,11 @@ export function buildSystemPrompt() {
         : '';
     return [
         basePrompt,
-        profilePrompt,
+        getProfilePrompt(),
         memorySection,
         memoryPrompt,
-        skillPrompt,
-        dateTimePrompt,
+        getSkillPrompt(),
+        getDateTimePrompt(),
         todoPrompt,
     ]
         .filter(Boolean)

package/dist/agent/skills.d.ts

@@ -3,6 +3,7 @@ export interface SkillInfo {
     description: string;
     dirPath: string;
 }
-export declare function discoverSkills(): SkillInfo[];
+export declare function discoverSkills(searchPaths?: string[]): SkillInfo[];
 export declare function loadSkill(name: string): string | null;
 export declare function getSkillsListText(): string;
+export declare function clearSkillsCache(): void;

package/dist/agent/skills.js

@@ -24,10 +24,10 @@ function parseFrontmatter(content) {
     }
     return { name: result.name || '', description: result.description || '' };
 }
-export function discoverSkills() {
+export function discoverSkills(searchPaths) {
     skills.length = 0;
     skillContentMap.clear();
-    const skillDirs = [
+    const skillDirs = searchPaths || [
         join(WORKSPACE_DIR, 'skills'),
         join(process.cwd(), '.agents', 'skills'),
     ];
@@ -81,3 +81,7 @@ export function getSkillsListText() {
     // console.log('skills...\n', lines.join('\n'))
     return lines.join('\n');
 }
+export function clearSkillsCache() {
+    skills.length = 0;
+    skillContentMap.clear();
+}

package/dist/agent/testing/fake_model.d.ts

@@ -0,0 +1,26 @@
+import { AIMessage, type BaseMessage } from '@langchain/core/messages';
+/**
+ * Fake LLM for testing the agent loop without calling real APIs.
+ * Replays a scripted sequence of messages on each invoke() call.
+ */
+export declare class FakeChatOpenAI {
+    private responses;
+    private callIndex;
+    constructor(responses: BaseMessage[]);
+    bindTools(_tools: any[]): this;
+    invoke(_messages: BaseMessage[], _config?: any): Promise<BaseMessage>;
+    stream(_messages: BaseMessage[], _config?: any): AsyncGenerator<BaseMessage>;
+    getCallCount(): number;
+}
+/**
+ * Helper to build an AIMessage with tool_calls.
+ */
+export declare function aiMessageWithTools(toolCalls: Array<{
+    name: string;
+    args: Record<string, any>;
+    id?: string;
+}>): AIMessage;
+/**
+ * Helper to build a simple text AIMessage.
+ */
+export declare function aiMessage(content: string): AIMessage;

package/dist/agent/testing/fake_model.js

@@ -0,0 +1,48 @@
+import { AIMessage, } from '@langchain/core/messages';
+/**
+ * Fake LLM for testing the agent loop without calling real APIs.
+ * Replays a scripted sequence of messages on each invoke() call.
+ */
+export class FakeChatOpenAI {
+    constructor(responses) {
+        this.callIndex = 0;
+        this.responses = responses;
+    }
+    bindTools(_tools) {
+        return this;
+    }
+    async invoke(_messages, _config) {
+        const response = this.responses[this.callIndex];
+        if (!response) {
+            throw new Error(`FakeChatOpenAI: no more responses (call ${this.callIndex})`);
+        }
+        this.callIndex++;
+        return response;
+    }
+    async *stream(_messages, _config) {
+        const response = await this.invoke(_messages, _config);
+        yield response;
+    }
+    getCallCount() {
+        return this.callIndex;
+    }
+}
+/**
+ * Helper to build an AIMessage with tool_calls.
+ */
+export function aiMessageWithTools(toolCalls) {
+    return new AIMessage({
+        content: '',
+        tool_calls: toolCalls.map((tc, i) => ({
+            name: tc.name,
+            args: tc.args,
+            id: tc.id || `fake-call-${i}`,
+        })),
+    });
+}
+/**
+ * Helper to build a simple text AIMessage.
+ */
+export function aiMessage(content) {
+    return new AIMessage({ content });
+}

package/dist/index.d.ts

@@ -1,2 +1,10 @@
 #!/usr/bin/env node
-export {};
+export interface RunOptions {
+    autoApprove?: boolean;
+    threadId?: string;
+}
+export declare function runOneShot(query: string, options?: RunOptions): Promise<{
+    response: string;
+    usageMetadata?: any;
+    threadId: string;
+}>;

