@fengrru/guardrail

Experimental — API may break in minor versions. See STABILITY.md.

Runtime safety guard with risk classification for AI agent operations.

Install

npm install @fengrru/guardrail

Quick Start

import { GuardRail, EntropyController } from "@fengrru/guardrail"

const guard = new GuardRail()

// Check risk before execution
const risk = guard.classifyRisk({
  type: "file_write",
  path: "src/index.ts",
})

if (risk.level > 1) {
  await guard.requestConfirmation(risk)
}

// Monitor entropy
const entropy = new EntropyController({ tokenBudget: 1000000 })
const action = entropy.evaluate(metrics)

Risk Levels

| Level | Description | Example | |-------|-------------|---------| | 0 | Read-only | Read file | | 1 | Local modification | Edit file | | 2 | Global impact | Delete directory | | 3 | Destructive | Drop database |

Features

• 4-level risk classification: 0-3 scale
• Entropy monitoring: token budget, failure rate
• Action confirmation: high-risk operations require approval
• Research mode: consecutive failures trigger pause

Documentation

• API Reference — TypeDoc-generated API docs
• Source Code
• Examples

License

MIT