Hacker News new | past | comments | ask | show | jobs | submit login
HypeScript: Simplified TypeScript type system in TypeScript's own type system (github.com/ronami)
202 points by kerneloops on July 28, 2022 | hide | past | favorite | 82 comments



“I'm not about to start using it for real projects, it's really just a thought experiment about how nice JavaScript could be with an alternative syntax.”

- jashkenas 2009

https://news.ycombinator.com/item?id=1014225


"Hello everybody out there using minix -

I'm doing a (free) operating system (just a hobby, won't be big and professional like gnu) for 386(486) AT clones. This has been brewing since april, and is starting to get ready. I'd like any feedback on things people like/dislike in minix, as my OS resembles it somewhat (same physical layout of the file-system (due to practical reasons) among other things)."

https://www.cs.cmu.edu/~awb/linux.history.html


That message's 30th birthday is in 2 days


> Please note that this project is meant to be used for fun and learning purposes and not for practical use.

https://github.com/ronami/HypeScript#introduction


I still use coffeescript over javascript because I feel like I can be as expressive as I want. I understand why most left for typescript (type safety), but I feel like everyone just gave up on reducing boilerplate for things in javascript. Switch statements, list comprehensions, array slicing, better equality operators... a lot of things that made coffeescript a great experience have never carried over.


I never tried CoffeeScript since nobody pays me for it, though I am curious about ReasonML as an alternative, there's a Neovim front-end[0] coded in Reason that compiles natively[1], and supports existing VS Code plugins from the VSCodium plugin repository[2] which I still have yet to look at how the heck they pulled that bit off, but it is pretty interesting.

My thinking is, not only can I do amazing front-end stuff on the web, but I can do front-end GUI applications that aren't reliant on Electron thanks to OCaml magic.

[0]: https://github.com/onivim/oni2#introduction

[1]: https://github.com/revery-ui/revery

[2]: https://open-vsx.org/


I loved it to. I'd love to make a typed coffeescript variant, which transpiles to TypeScript rather than JavaScript. My pet name for this is `tycoscript`.


I like the idea, but I'd recommend the name "ToffeeScript".


This already exists, see [1]. Development has a bit stalled, but afaik the author wants to keep pursuing it. There's even a Discord server for this kind of project, see also this GitHub thread [2] where CS+TS is being discussed in general. Alternatively, you can code in Coffee with JSDoc type annotations, yielding a near-TypeScript experience in both IDE and output, albeit the annotation syntax can be a bit annoying, ironically.

[1] https://github.com/edemaine/coffeescript/blob/typescript/tes... [2] https://github.com/jashkenas/coffeescript/issues/5307


I used LiveScript for some time, but I went with the crowd eventually.


> Switch statements

JS has had switch statements for a long time. Do you mean switch "expressions" that can be used inline? It's not pretty but JS has long supported that too using either eval() "hacks" or IIFEs: https://stackoverflow.com/questions/32451049/is-eval-the-onl...

Do you mean pattern matching? There are two Stage 1 TC39 proposals with different approaches to it:

More hands-on/direct: https://github.com/tc39/proposal-pattern-matching

More indirect: https://github.com/tc39/proposal-do-expressions

> list comprehensions

There are several libraries with strong Iterator support today that while not syntactically gorgeous do great jobs at allowing you to write simple list operations. Off the top of my head are IxJS and iter-tools, but also classic mainstays like up to date versions of lodash and Ramda.

There's a Stage 2 proposal to bring more of them into "the standard library": https://github.com/tc39/proposal-iterator-helpers

There's also a Stage 2 proposal for an operator to make the syntax generally nicer no matter the library: https://github.com/tc39/proposal-pipeline-operator

> array slicing

Array.prototype.slice() was added back in ES2015: https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...

Array.prototype.at() was "just added" in ES2022 for unary (single argument) "slice" (ability to use negative indexes): https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...

If you prefer syntax to function calls, there's a Stage 1 proposal still live for that: https://github.com/tc39/proposal-slice-notation

> better equality operators

That ship has probably sailed. Most projects at this point just have linter rules to standardize everything to the more reliable "triple equals" (===, and !==).

There are proposals for more structural equality functions in the standard library, though, for instance: https://github.com/tc39/proposal-array-equality


Man, I know this wasn't really the intention of the comment, but I just read through that thread and got a little nostalgic. Crazy that was like 13 years ago. Sheesh.


I also wonder how much of his 'tiny language' filtered through in ES6.


I wouldn’t say they are related but a lot of things he was doing back then ended up in ES6. Great minds think alike I guess.


Ah ah, another instance of "just a hobby, won't be big and professional like gnu".


Is that not referring to Coffeescript? I don't see the relevance.


