The idea
A value of type Type<A, O, I>
(called "runtime type") is the runtime representation of the static type A
.
Also a runtime type can
- decode inputs of type
I
(throughdecode
) - encode outputs of type
O
(throughencode
) - be used as a custom type guard (through
is
)
export type mixed = object | number | string | boolean | symbol | undefined | null
class Type<A, O = A, I = mixed> {
readonly _A: A
readonly _O: O
readonly _I: I
constructor(
/** a unique name for this runtime type */
readonly name: string,
/** a custom type guard */
readonly is: (v: mixed) => v is A,
/** succeeds if a value of type S can be decoded to a value of type A */
readonly validate: (input: I, context: Context) => Either<Errors, A>,
/** converts a value of type A to a value of type S */
readonly encode: (a: A) => O
) {}
/** a version of `validate` with a default context */
decode(i: I): Either<Errors, A>
}
Note. The Either
type is defined in fp-ts, a library containing implementations of
common algebraic types in TypeScript.
Example
A runtime type representing string
can be defined as
import * as t from 'io-ts'
export class StringType extends Type<string> { // equivalent to Type<string, string, mixed> as per type parameter defaults
readonly _tag: 'StringType' = 'StringType'
constructor() {
super(
'string',
(m): m is string => typeof m === 'string',
(m, c) => (this.is(m) ? success(m) : failure(m, c)),
a => a
)
}
}
A runtime type can be used to validate an object in memory (for example an API payload)
const Person = t.type({
name: t.string,
age: t.number
})
// ok
Person.decode(JSON.parse('{"name":"Giulio","age":43}')) // => Right({name: "Giulio", age: 43})
// ko
Person.decode(JSON.parse('{"name":"Giulio"}')) // => Left([...])
Error reporters
A reporter implements the following interface
interface Reporter<A> {
report: (validation: Validation<any>) => A;
}
This package exports two default reporters
PathReporter: Reporter<Array<string>>
ThrowReporter: Reporter<void>
Example
import { PathReporter } from 'io-ts/lib/PathReporter'
import { ThrowReporter } from 'io-ts/lib/ThrowReporter'
const result = Person.decode({ name: 'Giulio' })
console.log(PathReporter.report(result))
// => ['Invalid value undefined supplied to : { name: string, age: number }/age: number']
ThrowReporter.report(result)
// => throws 'Invalid value undefined supplied to : { name: string, age: number }/age: number'
Community
- io-ts-types - A collection of runtime types and combinators for use with io-ts
- io-ts-reporters - Error reporters for io-ts
- geojson-iots - Runtime types for GeoJSON as defined in rfc7946 made with io-ts
TypeScript integration
Runtime types can be inspected
This library uses TypeScript extensively. Its API is defined in a way which automatically infers types for produced values
Note that the type annotation isn't needed, TypeScript infers the type automatically based on a schema.
Static types can be extracted from runtime types with the TypeOf
operator
type IPerson = t.TypeOf<typeof Person>
// same as
type IPerson = {
name: string,
age: number
}
Implemented types / combinators
import * as t from 'io-ts'
Type | TypeScript | Flow | Runtime type / combinator |
---|---|---|---|
null | null |
null |
t.null or t.nullType
|
undefined | undefined |
void |
t.undefined |
string | string |
string |
t.string |
number | number |
number |
t.number |
boolean | boolean |
boolean |
t.boolean |
any | any |
any |
t.any |
never | never |
empty |
t.never |
object | object |
✘ | t.object |
integer | ✘ | ✘ | t.Integer |
array of any | Array<mixed> |
Array<mixed> |
t.Array |
array of type | Array<A> |
Array<A> |
t.array(A) |
dictionary of any | { [key: string]: mixed } |
{ [key: string]: mixed } |
t.Dictionary |
dictionary of type | { [K in A]: B } |
{ [key: A]: B } |
t.dictionary(A, B) |
function | Function |
Function |
t.Function |
literal | 's' |
's' |
t.literal('s') |
partial | Partial<{ name: string }> |
$Shape<{ name: string }> |
t.partial({ name: t.string }) |
readonly | Readonly<T> |
ReadOnly<T> |
t.readonly(T) |
readonly array | ReadonlyArray<number> |
ReadOnlyArray<number> |
t.readonlyArray(t.number) |
interface | interface A { name: string } |
interface A { name: string } |
t.type({ name: t.string }) or t.type({ name: t.string })
|
interface inheritance | interface B extends A {} |
interface B extends A {} |
t.intersection([ A, t.type({}) ]) |
tuple | [ A, B ] |
[ A, B ] |
t.tuple([ A, B ]) |
union | A | B |
A | B |
t.union([ A, B ]) or t.taggedUnion(tag, [ A, B ])
|
intersection | A & B |
A & B |
t.intersection([ A, B ]) |
keyof | keyof M |
$Keys<M> |
t.keyof(M) |
recursive types | see Recursive types | see Recursive types | t.recursion(name, definition) |
refinement | ✘ | ✘ | t.refinement(A, predicate) |
strict/exact types | ✘ | $Exact<{{ name: t.string }}> |
