SemVer for TS in Practice

Showing how the recommendations from www.semver-ts.org can actually work in the real world.

Assumed Audience: Library maintainers in or adjacent to the TypeScript ecosystem, especially anyone skeptical of the practicality of the Semantic Versioning for TypeScript Types proposal.

Over the last couple weeks, I have been making some noise about the Semantic Versioning for TypeScript Types specification I authored as part of my work on TypeScript at LinkedIn. All well and good to talk about the value of that work in theory… but does it work in the real world? Well, I wouldn’t be writing this blog post if the answer were no.

In TypeScript PR #43183, the TypeScript team introduced a breaking change, which is landing in TypeScript 4.7. The change, like most breaking changes in TypeScript, is a good one: it catches bugs which were not caught previously. However, it is a breaking change! I noticed because the TS nightly tests for True Myth started failing a few weeks ago when the PR landed on TS.

This morning, I sat down with a cup of espresso at a local coffee shop and started working on the library. The actual failure case was fairly straightforward. The Maybe and Result types in True Myth implement toString() so that you get something like Just(42) out instead of "[object Object]" when calling toString() on the types. A while back, I had added a type constraint on the shared toString() function to make it type safe for True Myth’s types to call .toString() on their wrapped values directly:

function toString<T extends { toString(): string }>(maybe: Maybe<T>): string {
  // implementation...
}

In TypeScript 4.7, this stops working, because T used to implicitly get resolved to extends object further up the chain as part of TSs inference, but it doesn’t anymore, which means that writing Maybe.of(42) now produces a type which does not satisfy this constraint. As a result, where the Maybe and Result classes’ .toString() methods called the standalone toString() function, TS was quite reasonably complaining that the types weren’t compatible:

    src/maybe.ts:244:21 - error TS2345: Argument of type 'Maybe' is not assignable to parameter of type 'Maybe<{ tostring(): string; }>'.
      Type 'Just' is not assignable to type 'Maybe<{ tostring(): string; }>'.
        Type 'Just' is not assignable to type 'Just<{ tostring(): string; }>'.
          Type 'T' is not assignable to type '{ toString(): string; }'.
    244     return toString(this);
                            ~~~~
      src/maybe.ts:53:17
        53 class MaybeImpl {
                           ~
        This type parameter probably needs an `extends object` constraint.

I started by just that kind of type constraint to the type parameters for Maybe and Result, and that worked… but then there was another issue: it meant that I had to add type annotations at a couple points in the test suite to tell TS how to resolve the types in a way that was actually safe. That would work, but it would also be a breaking change! I got as far as writing a commit message which explained it before realizing there was no way I could justify making this and not calling it a breaking change.

One option would be to say that the current version only supports up through TS 4.6, and bundle it into a breaking change as part of the already-planned upcoming major release of the library. That feels bad, though: there is no other reason not to support TS 4.7 with the current version of the library. More, the result would be making the experience of working with those cases worse.

So I took a step back and thought about the problem a bit more. It occurred to me that if, instead of adding constraints to the class methods, I removed them from the standalone function type, the problem would go away. To make that work, I would need to make the implementation of the function slightly more generic, to handle any type at all, whether or not it has its own .toString() implementation… but that sounded like a net win. So that’s what I did: I re-implemented the toString() functions for Maybe and Result to do the right thing regardless of what the value they were dealing with was:

  • If the wrapped value has a toString() method which does indeed return a string, return the string it produces.
  • If the wrapped value doesn’t have a toString() method, use JSON.stringify() on the wrapped value.
  • If the wrapped value has a toString() method but it returns something which isn’t a string, use JSON.stringify() on the wrapped value instead.1
  • If the wrapped value has some other property named toString on it which isn’t a function at all, use JSON.stringify() on the wrapped value.

The result is an implementation which is much robust and more useful than what we had before — and it works on TypeScript 4.7. I published it as True Myth 5.3.0 just a few minutes ago. (Now, to see about fixing the other three libraries I maintain which have the same issue…)

The Semantic Versioning for TypeScript Types proposal works for real libraries in the real world. Following the guidance we came up with really does make it possible to insulate users from breaking changes in TypeScript. It’s a good feeling!

Thoughts, comments, or questions? Discuss on LinkedIn, Hacker News, lobste.rs, or Twitter!


  1. You could make an argument that this is ignoring the design intent of the user who implemented a method like toString() { return 123 }. Fair enough. The only type safe option would be to call JSON.stringify() on the result of that, though: 123 isn’t a valid string! If you use True Myth and think we should use JSON.stringify() on the result of any .toString() implementation, rather than calling it on the original object, we’re open to being persuaded,s o open an issue or discussion! ↩︎