This is epic. There’s a parser in there.

Edit: this is explained a bit here [1]. It uses Template Literal Types[2].

1 https://blog.joshuakgoldberg.com/type-system-game-engines/

2 https://www.typescriptlang.org/docs/handbook/2/template-lite...


I'm impressed by how readable the live demo is.

I have to say, the fact that you can do something so... obscene, but in a way that's actually surprisingly readable, speaks highly of Typescript's type system and design.


Please tell me this means you can do dependent types! I have no idea how this works, seems like magic and seems like anything is possible.

Maybe "dependent-but-stringly typed" is possible?


TypeScript has a very limited form of dependent typing made available through `typeof`, but it does not have dependent typing as e.g. Idris does. However, in general, these sorts of demonstrations do nothing to show that a language can be dependently typed. These sorts of demonstrations show that a type system is Turing complete. However, a type system can be Turing complete without being dependent. For example, type-level integers might be completely different from run-time integers with no way to things from the latter to the former.


Sure, but plenty of type systems are Turing complete. What makes this demo impressive is the (unique?) feature of string literal types and template literal types, which lets you operate on text (like, actual text, with no weird type-level encodings).


Well, certainly string literal types have been around in GHC Haskell for quite a while (7.10.x). They've also been a thing in Scala 2.12.x although the 'implementation' was some sort of weird type checker 'hack'. I believe Scala 3 supports them natively.

I can't speak to the ergonomics of actually implementing anything which uses them internally since I've never really had much use for them outside using the surface-level API of a couple of libraries.


This is insanely cool voodoo. But... I may be lacking in imagination here... I'm trying to think of how I'd actually use it. Maybe it'll come to me in the middle of the night 8)

I do really wish TS had actual inbuilt reflective type checking along the lines of this: https://github.com/Hookyns/tst-reflect ...I can hazily see some kind of monstrosity that could come from linking these things together hah


This linked project is very cool! I wonder if this path is expressly a non-goal of the TypeScript team though…?


Yeah, giving types a runtime footprint is on the official list of non-goals for the TypeScript project: https://github.com/Microsoft/TypeScript/wiki/TypeScript-Desi...


Interesting.

> Instead, encourage programming patterns that do not require run-time metadata.

I can understand why. But there are times when any alternative pattern is a worse choice. Then it's left to the coder to hack together type checks however it's convenient. Does the Dog class really need a static variable typed to Dog|Cat... ? It would be nice if instead of a dozen nonstandard ways to do that, some metadata could be emitted if you wanted to. TS does weird stuff like emitting functions to define enums, for instance.


> Does the Dog class really need a static variable typed to Dog|Cat... ?

It doesn't- because classes are a native JS concept and not a TS concept; they already have the metadata to distinguish them

> TS does weird stuff like emitting functions to define enums, for instance

Not sure what's meant by this; TypeScript does introduce enum syntax which technically generates runtime code (one of the few instances), though it's just very light sugar over constants, and using it is also sorta advised against these days


> classes are a native JS concept

Recently, with ES6. But JS still lacks a way to get the type of an instance. You can check it but you can't simply get it, as in `new typeof(this)`. So you have to store it somewhere in the instance or in the type, as you would for a manual typeguard. And much of TS usage is still not for generating ES6 code. TS overlays boilerplate to mimick classes on ES5 targets. As it mimicks enums. I'm just saying I don't see why it couldn't mimick type reflection as well.


Am I the only one that feels uncomfortable with all that usage of strings in TS’s type system? Why not use pure literals instead of string literals? This is a genuine question, I’m trying to find out what the pros and cons where in the decision making process.


In the context of type definitions, strings aren't really regular bare strings.

I.e. if I write

  type foo = {
     bar: "Vodka"
  }
I'm guaranteeing that the value of 'bar' on any object of that type must be "Vodka" - the type of bar is not string, it's literally "Vodka" because string values can be types.

This seems a little obscure and pointless, but you can put these string types in a union, and enforce that a value must be one of many values.

  function doTheThing(color: "RED" | "GREEN") {
      //do stuff
  }

  doTheThing("RED"); //ok
  doTheThing("GREEN"); //ok
  doTheThing("ORANGE"); //doesn't compile, because the type of the param *isn't* string, it's "RED" | "GREEN"
You can also define a type like

  type Mountain = { name: "EVEREST", height: 8848 } | { name: "K2", height: 8611 };
and then at compile time know that if mountain.name === "EVEREST" then height === 8848 because the types of name and height aren't string and number, they're "EVEREST" | "K2" and 8848 | 8611, and the compiler is smart enough to work out one based on the other.

