Extract not narrowing type union in mapped type passed as generic

[mikeauclair]

🔎 Search Terms

Extract, type union

🕗 Version & Regression Information

• This changed between versions 5.33 and 5.4

⏯ Playground Link

https://www.typescriptlang.org/play?ts=5.4.2#code/C4TwDgpgBAysBOBLAdgcwNIRAZwDwBUA+KAXigFEAPBAQwGNhcBrLAewDMoBvAWACgogqAG0AClBRQWIDlHxQa2OWIC6UCNQjIAJkuwIUqKAH4oogFxRkEAG4R4KyzWQh+AXwA0UfUjSF+-HSsyPpQAEYANqysYKRQBB6EABQAXpZwvhhYeEQAlKTEvAJQQSHAUOyWPoZxKe4BfCjA9uz00Ira3PyCAI4A7hBVBmj1fPwREOU06cNZOLgd-nxAA

💻 Code

type StringKeys<T> = Extract<keyof {
    [P in keyof T as T[P] extends string ? P: never]: any
}, string>

const bloop = <T,>(z: StringKeys<T>) => {
  const f: string = z
}

interface asd {
  qwe: string
}

let a: StringKeys<asd>

🙁 Actual behavior

z is not assignable to string in the bloop function

🙂 Expected behavior

z should be assignable to string, because we extract string-extending keys

Additional information about the issue

This was working just fine on 5.3.3, and continues to work when the generic is concretized (hence the let at the bottom of the example)

[Andarist]

Conditional types are tricky as they often stay deferred in generic contexts. This version could somewhat easily work though (and it doesn't):

type StringKeys<T> = keyof {
  [P in keyof T as T[P] extends string ? P & string : never]: any;
};

const bloop = <T,>(z: StringKeys<T>) => {
  const f: string = z; // errors but shouldnt
};

As a workaround, you can use this version:

type StringKeys<T> = keyof {
  [P in keyof T as T[P] extends string ? P : never]: any;
} &
  string;

const bloop = <T,>(z: StringKeys<T>) => {
  const f: string = z;
};

[mikeauclair]

That workaround works in my real codebase, so I'll cut over to that approach for now - is this a matter of the intersection (instead of Extract) causing non-deferred evaluation?

[fatcerberus]

I would guess it’s just a matter of the compiler knowing that T & string is always assignable to string regardless of what T is (by definition, as A & B is the set of values assignable to both)

[RyanCavanaugh]

Bisects to #56742