t.strict({ name: t.string }) |
Recursive types
Recursive types can't be inferred by TypeScript so you must provide the static type as a hint
// helper type
type ICategory = {
name: string,
categories: Array<ICategory>
}
const Category =
t.recursion <
ICategory >
('Category',
self =>
t.type({
name: t.string,
categories: t.array(self)
}))
Tagged unions
If you are encoding tagged unions, instead of the general purpose union
combinator, you may want to use the
taggedUnion
combinator in order to get better performances
const A = t.type({
tag: t.literal('A'),
foo: t.string
})
const B = t.type({
tag: t.literal('B'),
bar: t.number
})
// the actual presence of the tag is statically checked
const U = t.taggedUnion('tag', [A, B])
Refinements
You can refine a type (any type) using the refinement
combinator
const Positive = t.refinement(t.number, n => n >= 0, 'Positive')
const Adult = t.refinement(Person, person => person.age >= 18, 'Adult')
Strict/Exact interfaces
You can make an interface strict (which means that only the given properties are allowed) using the strict
combinator
const Person = t.type({
name: t.string,
age: t.number
})
const StrictPerson = t.strict(Person.props)
Person.decode({ name: 'Giulio', age: 43, surname: 'Canti' }) // ok
StrictPerson.decode({ name: 'Giulio', age: 43, surname: 'Canti' }) // fails
Mixing required and optional props
Note. You can mix required and optional props using an intersection
const A = t.type({
foo: t.string
})
const B = t.partial({
bar: t.number
})
const C = t.intersection([A, B])
type CT = t.TypeOf<typeof C>
// same as
type CT = {
foo: string
bar?: number
}
You can define a custom combinator to avoid the boilerplate
export function interfaceWithOptionals<RequiredProps extends t.Props, OptionalProps extends t.Props>(
required: RequiredProps,
optional: OptionalProps,
name?: string
): t.IntersectionType<
[
t.InterfaceType<RequiredProps, t.TypeOfProps<RequiredProps>>,
t.PartialType<OptionalProps, t.TypeOfPartialProps<OptionalProps>>
],
t.TypeOfProps<RequiredProps> & t.TypeOfPartialProps<OptionalProps>
> {
return t.intersection([t.interface(required), t.partial(optional)], name)
}
const C = interfaceWithOptionals({ foo: t.string }, { bar: t.number })
Custom types
You can define your own types. Let's see an example
import * as t from 'io-ts'
// represents a Date from an ISO string
const DateFromString = new t.Type<Date, string>(
'DateFromString',
(m): m is Date => m instanceof Date,
(m, c) =>
t.string.validate(m, c).chain(s => {
const d = new Date(s)
return isNaN(d.getTime()) ? t.failure(s, c) : t.success(d)
}),
a => a.toISOString()
)
const s = new Date(1973, 10, 30).toISOString()
DateFromString.decode(s)
// right(new Date('1973-11-29T23:00:00.000Z'))
DateFromString.decode('foo')
// left(errors...)
Note that you can deserialize while validating.
Custom combinators
You can define your own combinators. Let's see some examples
maybe
combinator
The An equivalent to T | null
export function maybe<RT extends t.Any>(
type: RT,
name?: string
): t.UnionType<[RT, t.NullType], t.TypeOf<RT> | null, t.OutputOf<RT> | null, t.InputOf<RT> | null> {
return t.union<[RT, t.NullType]>([type, t.null], name)
}
pluck
combinator
The Extracting the runtime type of a field contained in each member of a union
const pluck = <F extends string, U extends t.UnionType<Array<t.InterfaceType<{ [K in F]: t.Mixed }>>>>(
union: U,
field: F
): t.Type<t.TypeOf<U>[F]> => {
return t.union(union.types.map(type => type.props[field]))
}
export const Action = t.union([
t.type({
type: t.literal('Action1'),
payload: t.type({
foo: t.string
})
}),
t.type({
type: t.literal('Action2'),
payload: t.type({
bar: t.string
})
})
])
// ActionType: t.Type<"Action1" | "Action2", "Action1" | "Action2", t.mixed>
const ActionType = pluck(Action, 'type')
Recipes
Is there a way to turn the checks off in production code?
No, however you can define your own logic for that (if you really trust the input and the involved types don't perform deserializations)
import * as t from 'io-ts'
import { failure } from 'io-ts/lib/PathReporter'
const { NODE_ENV } = process.env
export function unsafeValidate<S, A>(value: any, type: t.Type<S, A>): A {
if (NODE_ENV !== 'production') {
return type.decode(value).getOrElse(errors => {
throw new Error(failure(errors).join('\n'))
})
}
// unsafe cast
return value as A
}
Known issues
Due to an upstream bug, VS Code might display weird types for nested interfaces
const NestedInterface = t.type({
foo: t.type({
bar: t.string
})
})
type NestedInterfaceType = t.TypeOf<typeof NestedInterface>
/*
Hover on NestedInterfaceType will display
type NestedInterfaceType = {
foo: t.InterfaceOf<{
bar: t.StringType;
}>;
}
instead of
type NestedInterfaceType = {
foo: {
bar: string;
};
}
*/