==============================================

For extra context - a lot of this is for interoptability with Javascript code - you want to call some Javascript function with a stringly typed enum, and enforce only passing in valid values, but the Javascript code still just deals with strings.

A lot of Typescripts kind of insane flexibility is so you can introduce type safety to all sorts of dynamic Javascript code, without having to make sacrifices on the dynamic-ness of it.

Turns out this flexibility is actually kinda awesome to have in general even when you're not trying to refactor an existing JS codebase.


And don't forget about [template literal types](https://devblogs.microsoft.com/typescript/announcing-typescr...)!

```

type Color = "red" | "blue"; type Quantity = "one" | "two";

type SeussFish = `${Quantity | Color} fish`; // same as // type SeussFish = "one fish" | "two fish" // | "red fish" | "blue fish";

```


To add to this, the reason it’s important for TypeScript to support string literals as values instead of using enums or something new is interop with existing JS. For example, think of an event handler where you pass the event name as the first argument; you can ensure at compile-time that an invalid event name isn’t passed.


Or more importantly, varying the event type based on the event name.

e.g. addEventListener("click", (MouseEvent): void) versus addEventListener("input", (InputEvent): void)


That's really interesting.

    doTheThing(readFromUser())
What does the compiler do in cases where you have strings coming in like this?


It gives a type error, because strings are not assignable to string literals.

https://www.typescriptlang.org/play?#code/FAEwpgxgNghgTmABAM...

However, if you check or assert that the returned value is one of the accepted literals, compiler accepts it:

https://www.typescriptlang.org/play?#code/FAEwpgxgNghgTmABAM...

This is one of the classic examples of flow-sensitive typing in TS.


it'll tell you that `string` cannot be assigned to type `"RED" | "GREEN"`, and so you'll need to write a function that refines the type correctly, e.g.

    function isValidInput(input: unknown): input is "RED" | "GREEN" {
      return input === "RED" || input === "GREEN";
    }

    const input = readFromUser();
    if (isValidInput(input)) {
      doTheThing(input); // no type error
    }


It's an error, as the plain `string` type doesn't satisfy the union. You'll need to perform type assertions to make the return type down before passing it as a parameter.


The academic term for such types is "singleton types". They allow for some quite mind-blowing possibilities as you've shown.


As others have pointed out, they're not really "strings" per se, or at least can be a lot stricter than a string might seem when used right. We use that constantly for useful things.

The bigger reason why it is this way is to ensure you can write type annotations for stringly-typed Javascript. While Typescript is it's own thing really at this point, it still is very focused on making it possible to type-annotate JS code.

    type Event = 'onClick' | 'onHover'

    // some stringly typed javascript can now be nicely typed
    something.triggerEvent('onClick', someData)
It lets you do some really nice things, and is quite strongly typed despite how it looks at first glace, while still letting me write some strongly-typed definitions for existing Javascript code that was very much __not__ written with types in mind.


In a similar case TypeScript even knows the signature of the event handler for addEventListener, and what type of Element document.createElement('div') returns. Someone even built typings for querySelector that parse the selector and returns the correct element type.

In general, however, I still think APIs written in TypeScript first are a lot cleaner type-wise than typings retrofitted onto a JS API. And having TypeScript as an afterthought these days doesn't seem like a good path to go anymore.


> In general, however, I still think APIs written in TypeScript first are a lot cleaner type-wise than typings retrofitted onto a JS API

Oh absolutely. Writing JS-first just seems like a lost cause. But there are still enough internal libraries and such written in very JS JS floating around that I come across at work that we need to interface with that I appreciate how much effort the TS team still puts into making sure I can drag those libs kicking and screaming into 2022.


I found it interesting that I would never use strings this way in JS, but TS seems to encourage it. I definitely would not want to strip typing out of my TS code and try to maintain it as plain JS.


What would you use for a field called "stage" that could be set to 'dev' or 'stage' or 'prod'? Or other enum-type things?

I've seen pure JS enums like {'dev' : Symbol('dev'), 'prod' : etc} but I never use them because you can't send them over the wire.

You could use a type like {dev:true} | {prod:true} | {stage:true} but you still wind up comparing strings.


> {dev:true} | {prod:true} | {stage:true}

This can't be accessed safely. Because all value in typescript can be sub type of how it typed.

It means code following will pass.

    var a = { dev: true, get prod() { throw Error() } }
    var b: { dev:true } | {prod:true} | {stage:true} = a

And there is no guarantee that access any of these fields is safe.


That did not compile for me with my TS settings;

> Type '{ dev: boolean; readonly prod: void; }' is not assignable to type '{ prod: true; }'

It's still a bad solution though; it allows nonsense values like {dev:true, prod:true}.


