Suggestion: type checking for JSX children

[jwbay]

Rationale

Props are an important part of a React component's interface, but children can be as well.
In applications it's not uncommon to want to restrict what kind of children a component will
accept. Currently this enforcement has to be done at runtime by throwing errors.

The Goal

Components can specify allowed children as a union type. Each member would correspond to
the tag, or 'type' field, of the resulting Element. Ideally this could be plucked from
the type of props.children, which would satisfy both component classes and SFCs.

Motivating Examples:

A modal component may want to make assumptions about its children to satisfy layout
constraints.

interface IModalProps {
    children: ModalHeader | ModalBody | ModalFooter;
}

function Modal({ children }: IModalProps) { ... }
function ModalHeader(props) { ... }
function ModalBody(props) { ... }
function ModalFooter(props) { ... }

<Modal>
    <ModalHeader />
    <div /> { /* Desired error: type 'div' does not exist in type 'ModalHeader | ModalBody | ModalFooter' */ }
    <ModalBody />
    <ModalFooter />
</Modal>

Similarly...

interface IButtonGroupProps {
    children: 'button';
}

function ButtonGroup({ children }: IButtonGroupProps) { ... }

<ButtonGroup>
    <button />
    <button />
    <a href='' /> { /* Desired error: type 'a' is not assignable to type 'button' */ }
</ButtonGroup>

An interesting emerging pattern is using JSX as a function call to turn imperative APIs into
declarative ones. Currently these 'render callbacks' can't be type-checked at all. A more complete
summary can be found at http://reactpatterns.com/#render-callback. A further example can be found at https://github.com/ReactTraining/react-media#usage.

interface IUser {
    Name: string;
}

interface IFetchUserProps {
    children(user: IUser): any;
}

class FetchUser extends React.Component<IFetchUserProps, any> {
    render() {
        return this.state
            ? this.props.children(this.state.result)
            : null;
    }

    componentDidMount() {
        api.fetchUser().then(result => this.setState({ result }));
    }
}

function UserName() {
    return (
        <FetchUser>
            { user => (
                <h1>{ user.NAme }</h1> // Desired error: property 'NAme' does not exist on type 'IUser'
            ) }
        </FetchUser>
    );
}

Lastly, I don't think any of this is necessarily specific to React. Other libraries leveraging the JSX
specification for component structure should be able to leverage this as well. The crux of this lies at
the relationship between the type of the 'children' attribute of a given tag and the children actually
passed to it at usage sites.

[joelday]

@jwbay @RyanCavanaugh Did a quick prototype of this: joelday@50870c1

I'm sure there are a bunch of little gotchas here, but it's a start.

    // A custom element type with an explicit children prop attribute type
    
    class MyParentClass extends React.Component<{
      children?: MyClass[];
    }, any> {
    }
    
    // OK - Child element matches the children prop
    var d1 = <MyParentClass><MyClass reqd={true} /><MyClass reqd={true}></MyClass></MyParentClass>
    // Error - Incorrect child element type
    var d2 = <MyParentClass><div /><div></div><MyClass reqd={true} /><MyClass reqd={true}></MyClass></MyParentClass>
                            ~~~~~~~
!!! error TS2322: Type 'HTMLProps<HTMLDivElement>' is not assignable to type 'MyClass'.
!!! error TS2322:   Property 'setState' is missing in type 'HTMLProps<HTMLDivElement>'.
                                   ~~~~~
!!! error TS2322: Type 'HTMLProps<HTMLDivElement>' is not assignable to type 'MyClass'.```

[jwbay]

@joelday Oh, nice! Many thanks for taking at look at an implementation!

[joelday]

@yuit @mhegazy Let me know if there's anything I can do to help with this. Thank you!

[joelday]

@yuit Any traction with this? Would love for this to still make it into 2.3. Please let me know if there's anything I can do.

[petdun2519]

Would also like to see this implemented!

