--strictAny

#24423

• Talked about this last time, but no implementation.

• Now we have an implementation and data.

• Came from conversation with the Hack team.

  • (which, funny, was inspired)

• Background: problem today is that any is assignable in both directions.

  • Problem is often that a user wrote any in a declaration file because they were lazy or just scaffolding things.
    • But that means that any consumer of the .d.ts file has no way to defensively avoid assigning any to anything else.
  • This PR removes the "bottom-ness" from any - makes any a top type.

• Tried running --strictAny on DefinitelyTyped

  • Not on tests - tests have too much breakage.

  • Some things got broken - declaration files were

  • Much of the time it's just declaration authors being a little lazy.

    • Instantiating generics with any
      • But hey, even we're lazy! We instantiate the prototype property on generic classes with any across.
    • Using any in derived classes.

  • Why is any not assignable to Record<string, any>>

    • Ehh, maybe it should be.

  • There are some issues with signatures like (x: any, y: any, ...zs: any[]) => any.

    • Now any is subject to contravariant checking.
    • Again, can tell users to use never instead of any.

  • Issues with SVGElement not being assignable to Element.

    • DOM in general isn't sound, so... :(

  • Then there's the "covariant event emitter pattern".

    once(event: string, listener: (...args: any[]) => void): this;
    once(event: "close", listener: () => void): this;
    once(event: "drain", listener: () => void): this;
    once(event: "error", listener: (err: Error) => void): this;
    once(event: "finish", listener: () => void): this;
    once(event: "pipe", listener: (src: Readable) => void): this;
    once(event: "unpipe", listener: (src: Readable) => void): this;

• never?

  • never isn't very intuitive
  • We've tried to help users reason about it as a value that can never occur.

• The hard-to-solve problems really come from any in contravariant positions.

• Tried running this on the compiler.

  • Finds some questionable things.
  • commandLineParser is sloppy.
  • Calls to Array() always return an any[].
    • Probably should move the generic overload to the top?
      • Does that work right?
  • isArray is a predicate on Array<{}>
  • Functions with overloads in derived classes with an any in the middle.
    • No.
  • Did this catch any legitimate bugs?
    • commandLineParser had a lot of questionable behavior.
  • Also, like prototypes of generic classes, there are some places where any never occurred but the type checker - for example, erased signatures.

• Really don't like as never

  • None of us does.

• Really two usages of any that have been conflated:

  1. I was lazy.
  2. "Ehh, shut up you stupid type-checker."

• The answer may just be to tell people to go through their files and return unknown.

• Have a list of places any just appears:

  • Type parameters of prototype property in generic class
  • String index signature in implied type of binding pattern (function f({...x}))
  • Implied type of omitted element in array binding pattern (function f([,,x]))
  • Catch clause variables
  • Shorthand ambient modules
  • Union signature with only some explicit 'this' types
  • Types of properties in object from inference to 'keyof T'
  • Type of 'super' in object literal expressions
  • 'yield' expressions

• Not sure about this feature.

  • Part of the reason we went in this direction is because any has become so conflated and prevalent.
    • Library authors don't always know how users intend to use these things.

• Some users really want a more capable unknown.

• Some users really want a dynamic instead of an any.

• What about per-file strict-ness?

  • May cause a lot of user confusion.

• If we did this all over, we could have considered unknown to be treated as "effectively any" for usage, still not assignable.

• Can't imagine we would use this in --strict

• Would we use this feature?

  • Yes.
    • But we're also the freaking TypeScript team.

• We'd like to get feedback from users based on the current behavior and see if it's truly paletable.

"Loose mode" in .ts files vs. TypeScript in .js files

#22665
#23906

• Hard to reason about the semantics
• Webpack migration to checkJs has been instructive
  • But we're not trying to solve that in these proposals.
• Two things in checkJs that make things different.
  • JSDoc comments
  • Change how we check things.
• So maybe instead of a "loose mode", it's more like allowJsDoc.
  • Also errors when your JSDoc and TypeScript annotations disagree! :D
• So is tooling not enough here?
• Now we actually have rename file support - maybe when you rename a .js file to a .ts file, we should offer to migrate the file.