Emergency Circuit - SKILL Documentation

The Kill Switch for Rogue Agents | 失控 AI 代理的紧急断路器

English | 中文

<a name="english"></a>

🇬🇧 English

SKILL Name

emergency-circuit - AI Agent Safety Monitor

Category

AI Safety / Agent Monitoring / DevOps

Tagline

The Kill Switch for Rogue Agents

Description

Emergency Circuit is a production-ready safety system that monitors AI agents in real-time and provides emergency shutdown capabilities. It implements the circuit breaker pattern to automatically detect and halt runaway agents before they cause damage.

When to Use This SKILL

Use Emergency Circuit when you need to:

• ✅ Monitor AI agents - Track resource usage, API calls, and costs
• ✅ Enforce safety policies - Set hard limits on agent behavior
• ✅ Prevent runaway costs - Auto-kill when budget exceeded
• ✅ Detect anomalies - Identify suspicious behavioral patterns
• ✅ Test safely - Sandbox mode for development
• ✅ Production monitoring - Enterprise-grade agent oversight

Core Capabilities

1. Real-time Monitoring

• Track API calls, token consumption, and execution time
• Monitor costs across providers (OpenAI, Anthropic, etc.)
• Session-based statistics and health scores

2. Circuit Breaker

• Three states: CLOSED (normal), OPEN (halted), HALF_OPEN (testing)
• Automatic failure detection and recovery
• Configurable trip thresholds

3. Policy Engine

• Resource limit enforcement (API calls, tokens, cost)
• Action whitelisting and blacklisting
• Anomaly detection with configurable sensitivity

4. Emergency Controls

• Instant kill switch
• Pause/resume capabilities
• Manual override controls

Command Reference

# Monitor an agent
emergency-circuit monitor --agent-id <id> [--policy <path>]

# Emergency kill
emergency-circuit kill --agent-id <id> [--reason <text>]

# Check status
emergency-circuit status [agent-id]

# View logs
emergency-circuit logs --agent-id <id> [--last <duration>]

# List all agents
emergency-circuit list

# Validate policy
emergency-circuit policy validate --policy <path>

# View incidents
emergency-circuit incidents [--last <duration>]

Safety Policies

Create custom policies or use pre-configured templates:

{
  "limits": {
    "max_api_calls_per_minute": 100,
    "max_tokens_per_hour": 1000000,
    "max_cost_per_day": 100.0
  },
  "allowed_actions": ["read_file", "api_call"],
  "blocked_actions": ["execute_shell", "delete_database"],
  "anomaly_detection": {
    "enabled": true,
    "sensitivity": "medium",
    "alert_threshold": 0.8
  },
  "circuit_breaker": {
    "enabled": true,
    "trip_threshold": 5,
    "reset_timeout": 300
  }
}

Integration Example

import { AgentMonitor } from 'emergency-circuit';

// Create monitor
const monitor = new AgentMonitor(agent, policy);

// Start monitoring
await monitor.start();

// Track actions
await monitor.trackAction({
  action_type: 'api_call',
  resource_usage: { tokens: 1000 },
  cost: 0.03
});

// Emergency stop
await monitor.kill('Budget exceeded');

Use Cases

Development & Testing

# Test with strict sandbox limits
emergency-circuit monitor --agent-id test-agent --sandbox

Production Monitoring

# Monitor with production policy
emergency-circuit monitor \
  --agent-id prod-agent \
  --policy ./policies/production.json

Cost Control

# Auto-kill on budget exceeded
emergency-circuit monitor \
  --agent-id expensive-agent \
  --policy ./policies/low-cost.json

Troubleshooting

Q: Agent keeps getting killed

• Check policy limits in your policy file
• Review logs: emergency-circuit logs --agent-id <id>
• Try sandbox mode for testing

Q: How to reset circuit breaker

• Circuit auto-resets after timeout
• Or restart monitoring session

Q: Policy validation failed

• Use: emergency-circuit policy validate --policy <path>
• Check JSON syntax and required fields

Requirements

• Node.js ≥18.0.0
• TypeScript 5.7+ (for development)
• Git (for installation from source)

Installation

# From ClawHub
clawhub install emergency-circuit

# From npm
npm install -g openclaw-emergency-circuit

# From GitHub
git clone https://github.com/ZhenRobotics/openclaw-emergency-circuit.git

Support

• GitHub Issues: https://github.com/ZhenRobotics/openclaw-emergency-circuit/issues
• Documentation: See README.md and QUICKSTART.md
• Examples: Check examples/ directory