[yuit]

@joelday @petdun2519 PR will go up soon. Thanks for your protoype

[Jessidhia]

Tangentially related, but I wonder if there is a way to enforce that the children props, internally, is accessed using the React.Children helpers (https://facebook.github.io/react/docs/react-api.html#react.children). The value of children, as seen from the component implementation, is supposed to be opaque until accessed with a helper.

[cchamberlain]

This change has broken quite a few things for a patch release. The errors are not that helpful that I'm seeing, tsc throws TS2710: 'children' are specified twice... on spread props. If I attempt to strip out a children prop and then pass the filtered props the editor shows squigglies indicating children does not exist on the type I'm trying to filter them out of.

[joscha]

This change has broken quite a few things for a patch release. The errors are not that helpful that I'm seeing, tsc throws TS2710: 'children' are specified twice... on spread props. If I attempt to strip out a children prop and then pass the filtered props the editor shows squigglies indicating children does not exist on the type I'm trying to filter them out of.

I've noticed this as well @cchamberlain , our code now looks like this:

const { children, ...rest } = this.props;
return <Bla {...rest}>{children}</Bla>;

[cchamberlain]

@joscha - That did work for me in some of the instances I tried but there was at least one spot that when I tried it, it complained that the variable I was spreading did not contain an element children. The element being spread was not {...this.props} directly, it was {...this.props.childProps} and did not have children defined on its interface. I reverted back to TS 2.3.0 so I don't have an exact representation at the moment unfortunately.

I know TypeScript can't follow traditional semver without making most releases major versions, but this change breaks every component using spread and children (majority of our components) so seems like it warrants a minor version.

[cchamberlain]

Don't want to beat a dead horse but I just installed VS 2017 update 15.2 per #14529 (comment) and it only allows selection of TypeScript minor version, meaning I have to revert all the way back to 2.2.x if breaking changes are released on a patch (VS is reporting the errors I get with TS 2.3.2 even though I have package.json / yarn locked at 2.3.0).

Update: I got my solution working on 2.3, we had many instances of children being typed as the wrong type, causing this update to throw many errors. I very much like the typed children support, just feel that if a TypeScript release is not meant to strictly patch a previous release it should get a minor version, especially if the only thing we can specify in VS 2017 is minor version.

[donaldpipowitch]

Looking at #15160 (comment) it seems that the original idea of this issue isn't implemented and that the issue should be re-opened (imho).

You can type your children as

• a string, boolean, etc.
• to have no element, one element or multiple elements
• to be a function (which is really nice)

But can't say that a <ButtonGroup/> can only have <Button/>'s or that <Modal/> can only have <ModalHeader/>, <ModalBody/> and <ModalFooter/> as children.

That's because <Button/>, <ModalHeader/> and so on return a generic JSX.Element type.

[joewood]

@donaldpipowitch agreed, the current implementation doesn't seem to work as advertised.

[donaldpipowitch]

The more I think of it I come to the conclusion that this could probably not be solved at a type checking level, because of the way TSX works currently. We can't use something like tagging JSX.Element:

enum FooterTag {}
type FooterElement = JSX.Element & FooterTag;
const Footer = () => <div>Footer.</div> as FooterElement;

interface ModalProps {
  children: FooterElement;
}
const Modal = ({ children }: ModalProps) =>
  <div>
    {children}
  </div>;

// `<Footer />` *needs* to be casted to `FooterElement`,
// because the return value is always `JSX.Element`
const UseModal = () =>
  <Modal>
    {<Footer /> as FooterElement}
  </Modal>;

When we useTSX the return value of our component will always be JSX.Element, because of the way it is transpiled. If we can't change this way, than the only solution would be to make this check inside a linter rule. With a linter rule we could easily say "Hey, you can only use <Footer /> inside <Modal />." or something like that. This would be possible, if #15589 could be integrated into the language.

