This seems more productive than what I've been up to with a self-documenting codec abstraction, which may be too ambitious for something in C++.

OK @roberth @fricklerhandwerk @infinisil the infra here is now set up! Not sure whether we want to

• Just merge it
• Hide not-yet-documented fields
• Go through and document everything

I vote to merge as is, and agree it's a useful step forward.

But I sigh loudly as well because it's adding notable tech debt. Now we have 5 (!) mechanisms for producing documentation:

1. plain markdown via mdBook
2. globals.hh, the Config object, and --dump-cli
3. pre-processing in Makefiles
4. processing in Nix language
5. this

Especially the additional document composition code path makes it substantially more messy.

The ideal end state for me would be documentation living exclusively in C++ headers (with plain markdown files for optional free-form blurbs) and processing structured output – templating and such – happening exclusively in the Nix language, decoupled from building the all of Nix. Currently it's hacks on top of hacks and it takes forever to compile.

I agree there are too many. I would like to replace the stuff we do with the interpreter inside Nix with this method, because then it will work with the store-only Nix.

replace the stuff we do with the interpreter inside Nix with this method

That would make it a lot harder to contribute to for most people, and I don't see why docs can't be built by a full Nix as a separate step, but let's discuss it somewhere else.

Discussed during the maintainers meeting today. Deferred for the time being

Our use case includes having a self-contained nix executable directly output some pretty-pretty documentation on the terminal without relying on other programs.

Personally I want to state that such style of documentation for CLI tools is worst-case - it is neither man nor help, you're forcing user to pipe output to pager and leaving out man convetions (no one interested in following them because "help files" is written in Markdown, and typical markdown user is rarely even avare of existence of manual pages, and their conventions).

Not to mention that you're forcing user to execute your command in order to get help (rather than "viewing static files" with man).

(From here: commonmark/cmark#362).

Lowdown represents one local optimum, where it's responsible for the whole pipeline. This is insufficient, as described by tribals just now.

Another optimum is one where we have

• A markdown parser
• A semantically rich document AST, suitable for man page generation
• A good markdown writer, from markdown AST => cmark
• A function from document AST to markdown AST
• A terminal layout library => a Wadler-style pretty printer
  • for printing values and expressions
  • for :doc

In this scenario, we've got rid of

• Ad hoc stringly markdown manipulation that requires a bunch of rendering and visual review
• The need for a direct markdown to terminal renderer. We reuse our existing rendering framework for expressions, and the translation from the limited number of markdown constructors to the layout+terminal framework should be fairly simple.

Do we want capable documentation tooling? If so, what are the steps to bridge the optima? Adding cmark could be step one, but we should only take it if we agree on such a vision.

For yet another optimum, we could consider minimizing the documentation related code in Nix, and generate the documentation as well as some code from a single source that isn't C++.

• Use/write an interface definition language for docs and codegen #9817

I do think we eventually want nice expression printing and such, so it feels like #9817 might be a detour.

Personally I, again, would recommend to move from Markdown (it is not Holly Graal) to something more suitable for writing documentation, eg. https://gitlab.com/ddevault/scdoc/-/blob/master/scdoc.5.scd

Please note that I'm random outsider, brought to this issue by accident, so don't take my musings for granted.

I have rare experience with Nix, and want to state that CLI tooling was one reason why I switched to Guix.

The Nix ecosystem has by and large switch to markdown in order to make writing easier, reasoning that lowering the contribution threshold leads to more end user improvements than fancy features such as table captions. Indeed we have recognized that it is not the Holy Grail, and we've decided that retrieving that artifact is not feasible, and focus on things that matter more.

However, it is a fact that markdown is of little use when writing highly structured documentation, such as manpages. It needs to be embedded in something richer, and/or it needs to be extended with custom syntax, like we do in Nixpkgs Markdown, with e.g. {option}`services.postgresql.enable` .

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/2024-01-25-documentation-team-meeting-notes-106/38792/1

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/2024-01-19-nix-team-meeting-minutes-116/38837/1

I can remove the patch now that commonmark/cmark#524 is merged --- thanks CMark maintainers! But I still need to do something about replacing lowdown.

As I suspected, cmark wasn't interested in accepting 2000 more lines of C from lowdown, which means there is not an immediate path to getting rid of the lowdown dependency.

I would still like to merge this as is, however:

• AST manipulation in C++ is a good thing to be doing, and indeed will help with other efforts like the store-only Nix CLI.

• Having two Markdown libraries might feel icky, but it doesn't actually pose any practical problem.

• @roberth's idea of replacing lowdown with a C++ pretty printing library using the cmark AST (which can include custom nodes, which we've leveraged in e.g. the Nixpkgs documentation) would make an excellent self-contained GSoC project

  • CC @ryantm who has done similar work in nixpkgs-manual-mmdoc, nixos-manual-mmdoc: init nixpkgs#108063 and whose usage of cmark/cmark-gfm inspired me to switch after lowdown wasn't working.