Discriminating a type union results in unexpected possible unsoundness

[denis-sokolov]

Bug Report

🔎 Search Terms

type guard, property checking, discriminated, narrowing, union, property existence, type assertion,

I’m sorry if this is a duplicate, I could not come up with a clear way to name this behavior and recognize it in the list of issues.

🕗 Version & Regression Information

• This is the behavior in every version I tried, and I reviewed the FAQ for entries about Common "Bugs" That Aren't Bugs and Type System Behavior

Earliest version tested: 3.3.3333
Last version tested: 4.6.0-dev.20211105

⏯ Playground Link

Playground link with relevant code

💻 Code

type Foo = { id: string } | { id: string; optional: number };
function noOptionalAllowed(param: { id: string; optional?: never }) {}

const b = { id: '', optional: 2 } as Foo;
noOptionalAllowed(b); // Error, as expected
noOptionalAllowed("optional" in b ? b : b); // No error?

🙁 Actual behavior

No error on the last line.

🙂 Expected behavior

Error on the last line. x ? b : b and b should be of the same type and behave the same (ignoring side-effects in x). On the last line of the example the type of b in the “if-true” branch includes the optional field, but the type of the entire expression "optional" in b ? b : b is { id: string } with the optional field missing!

[MartinJohns]

Your variable a is narrowed to the type { id: string; optional: number }, that's why you see the different behaviour. This narrowing is intended. Using the type assertion as Foo you prevent this narrowing.

[denis-sokolov]

@MartinJohns, thank you, this explains the case with a. I was misled by the type hint at the declaration line (new issue), but everywhere else it is consistent. This simplifies the issue considerably and leaves only the core of the question in here. Thank you!

[andrewbranch]

Subtype reduction strikes again. This example looks pretty unfortunate though 😕

[andrewbranch]

I don’t think there’s anything we can do here. The ternary does subtype reduction on the union of the narrowed type and the unnarrowed type of b (the true and false branch of the conditional), which is { id: string }. Then that is subsequently assignable to { id: string, optional?: never }. This is the same unsoundness as this aliasing: https://www.typescriptlang.org/play?jsx=4&ts=4.1.5#code/CYUwxgNghgTiAEEQBd7NgcxQLngb3ilwGdkYBLAOwwBp4AjAfl0oFcBbekGeAXwG4AUKEiwESVFWTd2IYOSjTcBIvFIVqfIYLAB7SqXi76AK3gBefIVwAiKGrJUMNuvVugAZlXLSIAT3hKXVR7Nk5uGz5BKRk5BWkLI1MhdBgsVEsYmFl5RRBtIA

so there’s nothing new here, just two things working as expected that combine to look very strange in this case.

[denis-sokolov]

Thank you for your investigation, @andrewbranch! That sounds like a very precise explanation of what’s happening. This helps me find this to be a duplicate of #46449.