[donaldpipowitch]

How could this happen?

yuit was unassigned by donaldpipowitch 18 minutes ago

[antanas-arvasevicius]

Transpilation is done from JSX to "createElement" function calls and custom implementations can return anything. I can declare function definition and compiler can/should exactly know what return type will be.

…

On Aug 8, 2017 8:39 AM, "Donald Pipowitch" ***@***.***> wrote: How could this happen? [image: image] <https://user-images.githubusercontent.com/1152805/29057701-a71543a2-7c0c-11e7-88db-5edef0b77456.png> — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#13618 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEAcnA_fYkJzzRqyyGGF09p61Rz9tXLLks5sV_SagaJpZM4LqJuY> .

[RyanCavanaugh]

@donaldpipowitch Yui's left the team but was still assigned the issue. GitHub fixes this invalid assignment whenever someone makes any change (incl. comment) on an issue

[jwbay]

Couldn't JSX.Element be made generic with props/attributes?

const i = <input value='42' />;
//yields JSX.Element<{ value: string }>

const c = <Component prop={ 42 } />
//yields JSX.Element<{ prop: number }>

At this point the restricted children case devolves slightly from 'you can only pass something shaped like ModalFooter here' to 'you can only pass something with props shaped like ModalFooter's props', but for components I think that's really all that matters.

[donaldpipowitch]

Yui's left the team but was still assigned the issue. GitHub fixes this invalid assignment whenever someone makes any change (incl. comment) on an issue

Ah, I see. Thanks. So if you ever need someone new (remote)... 😏

At this point the restricted children case devolves slightly from 'you can only pass something shaped like ModalFooter here' to 'you can only pass something with props shaped like ModalFooter's props', but for components I think that's really all that matters.

As far as I understand this issue and what I personally would like to see is the first case ("you can only pass something shaped like ModalFooter here"), not the second case ("you can only pass something with props shaped like ModalFooter's props").

I want to enforce ordering "rules" of components similar to native DOM element ordering "rules". E.g. from <colgroup>:

The <colgroup> must appear after any optional <caption> element but before any <thead>, <th>, <tbody>, <tfoot> and <tr> element.

[patsissons]

what has become of this issue, is it dead?

[donaldpipowitch]

I don't want to generate noise for the maintainers, but maybe we should just open a new issue for a focussed discussion and re-evaluation of this feature...?

[gitsupersmecher]

something like this is a mitigation: export interface IMenuPorps { children ? : any; }

[dontsave]

Curious about the status of this issue as well. Being able to restrict children to specific types would be very useful

[luixo]

Chances are I misunderstood the issue topic, but I have somehow related problem.
So I have some component with its props typed:

export interface MyProps {
    extraElement: React.ReactNode;
}

class MyInnerComponent extends React.Component<MyProps> {
    render(): React.ReactNode {
        return (
            <div>
                {this.props.extraElement}
                {this.props.children}
            </div>
        );
    }
}

export default MyInnerComponent;

Now I want to calculate children in another component.

import MyInnerComponent, {MyProps as InnerComponentProps} from '...';

class MyWrapperComponent extends React.Component<{}> {
    calculateProps(): InnerComponentProps {
        return {
            extraElement: <div>Hello!</div>,
            children: [<div>First</div>, <div>Second</div>]
        };
    }
    render(): React.ReactNode {
        return (
            <MyInnerComponent {...this.calculateProps()} />
        );
    }
}

In this I get TS2322: Type { extraElement: Element; children: Element[] } is not assignable to type 'MyProps'. Object literal may only specify known properties, and 'children' does not exist in type 'MyProps'.
Probably, I should extend MyProps with something that includes a proper check for children being React.ReactNode or something, but I didn't found anything that look like that in @types/react.

[jchitel]

Recent discussion in this issue seems to be about allowing the JSX.Element type to be generic, or at least having JSX typing be more flexible. These use cases are covered by this issue, which is still open. I'd suggest moving all conversation there so that we can move toward an implementable proposal.

[mhegazy]

#21699 should cover this scenario.

[elbeg]

OK