← 返回 Skills 市场

AI Agent 自省调试框架

Name: AI Agent 自省调试框架
Author: danihe001

作者 steve xia · GitHub ↗ · v1.1.0 · MIT-0

cross-platform ⚠ suspicious

305

总下载

当前安装

版本数

在 OpenClaw 中安装

/install introspection-debugger

功能描述

AI Agent 自省调试框架 - 让 AI Agent 具备自我诊断和自动修复能力。用于捕获错误、根因分析、自动修复、生成报告。

使用说明 (SKILL.md)

AI Agent 自省调试框架

让 AI Agent 具备自省和自我修复能力。

功能

功能	描述
全局错误捕获	自动拦截 uncaughtException 和 unhandledRejection
根因分析	基于规则库匹配常见错误 (80%+)
自动修复	自动创建文件、修复权限、安装依赖
报告生成	生成结构化自省报告
人类通知	无法修复时通知人类

支持的错误类型

文件缺失 (ENOENT)
权限错误 (EACCES)
模块缺失 (MODULE_NOT_FOUND)
连接超时 (ETIMEDOUT)
限流 (429)
服务器错误 (500-504)
内存溢出 (OOM)
进程被终止 (SIGKILL)
认证错误 (401/403)

使用方法

1. 引入模块

const IntrospectionDebugger = require('./introspection-debugger');

2. 创建实例

const debugger = new IntrospectionDebugger({
  workspace: process.cwd(),
  maxHistorySize: 100,
  notificationHook: async (report) => {
    // 通知人类
    console.log('需要人工:', report.recommendation.message);
  }
});

3. 手动捕获错误

try {
  // 你的代码
} catch (e) {
  debugger.catch(e, { source: 'my-code' });
}

4. 查看统计

console.log(debugger.getStats());
// { totalErrors: 10, totalFixes: 8, autoFixRate: 0.8 }

文件

introspection-debugger.js - 主框架代码
reliable-api-client.js - API 客户端黄金标准

相关文档

INTROSPECTION_DEBUGGER.md - 详细使用文档
EVOMAP_STANDARD.md - 胶囊发布规范

安全使用建议

This skill does what it says (auto-diagnose and auto-fix), but it performs filesystem writes, chmods, and runs shell commands (including npm install) based on parsed error messages. Before installing or using it: (1) Review the full source for helper functions not shown here (extractFilePath, extractModuleName, execAsync, ensureDir) to confirm they properly validate and sanitize inputs; (2) Do not enable auto-fix in environments that process untrusted inputs — prefer notificationHook + manual confirmation; (3) Run the module in a restricted workspace/least-privilege container or sandbox until you vet it; (4) Consider removing or disabling automatic npm installs and any shell execs, or require explicit human approval for fixes; (5) If you need higher assurance, request the complete (untruncated) source for review so we can check for command injection, remote endpoints, or hidden network calls.

功能分析

Type: OpenClaw Skill Name: introspection-debugger Version: 1.1.0 The skill implements an auto-repair framework in `introspection-debugger.js` that uses `child_process.exec` to run shell commands (`npm install`, `chmod`) and `fs.writeFileSync` to create files based on strings extracted from error messages. This creates a significant risk of command injection and unauthorized file system modification if error messages are influenced by untrusted input. Additionally, the `notifyHuman` function can exfiltrate detailed error reports, including stack traces and execution context, to an external URL via a webhook, which could be abused for data exfiltration.

能力评估

✓ Purpose & Capability

The name/description (introspection + auto-repair) align with the shipped code and SKILL.md: the module captures global errors, analyzes root causes, writes files, fixes permissions, and installs missing modules — all expected for an auto-debugger.

ℹ Instruction Scope

SKILL.md instructs importing and instantiating the module and lists features (global capture, auto-fix). That matches the implementation, but the runtime behavior has broad side effects (registers process-level uncaughtException/unhandledRejection handlers and attempts automatic fixes) that the README does not tightly constrain (for example: when auto-fix runs, whether user confirmation is required, or limits on what paths/commands it will act on).

✓ Install Mechanism

This is an instruction-only skill with a code file and no install spec — nothing is downloaded or installed at skill install time. Risk comes from runtime operations (the code invokes shell commands and npm install when fixing).

⚠ Credentials

The skill requests no environment variables, which is appropriate, but it executes shell commands derived from parsed error messages (e.g., `chmod +x "${filePath}"`, `npm install ${moduleName}`). If error messages or workspace contents are attacker-controlled, these operations could lead to command injection, arbitrary package installation, or unwanted filesystem changes. There is no declared limit or sanitization visible in the provided code excerpt.

ℹ Persistence & Privilege

always:false and user-invocable:true — normal. The skill does, however, register global process handlers when instantiated, so simply requiring/constructing it will alter process-wide behavior; that's coherent with its purpose but important to be aware of.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install introspection-debugger
安装完成后，直接呼叫该 Skill 的名称或使用 /introspection-debugger 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v1.1.0

v1.1.0 优化：修复根因分析字段、增强调试日志、修复 installDependency 返回值、完善测试报告

元数据

Slug introspection-debugger

版本 1.1.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 1

常见问题