@zanii/policy

Decide whether an agent action is allowed before it runs — on conditions a delegation scope can't express: amount caps, allow/deny lists, time windows, rate limits. It complements @zanii/runtime (which enforces the tool manifest, delegation scope, and human-gate); this layer reasons about the payload.

npm install @zanii/policy

Usage

import { createPolicyEngine } from '@zanii/policy';

const engine = createPolicyEngine({
  default: 'allow',
  rules: [
    { id: 'sanctions', effect: 'deny', targets: ['pay.*'],
      where: [{ field: 'to.country', op: 'in', value: ['IR', 'KP'] }] },
    { id: 'big-transfer', effect: 'require_approval', targets: ['pay.transfer'],
      where: [{ field: 'amount', op: 'gt', value: 1000 }] },
    { id: 'email-flood', effect: 'deny', targets: ['email.send'],
      rateLimit: { max: 100, perSeconds: 3600 } },
  ],
});

const { decision, rule, reason } = engine.evaluate({
  target: 'pay.transfer',
  payload: { amount: 5000, to: { country: 'US' } },
  key: agentDid, // rate-limit bucket
});
// decision → 'require_approval'

Ordered rules, first match wins, explicit default. Predicate ops: eq ne lt lte gt gte in nin over dot-path payload fields. Target globs use the same syntax as delegation scopes (pay.*, *).

Wire decision into your flow: deny → don't act, require_approval → route to a human gate, allow → proceed and record the proof.

Changelog

• 0.1.0 — initial release: createPolicyEngine, rules with predicates + fixed-window rate limits.

License

Apache-2.0.