That is because I forgot to assert the true as literal true, so it defaults to be boolean (where both true and false are allowed)


I think it's kind of cool. It reminds me a bit of C++ templates, which are "compile time duck typed". The benefit is that you get lose typing constructs that are evaluated strictly and at type check time, which is kinda nutty. It means that you can lean really heavily on the compiler.

Sort of like using Python to generate Typescript, as random example - my Python code doesn't have to typecheck, but its output does, so I can do absurd shit in Python and still feel good about the output.

One example is something like this: https://www.typescriptlang.org/docs/handbook/2/template-lite...

These are types that are built from string templates. Since strings are loose and can be manipulated in crazy ways like appending, we can now manipulate types in the same way. We can write types that are themselves parsers.

So idk if that's good or not, the downside in C++ is that TMP errors are fucking insane, but the upside is that I can have a function that says "pass me something and I'll call "iter" on it and I don't care what that thing is".

It also feelsy kinda more "typey". Types are just values. Types are just strings or numbers or whatever. They're things, with the constraint being that they must exist concretely (or be inferrable) when the type checker runs. No distinction between types and values seems like it's the ideal.


Which allows for things like this type that implements a simplified SQL query parser checked against a provided 'database' object:

https://github.com/codemix/ts-sql

This project was my go-to "nifty but pointless" example for TS string literal types before this article :)


Yeah I saw that was cited and I love it. It's insane but super cool and I could imagine the approach being used for something similar to https://crates.io/crates/sqlx although there's probably better ways to do it lol


Yo dawg, I heard you like Typescript ... you know the rest.


  interface TypeScript {
    typeScript: TypeScript;
  }


This article is intriguing, I liked the playful tone, particularly in light of the subject. I’m hoping the talk gets streamed to YouTube.

This also makes me interested in the formalization attempts I have heard are ongoing for TS’s type system. Anyone with good paper/videos on that development feel free to drop a link.


Which article were you referring to?

The link in the post goes to a GitHub repo.


This guy's like, "I see your Lisp in Lisp and raise you..."


Incredible, one would expect this to be theoretically possible since Typescript's type system is Turing complete, but it is certainly different to see it done in practice. Wow!



Tetris is challenging because TS escapes newlines in type output, so you'd end up with a bunch of\nwhich makes it\nhard to read the output. It could probably be done using arrays instead, but that gets truncated so you'd have to have a very small "screen"


For people used to working in TypeScript who suddenly find themselves having to work on a vanilla js app... and don't wanna use JSDoc?


I'd convince the team to switch to typed or get out of the project


Do everything you can to introduce TS incrementally. JSDoc simply isn’t enough.


Use TS locally


You can secretly have super powers.


What does this mean?


I think OP means that if your project / job doesn't support typescript, you can write .ts files on your local machine and then compile them to .js. I think it would be very challenging though to be the only person on a project using TS though.


Yeah that would be weird.

Using tsc to type check javascript - with JSDoc type annotations - works fine. You're really not missing much from "real" typescript, and you can save yourself a transpile.


It doesn't really work fine beyond a single file.

I tried the experiment of .js files driven by typescript in JSDocs for around 10k lines and there is a night and day difference in what you can express, how and reuse it at the type level.


I'm experienced in doing this too.

Name one thing you can't do.

To clairfy, if I am sharing types between modules I tend to use .d.ts files. Still no transpiling needed.


until you need to type the shape of objects, and want to share that shape with other files. You can't import JSDoc types from other files. At which point, you would be forced to create `types.d.ts` file or similar. Might as well be writing typescript.

The limitations are too much for complex projects at scale. I strongly believe this is why Closure types didn't take off.

That and Google's lack of developer relations around the project


> You can't import JSDoc types from other files

Yes, you can. Using `import('./someFile').SomeType`

Though I do agree, JSDoc works, but isn't as nice as directly writing types in a .ts file


You can use types written in a .d.ts file in javascript. You reference them with the JSDoc annotations.

And no it's not the case you may as well write typescript at that point - because you're still using the actual language the browser runs, and not transpiling.


Can we read the source file from filesystem, then pass into type checker like this ?


that sounds increasingly like a build-time macro system...


This is mindblowing


I love TypeScript but its complexity is getting ridiculous.


What complexity do you feel is unneeded?


> Please note that this project is meant to be used for fun and learning purposes and not for practical use.


Oh thank God. There aren't nearly enough ways to write javascript.


There's no Javascript (or any other runtime) code in this at all. It all happens at compile-time.


But you probably are going to be running this as JS eventually


There's literally nothing to run. The source file is just a bunch of type declarations.


If you transpile it to js, you'll get empty file.


The real lesson isn't the output of our build system, it's the types we checked along the way.




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: