@aroha-sdk/policy

Pre-LLM policy guardrails for Aroha agents — declarative rules that run before the model is called, complementing cryptographic mandate bounds.

Install

npm install @aroha-sdk/policy

Quick example

import { PolicyEngine } from "@aroha-sdk/policy";

const policy = new PolicyEngine({
  rules: [
    // Block requests that exceed the mandate's allowed actions
    {
      name: "allowed-actions",
      check: (req, mandate) =>
        mandate.constraints.allowedActions.includes(req.action),
      onViolation: "reject",
      message: "Action not permitted by mandate",
    },
    // Rate-limit to protect the spending cap
    {
      name: "rate-limit",
      check: (req, mandate) => rateLimiter.allow(mandate.granteeDID),
      onViolation: "reject",
      message: "Rate limit exceeded",
    },
    // Flag suspicious prompts for review before calling the LLM
    {
      name: "prompt-injection-guard",
      check: (req) => !containsInjectionPattern(req.message),
      onViolation: "flag",
    },
  ],
});

serve({
  name: "my-agent", did, privateKey,
  onRequest: async (req) => {
    const { allowed, reason } = await policy.evaluate(req, req.mandate);
    if (!allowed) return { error: reason, status: 403 };
    // ... call LLM
  },
});

Why this exists in the mandate chain

Cryptographic mandate bounds enforce what an agent is authorized to do. Policy rules enforce what it should do — prompt injection guards, rate limits, content filters, and action-allow-lists that run before the LLM ever sees the request.

API

• PolicyEngine(opts) — create a policy engine with a list of named rules.
  • engine.evaluate(req, mandate) — run all rules. Returns { allowed, reason?, flagged? }.
• PolicyRule — { name, check(req, mandate): boolean | Promise<boolean>, onViolation: "reject" | "flag", message? }.
• buildDefaultPolicy() — sensible default ruleset: prompt injection guard, action allowlist, and spend-cap pre-check.

License

MIT