Author Policies with the SDK

Define deterministic executable Policy decisions and use canonical identity, action, Resource, and constraint contracts.

Use policy() for every public Policy package. It accepts executable Policy definitions only; declarative statements are not a public authoring or Resource-creation API. Policy code returns one outcome value. The execution environment owns request and outcome hashes.

Executable Policy

import { policy } from "@constal/sdk";

export default policy({
  id: "ticket-read", version: "1.0.0",
  evaluate(input) {
    const allowed = input.action === "resource:invoke"
      && input.context["resource.operation"] === "ticket.read";
    return allowed
      ? { kind: "allow" as const }
      : { kind: "deny" as const, code: "operation-denied", reason: "operation denied" };
  },
});

The evaluator has no network, secret, storage, clock, random, timer, or trusted-runtime capability. It receives normalized identity, exact action and Resource, bounded context, optional invocation data, and a requested grant when applicable. Return exactly one allow, deny, require-approval, substitute, constrain, or grant outcome. Explicit deny wins; scopes only narrow existing authority. Unsupported outcomes or constraints deny at the consuming boundary.

Package the definition with constal.policy.json. Its required target names exactly one resourceKind and a normalized label selector. The selector determines attachment; Policy code does not declare Resource patterns and invocation arguments cannot change membership.

Built-in governance helpers such as sandboxControls(), sandboxLimit(), modelControls(), modelLimit(), and agentLimit() create typed constraints for an exact Resource governance contract. The platform validates them against the target contract, derives customer/subject/tenant scope from trusted authority, reserves hard limits before dispatch, and passes only effective controls to the provider. Continue with Govern Resource usage, Author a Policy, and CLI Policy evaluation.