package/dist/index.js

@@ -1,11 +1,41 @@
 #!/usr/bin/env node
 import { Command } from 'commander';
-import { readFileSync, existsSync } from 'fs';
+import { readFileSync, existsSync, realpathSync } from 'fs';
 import { join, dirname } from 'path';
-import { fileURLToPath } from 'url';
+import { fileURLToPath, pathToFileURL } from 'url';
 import { CONFIG_PATH } from './agent/config.js';
 const __dirname = dirname(fileURLToPath(import.meta.url));
 const pkg = JSON.parse(readFileSync(join(__dirname, '../package.json'), 'utf-8'));
+async function initApp() {
+    const { initColors } = await import('./agent/colors.js');
+    const { initDb } = await import('./agent/db.js');
+    const { initAgent } = await import('./agent/agent.js');
+    const { checkModel } = await import('./agent/model.js');
+    const { shutdownMcp } = await import('./agent/mcp/index.js');
+    await initColors();
+    await checkModel();
+    initDb();
+    await initAgent();
+    return {
+        shutdown: async () => {
+            await shutdownMcp();
+        },
+    };
+}
+export async function runOneShot(query, options = {}) {
+    const threadId = options.threadId || `oneshot-${Date.now()}`;
+    const { shutdown } = await initApp();
+    try {
+        const { runAgent } = await import('./agent/agent.js');
+        const result = await runAgent(query, threadId, {
+            autoApprove: options.autoApprove,
+        });
+        return { response: result.response, usageMetadata: result.usageMetadata, threadId };
+    }
+    finally {
+        await shutdown();
+    }
+}
 async function main() {
     const program = new Command();
     program.name(pkg.name).description(pkg.description).version(pkg.version);
@@ -16,30 +46,70 @@ async function main() {
         console.log(`配置文件位置: ${CONFIG_PATH}`);
         console.log('请使用编辑器打开并修改配置，修改完成后重启 zhitalk');
     });
-    if (process.argv.length > 2) {
-        program.parse();
-        return;
+    program
+        .command('run <query>')
+        .description('Run a single query and exit')
+        .option('-a, --auto-approve', 'Auto approve all tool calls without confirmation', false)
+        .option('-j, --json', 'Output result as JSON', false)
+        .option('-t, --thread-id <id>', 'Thread ID for the session')
+        .action(async (query, options) => {
+        if (!existsSync(CONFIG_PATH)) {
+            console.error(`找不到配置文件 ${CONFIG_PATH}`);
+            console.error('请先运行 zhitalk 进行初始化');
+            process.exit(1);
+        }
+        try {
+            const result = await runOneShot(query, {
+                autoApprove: options.autoApprove,
+                threadId: options.threadId,
+            });
+            if (options.json) {
+                console.log(JSON.stringify({
+                    response: result.response,
+                    usage: result.usageMetadata || null,
+                    thread_id: result.threadId,
+                }, null, 2));
+            }
+            else {
+                console.log(result.response);
+            }
+        }
+        catch (err) {
+            console.error(`Error: ${err.message}`);
+            process.exitCode = 1;
+        }
+        finally {
+            process.exit();
+        }
+    });
+    program.action(async () => {
+        if (!existsSync(CONFIG_PATH)) {
+            const { runInstall } = await import('./install.js');
+            await runInstall();
+            return;
+        }
+        const { interactiveChat } = await import('./agent/cli.js');
+        const { shutdownMcp } = await import('./agent/mcp/index.js');
+        const { shutdown } = await initApp();
+        process.once('SIGINT', async () => {
+            await shutdown();
+            process.exit(0);
+        });
+        await interactiveChat();
+        await shutdown();
+    });
+    program.parse();
+}
+function isMainModule() {
+    try {
+        const realArgvPath = realpathSync(process.argv[1]);
+        return import.meta.url === pathToFileURL(realArgvPath).href;
     }
-    if (!existsSync(CONFIG_PATH)) {
-        const { runInstall } = await import('./install.js');
-        await runInstall();
-        return;
+    catch {
+        return false;
     }
-    const { initColors } = await import('./agent/colors.js');
-    const { initDb } = await import('./agent/db.js');
-    const { initAgent } = await import('./agent/agent.js');
-    const { interactiveChat } = await import('./agent/cli.js');
-    const { shutdownMcp } = await import('./agent/mcp/index.js');
-    const { checkModel } = await import('./agent/model.js');
-    await initColors();
-    await checkModel();
-    initDb();
-    await initAgent();
-    process.once('SIGINT', async () => {
-        await shutdownMcp();
-        process.exit(0);
-    });
-    await interactiveChat();
-    await shutdownMcp();
 }
-main();
+// 仅在直接运行本文件时执行 main，避免被测试导入时自动执行
+if (isMainModule()) {
+    main();
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zhitalk",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "A personal AI Agent like OpenClaw, including tools, skills, memory, hook, sub-agent, MCP server, etc. Run in the terminal.",
   "type": "module",
   "main": "dist/index.js",