Design Meeting Notes, 4/30/2024

[Daniel Rosenwasser]

TReturn and TYield for Iterators and Other Types

#58243

• TReturn is the type of values returned by a generator.
• TYield is the type of values yielded by a generator.
• When we added these, we only added them to certain types like IteratorResult
  • Since then, we have a PR that adds type parameters to Iterator, IterableIterator, AsyncIterable, etc.
  • What this has meant since then was that if you ever try to grab value off of most IteratorResults, then you'll get any.
    • Good as a fall-back, but usually not what people want.
• Changing this is a breaking change - so how can we avoid breaks?
  • First, each of these types should have a default value for TReturn and TYield that is any.
    • But when a user writes Iterator<number>, TypeScript should preserve that type as-is.
  • What is the level of breakage as is?
    • Almost every case is an example of the type system not knowing that size > 0 implies value is not undefined.
    • There's also a lot of breakage around DefinitelyTyped's ExpectType and ExpectError utility types.
• Several new errors introduced. Previously, people expecting .next() to return a T now get T | void.
  • e.g. if (something.size === 1) { something.values().nexT().value.getTasks(); }
  • Why is it void and not undefined?
    • Should it be undefined?
    • We really don't want to relitigate all the issues around non-returning functions returning void instead of undefined.
    • It's possible we carve things out for generators.
  • Should we allow postfix ! to get rid of void?
• How many upvotes does resolution for this issue have?
  • Old issue had around 15 upvotes.
  • This is so breaky that most of this has to go under a switch.
• What happens if we don't do this?
  • People trying to thread these into the built-in iterator types.
  • If you have a generator, you can now write myGenerator().map(x => x * x) and you want the TReturn preserved.
• If the worst thing you have to do is add a !, is that so bad?
• Feel like there should be an approach where this is flagged.
  • Could be like strictBindCallApply.
  • We fret about breaks for much more widely-requested features.
  • But the people it breaks are using iterators, probably want to use the feature.
  • The flag that this conceptually aligns with is noUncheckedIndexedAccess.
    • Ah yes, .next(), the famous indexed access syntax.
    • Okay, fine, a separate flag with an equally bad name (noUncheckedIteratorAccess or noUncheckedIteratorResult).

tsc --init Updates?

• We noticed tsc --init adds "module": "commonjs" - which our docs said to never ever ever do.
• Okay, what else do we need to fix in --init?
• Ryan asked on Twitter what people would like to change.
• People said lots of stuff
  • Big one that resonates: remove the wall of text comments
• Are we okay wiht tsc --init generating a file today?
  • Uh yes.
• Okay, that's a good start.
• What else?
  • Make it minimal
  • People seem opposed to interactive.
• Okay, what are things that need to be hashed through?
  • target: esnext?
  • module - node16 or bundler?
  • verbatimModuleSyntax?
  • skipLibCheck?
  • We need to hash through these.
• Should we still have some comments on defaults?
  • Maybe.
  • Ryan also wrote a wizard, someone else wrote something too.

[Martin Johns]

Related: #46150