Structured data and real checks

Give the next step a usable contract without confusing valid JSON with a correct answer.

A structured result lets code use an agent's answer. It does not establish that the answer is true. Treat shape validation and acceptance checks as separate steps.

Agent response → schema validation → domain checks → decision
                  usable shape       supported result

Define what the next step needs

review.ts

import { Schema } from "effect";

export const Review = Schema.Struct({
  approved: Schema.Boolean,
  feedback: Schema.String,
  evidence: Schema.Array(Schema.Struct({
    claim: Schema.String,
    source: Schema.String,
  })),
});
export type ReviewResult = typeof Review.Type;

export function canAccept(review: ReviewResult): boolean {
  return review.approved && review.evidence.length > 0;
}

Pass Review as output on Session.run or session.ask. Your result now has boolean approved and an array of evidence entries. canAccept rejects an empty evidence list; it still cannot tell whether the cited source supports the claim. A source-checking step or a person must establish that.

See what the schema rejects

review.test.ts

import { expect, test } from "bun:test";
import { Schema } from "effect";
import { Review, canAccept } from "./review";

test("a valid shape can still fail the acceptance rule", () => {
  const review = Schema.decodeUnknownSync(Review)({
    approved: true, feedback: "Looks good", evidence: [],
  });
  expect(canAccept(review)).toBe(false);
  expect(() => Schema.decodeUnknownSync(Review)({
    approved: "yes", feedback: "Looks good", evidence: [],
  })).toThrow();
});

The first object decodes but fails the domain check. The second object fails schema decoding because "yes" is not a boolean.

Decide what failure means

Outcome	Response in your workflow
Invalid schema	Stop or take an explicit repair branch
Valid negative review	Revise, decline or escalate according to policy
Provider failure	Preserve the error; a review did not complete
Nonzero verification command	Reject the candidate using actual command evidence

Laufwerk turns your output schema into prompt guidance and decodes the response. It does not use provider-enforced constrained generation or automatically rerun a failed structured turn: the original turn may already have used tools or edited files. Keep acceptance checks outside the agent's editable material where possible.

Reference: Schema types · Session output semantics

← Sessions and shared files · Next: Branches, loops and parallel work →

Structured data and real checks

Define what the next step needs

See what the schema rejects

Decide what failure means

On this page