Overview

@amritk/generate-examples turns a JSON Schema into test data. Where the other mjst generators give you code that consumes data at runtime (parsers, validators, types), this one closes the loop by giving you data to exercise that code with.

Each generated file exports:

• A TypeScript type definition for the schema
• A fast-check arbitrary (FooArbitrary) that produces schema-valid values — ideal for property-based testing
• A concrete, self-contained example value (fooExample) — ideal for fixtures, seeds, and documentation

An index.ts barrel re-exports everything.

[!NOTE] The generated arbitraries import fast-check, so consumers need it installed (npm i -D fast-check). An arbitrary whose schema uses a keyword no fc.* combinator captures on its own (if/then/else, not, exclusive oneOf, contains, and the presence-gated object keywords — patternProperties, propertyNames, dependent*, min/maxProperties) also imports @amritk/runtime-validators for a post-generation validating filter; files that need no such filter don't. The static fooExample values have no runtime dependencies.

@amritk/runtime-validators is a dependency here rather than a peer, because this generator imports it itself. That resolves it for the generator, but not necessarily for the generated file — that file lands in your source tree, so under pnpm's strict layout or Yarn PnP it resolves from your project, not from this package's. If your schemas use any of those keywords, install it directly (npm i @amritk/runtime-validators). It cannot also be declared a peer: Bun rejects a workspace package listed as both, and --frozen-lockfile then fails for the whole repo.

Installation

npm install @amritk/generate-examples
# or
pnpm add @amritk/generate-examples
# or
yarn add @amritk/generate-examples
# or
bun add @amritk/generate-examples

Usage

import { buildExampleSchema } from '@amritk/generate-examples'

const schema = {
  type: 'object',
  properties: {
    id: { type: 'string', format: 'uuid' },
    age: { type: 'integer', minimum: 0 },
  },
  required: ['id'],
} as const

const files = await buildExampleSchema(schema, 'User')
// → [{ filename: 'user.ts', content: '...' }, { filename: 'index.ts', content: '...' }]

The generated user.ts looks like:

import * as fc from 'fast-check'

export type User = { id: string; age?: number }

export const UserArbitrary: fc.Arbitrary<User> = fc.record(
  { "id": fc.uuid(), "age": fc.integer({ min: 0 }) },
  { requiredKeys: ["id"] },
)

export const userExample: User = { "id": "00000000-0000-0000-0000-000000000000", "age": 0 }

Use the arbitrary in a property test:

import { test, fc } from '@fast-check/vitest'
import { UserArbitrary } from './generated'
import { parseUser } from './parsers'

test.prop([UserArbitrary])('parseUser round-trips any valid User', (user) => {
  expect(parseUser(user)).toEqual(user)
})

…or grab the static example as a fixture:

import { userExample } from './generated'

const res = await fetch('/users', { method: 'POST', body: JSON.stringify(userExample) })

Lower-level API

Supported keywords

type — including multi-type unions like ['string', 'null'] — (string/number/integer/boolean/null/array/object), properties, required, additionalProperties, items, prefixItems, minItems/maxItems, uniqueItems, minLength/maxLength, pattern, format, minimum/maximum, exclusiveMinimum/exclusiveMaximum, multipleOf, enum (filtered by sibling constraints), const, minProperties/maxProperties, patternProperties, propertyNames, dependentRequired, dependentSchemas, contains, minContains, allOf (merged into one schema), oneOf/anyOf, if/then/else, not, $ref, and the x-mjst extension (Date, bigint). if/then/else, not, and oneOf exclusivity are enforced by validating generated candidates against the schema and retrying/rejecting. Unsupported constructs degrade to fc.anything() in arbitraries and null in static examples.

Static examples cover every format @amritk/runtime-validators knows how to check: email, idn-email, date, date-time, time, duration, uuid, uri, iri, uri-reference, iri-reference, uri-template, json-pointer, relative-json-pointer, hostname, idn-hostname, ipv4, ipv6, regex, plus OpenAPI's url. An unrecognized format falls back to "string".