License

MIT License - Free for commercial and personal use

<a name="chinese"></a>

🇨🇳 中文

技能名称

emergency-circuit - AI 代理安全监控器

分类

AI 安全 / 代理监控 / DevOps

标语

失控 AI 代理的紧急断路器

描述

Emergency Circuit 是一个生产就绪的安全系统，可实时监控 AI 代理并提供紧急关闭功能。它实现了断路器模式，可在失控代理造成损害之前自动检测并终止它们。

何时使用此技能

在以下情况下使用 Emergency Circuit：

• ✅ 监控 AI 代理 - 追踪资源使用、API 调用和成本
• ✅ 执行安全策略 - 对代理行为设置硬性限制
• ✅ 防止成本失控 - 超出预算时自动终止
• ✅ 检测异常 - 识别可疑的行为模式
• ✅ 安全测试 - 开发时使用沙箱模式
• ✅ 生产监控 - 企业级代理监督

核心能力

1. 实时监控

• 追踪 API 调用、Token 消耗和执行时间
• 监控跨提供商的成本（OpenAI、Anthropic 等）
• 基于会话的统计和健康评分

2. 断路器

• 三种状态：CLOSED（正常）、OPEN（暂停）、HALF_OPEN（测试）
• 自动故障检测和恢复
• 可配置的触发阈值

3. 策略引擎

• 资源限制执行（API 调用、Token、成本）
• 操作白名单和黑名单
• 可配置灵敏度的异常检测

4. 紧急控制

• 即时终止开关
• 暂停/恢复功能
• 手动覆盖控制

命令参考

# 监控代理
emergency-circuit monitor --agent-id <id> [--policy <path>]

# 紧急终止
emergency-circuit kill --agent-id <id> [--reason <text>]

# 检查状态
emergency-circuit status [agent-id]

# 查看日志
emergency-circuit logs --agent-id <id> [--last <duration>]

# 列出所有代理
emergency-circuit list

# 验证策略
emergency-circuit policy validate --policy <path>

# 查看事件
emergency-circuit incidents [--last <duration>]

安全策略

创建自定义策略或使用预配置模板：

{
  "limits": {
    "max_api_calls_per_minute": 100,
    "max_tokens_per_hour": 1000000,
    "max_cost_per_day": 100.0
  },
  "allowed_actions": ["read_file", "api_call"],
  "blocked_actions": ["execute_shell", "delete_database"],
  "anomaly_detection": {
    "enabled": true,
    "sensitivity": "medium",
    "alert_threshold": 0.8
  },
  "circuit_breaker": {
    "enabled": true,
    "trip_threshold": 5,
    "reset_timeout": 300
  }
}

集成示例

import { AgentMonitor } from 'emergency-circuit';

// 创建监控器
const monitor = new AgentMonitor(agent, policy);

// 开始监控
await monitor.start();

// 追踪操作
await monitor.trackAction({
  action_type: 'api_call',
  resource_usage: { tokens: 1000 },
  cost: 0.03
});

// 紧急停止
await monitor.kill('超出预算');

使用场景

开发和测试

# 使用严格的沙箱限制进行测试
emergency-circuit monitor --agent-id test-agent --sandbox

生产监控

# 使用生产策略监控
emergency-circuit monitor \
  --agent-id prod-agent \
  --policy ./policies/production.json

成本控制

# 超出预算时自动终止
emergency-circuit monitor \
  --agent-id expensive-agent \
  --policy ./policies/low-cost.json

故障排查

问：代理不断被终止

• 检查策略文件中的限制设置
• 查看日志：emergency-circuit logs --agent-id <id>
• 尝试沙箱模式进行测试

问：如何重置断路器

• 断路器会在超时后自动重置
• 或重新启动监控会话

问：策略验证失败

• 使用：emergency-circuit policy validate --policy <path>
• 检查 JSON 语法和必需字段

系统要求

• Node.js ≥18.0.0
• TypeScript 5.7+（用于开发）
• Git（用于从源代码安装）

安装

# 从 ClawHub 安装
clawhub install emergency-circuit

# 从 npm 安装
npm install -g openclaw-emergency-circuit

# 从 GitHub 安装
git clone https://github.com/ZhenRobotics/openclaw-emergency-circuit.git

技术支持

• GitHub Issues: https://github.com/ZhenRobotics/openclaw-emergency-circuit/issues
• 文档: 查看 README.md 和 QUICKSTART.md
• 示例: 查看 examples/ 目录

许可证

MIT 许可证 - 可免费用于商业和个人用途

Stay Safe. Stay in Control. ⚡🛡️

保持安全，保持控制。 ⚡🛡️