I think you're missing that `useWindowWidth()` could have more than one `useState()` and thus you'd need to somehow compose Symbols. Which is what the next section is about.
https://gist.github.com/politician/5f03c169a4119a63abb785b1c...
EDIT: I still think it's a hard design problem and the concerns in flaw #8 are _very real_ to a broad range of developers. When you're just trying to get it done, proper bookkeeping of Symbols and injecting them into hooks composed of other hooks might cross the line. It's a shame that relying on call-order indexing is the solution because it's magical. But at the end of the day, engineering is about trade-offs. Time will tell whether deeply nested composition of automatically-managed hooks was a good feature to expose.
I've been thinking about this a lot since Hooks were introduced, and I'm increasingly of the opinion that it isn't that magical. Order of operations is incredibly important in the functions we write, especially in a language like JS that makes no effort for strict "pure" side-effect free functions. The order of a console.log or a return versus an increment matters in JS. We write a lot of procedural code in JS where order matters (a lot) already.
In that matter, Hooks can just melt into the "procedural" background of JS.
That said, I still feel like I want a better solution than "lint errors" for things like accidental branches of a Hook. I don't have any more of a proposal for how that would work or what that would mean than the article here, though, unfortunately.
The best I can come up with is Sweet.js (hygenic macros), but can you imagine what outrage that would provoke?
I keep trying to figure out if there is a way to push it into a type safety problem in Typescript, at least. A compile error would be preferable to a lint error, even if not everyone uses a typing compiler like Typescript.
If there were some way that you could get use* functions to narrow to `never` in any branch or branch-like position, that would be cool. Unfortunately, I can't think of a natural way in the type system to do that that would work in even a plurality of cases. Which drops back to it needing to be a context-sensitive linter rule.
Basically, lint, but at transpile time.
That said, the presentation of the third and forth flaws seem a bit weak. useState accepts a key as the first parameter, but in both of the composite functions that input parameter disappears. That looks like a refactoring error.
From a design standpoint, if `useState(symbol)` is acceptable (and I'm not saying it is), then `useWindowWith(symbol)` would likewise be acceptable in that it's not adding any more requirements to the interface than the vanilla version.
Proper bookkeeping of symbols resolves the issues here. However, if the design objective is to reduce the effort on the developer to do their own bookkeeping, then call-order indexing does make sense.
(Not to mention that Symbols are not supported by IE11 which remains an important target for bigger companies.)
I'm not sure there's a good way out. When you look at solutions like Protocol Buffers, they just bite the bullet and require the developer to supply the indexing. If JavaScript had something like Go's iota, then you could imagine using an enumeration to supply the indexing without requiring everyone to type 1,2,3,4... etc. But it doesn't so, that's wishful thinking. Nevertheless, the react codebase itself does contain a giant list of assignments of Symbol() || number to constants, so it's a pattern you're already aware of.
A tough nut to crack.