Known limits

Every fooExample is validated against its own schema before it is written. When the value does not satisfy the schema it is still emitted — so the module always compiles — but the generator prints a console.warn naming the type. Reach for FooArbitrary in those cases: the arbitrary carries a runtime validating filter and stays correct where the static value cannot.

The value falls short for three reasons:

• The schema has no instance. { pattern: '^ab$', minLength: 5 }, uniqueItems over booleans with minItems: 3, a required key that additionalProperties: false forbids, or a oneOf whose branches every value matches twice. Nothing correct exists to emit; the warning is pointing at the schema, not the generator.
• The constraint is beyond the deriver. pattern is sampled by a best-effort recursive-descent walk of the regex, so lookarounds and backreferences fall back to "string"; an unrecognized format does the same.
• The bound is larger than any fixture should be. A derived string, array, or object stops growing at 10,000 characters / elements / keys, so a document asking for minLength: 50000000 yields a capped value and a warning rather than a 50 MB literal. FooArbitrary still honours the real bound.

A few more shapes worth knowing about, each of which keeps the generated file compiling rather than making it correct:

• A schema can require a key its generated type never declares — required naming something absent from properties, a dependentRequired / dependentSchemas dependency, or a minProperties filler on an object with no index signature. The example keeps the key (a fixture missing what its schema demands is broken data) and is emitted as … as Foo, since a bare object literal with an excess property fails to compile.
• An authored examples[0] (or, when there are no examples, the default) is used only when it satisfies its own schema. A hint that does not ({ type: 'string', default: 42 } — common in documents whose field types changed after the hint was written) is ignored in favour of a structurally derived value, because the generated type follows the schema and would reject the hint outright. const is always honoured: the type is the const's own literal type, so the two cannot disagree.
• An unsatisfiable range (minLength: 10, maxLength: 2) collapses onto its upper bound in the arbitrary. Every bounded fc.* combinator asserts min <= max and throws at import, which would take down every other export in the file alongside it. Integer bounds are also confined to fast-check's own 32-bit range, and length/count bounds to non-negative integers.
• A recursive definition's example has to stop somewhere and stops with null, which the non-nullable type does not admit — so it is emitted as … as unknown as Node. NodeArbitrary ties the recursion properly through fc.letrec and needs no such escape.
• A pattern that is not a valid JavaScript regex, or that uses a lookahead or lookbehind, falls back to a plain fc.string(). fc.stringMatching compiles the pattern at module scope and cannot generate from an assertion, so honouring it would throw where the whole file becomes unusable rather than just that one arbitrary being loose.
• Nesting deeper than 400 levels is refused with an error naming the limit. Building an arbitrary costs several stack frames per schema level, so a deeper document exhausts the stack — the cap turns that into a message that says what is wrong.

Two shapes stay impossible to generate from, and the arbitrary will retry forever if you sample it. Both are schemas with no instance, and the example warns:

• A pattern no string of the required length can match ({ pattern: '^[a-z]{2}$', minLength: 5 }). A satisfiable-but-narrow pairing ({ pattern: '^[a-f0-9]+$', minLength: 32, maxLength: 32 }) is slow for the same reason — fc.stringMatching rarely lands on the exact length.
• A minLength/minItems so large that no value of that size can be built.

One more gap is not this package's to close: a $ref that resolves nowhere in the document is typed by its name (Nope) but never imported, because it was never generated as a file. The arbitrary degrades to fc.anything(), but the type still names it, so the file does not compile. Same for { "type": [] }, which types as export type Foo = ;. Both come from @amritk/helpers/generate-type-definition.

[!TIP] The example for a $ref is inlined by value, so a definition graph with wide fan-out produces a correspondingly large literal. That cost is in the output size, not in generation time — each definition is derived once per document.

License

MIT