Nanyx
PlaygroundPackages

Info view and library browser

What is still open about the info view, doc comments, the library browser and search by type

Doc comments, the info view, the library browser and search by type are described for users on VS Code extension, Basics and The nanyx command, and how they are built in Language server. This page keeps what is still open and the open questions. Work is tracked under the lsp label.

Still open

  • Doc comments by symbol. Documentation.Reader parses the file a symbol is declared in for its doc comments. Recording them by symbol in TypedModule would let the info view and the browser read them as they read types.
  • A tag’s and a record field’s doc comment. A tag has no symbol, so its doc comment is attached but not shown anywhere; nor is a record field’s on a member completion.
  • A memory block is part of the scope it is in. A block can name its Memory, memory { context m -> … }, so it may become a scope of its own in the info view, since it is where a scoped context lives.
  • A module’s types. The info view lists no module’s types, so it shows no kind; a library browser that lists them should show which are nominal, protected or private.
  • A pinned position does not move with edits above it.
  • Derived instances in search. Search by type does not look for a context instance derived from a type’s parts, such as a list’s ToString, so such a match is shown as still needing it.
  • Other packages. The browser and search list only the standard library. The packages a workspace depends on could be listed the same way.
  • Completion by type. Completion could use the same ranking as search: offer the functions whose type fits the hole the cursor is in.
  • Bigger libraries. Each query is a unification against every export, which is fast for a library the size of the standard library. For a much bigger one, the index can be keyed by the type’s shape (its result’s head, its number of inputs) first.

Open questions

  • Should the info view show types as trees, so each part can be hovered and followed to its definition, as Lean’s does? That changes the request from strings to a typed tree, and the type printer would need to emit one.
  • How loose should search be? Hoogle also matches a query in which an input is curried, or where a list could be a maybe. Each looser rule finds more, but buries the exact matches further down.
  • Should the browser also be a page on the docs site, generated from the same data, so the library’s reference is readable without an editor?
  • Should user widgets come later? In Lean, a library can ship its own UI into the infoview. That is a long way off here, but nanyx/infoView returning sections of items leaves room for it.
  • Should other editors get the info view? The request is plain LSP, so another client can show the same thing; a text rendering in an output buffer would do for Neovim.