Readonly<[...unknown[], infer L]> and readonly [...unknown[], infer L] exhibit inconsistent behavior.

[13OnTheCode]

🔎 Search Terms

Readonly<[...unknown[], infer L]> and readonly [...unknown[], infer L] exhibit inconsistent behavior.

🕗 Version & Regression Information

⏯ Playground Link

https://www.typescriptlang.org/play?ts=5.3.2#code/C4TwDgpgBAglC8UBKECGATA9gOwDYgB4BtAOjIFdsBrbTAd2yIF0AaKSm+7JgPgFgAUFCgB6EVAB6AfkGDQkKACEEUAE5oseEFFIVqtBszYcD3QcLGSZAueGgAVAIwqiAclSu2rgEaeorgGNXJigIAA9gCGx0AGc1DRx8HQBLbAAzCFUoAAk2MhITLmYQqRyoAC4obAgAN0zzUXFpWQF5BwAmF3c-Hx6gkPDI6LiUDETCIlSMrNyofMLDJl4oUuyKqtr6oUarFraoewBmLo8vXy9+0Iio2PixrR15-SLWKCnMqAAZEq-16rrVA1LM0bK07AcACwnHrnfyXQY3EYJLTEJ6cRZsd5Zb48Fa-Sr-LYWJrWYmwALAcioXCVBbYIHiACiYUgFIg6EqgVcgiAA

💻 Code

type A = Readonly<[...unknown[], unknown]>
  // ^? type A = readonly [...unknown[], unknown]

type B = readonly [...unknown[], unknown]
  // ^? type B = readonly [...unknown[], unknown]

type T1 = ['a', 'b', 'c'] extends readonly [infer H, ...unknown[]] ? H : never
  // ^? type T1 = 'a'

type T2 = ['a', 'b', 'c'] extends Readonly<[infer H, ...unknown[]]> ? H : never
  // ^? type T2 = 'a'

type T3 = ['a', 'b', 'c'] extends readonly [...unknown[], infer L] ? L : never
  // ^? type T3 = 'c'

type T4 = ['a', 'b', 'c'] extends Readonly<[...unknown[], infer L]> ? L : never
  // ^? type T4 = unknown
  // Actual: unknown
  // Expected: 'c'

🙁 Actual behavior

🙂 Expected behavior

Infer the correct type.

Additional information about the issue

No response

[Andarist]

Nice find. I find it hard to believe that some problems like this were not already reported but I can't find any issues related to this.

The problem can be shown in a much simpler example that doesn't involve any infer type parameters:

type Result = Readonly<[...string[], number]>
//      ^? type Result = readonly [...(string | number)[], string | number]

Since Readonly is a homomorphic mapped type and this goes through the homomorphic type instantiation, I believe that type arguments of this tuple should be iterated "as is". "Indexing" an element past the fixed element count shouldn't combine it with preceding rest etc

[fatcerberus]

That homomorphic behavior makes me think this is related to #56885

[Andarist]

The underlying mechanism that is responsible for both is currently the same. I think I've read some comments very recently about this other case being the limitation of the current design - it could potentially be improved but it's a known behavior of the current design and it's considered to be the correct behavior per the current implementation.

The issue here feels different because the user doesn't interact explicitly with property access like this - it's just coincidental that when the mapping is done internally it goes through getPropertyTypeForIndexType "as is" and thus it leads to this issue.

[Andarist]

I re-read the @craigphicks’s comment there and there is definitely something off there. I distinctly remember reading a comment about this behavior with variadic+trailing required elements to be by design today but i might be misremembering what i’ve read or maybe it was an outdated comment or smth

[fatcerberus]

Not to derail things too much, but...

I distinctly remember reading a comment about this behavior with variadic+trailing required elements to be by design

The relevant comment was recently linked by @jcalz and is from Anders's original PR that implemented leading/middle rest elements: #41544

Also note that indexing a tuple type beyond its starting fixed elements (if any) yields a union type of all possible element types (in this case string | number).

I can't find anything that suggests this behavior has (or should have) changed. Craig points to #49138 being fixed in 5.0.4 but that issue is about assignability, not indexing.

[craigphicks]

@fatcerberus

The fix there was introducing bi-directional matching - from both end towards the middle.
As you say

that issue is about assignability, not indexing
so it is different code. Obviously it would be "nice" if the behaviors of assignment and indexing corresponded with each other.

#56885 started out as a bug report #56883, but that was closed and reopened as a feature request. Currently it doesn't look like a feature request, and I have recommended to @13OnTheCode that it be rewritten as a proper feature request.

Hopefully that same indexing feature improvement would eventually be reflected here as well.

[ahejlsberg]

Simpler repro:

type R1<T> = readonly [...unknown[], T];  // readonly [...unknown[], T]
type R2<T> = Readonly<[...unknown[], T]>;  // readonly [...unknown[], unknown]