Indeed. I'm glad for us to have learned Python's 20 year lesson on the dangers of an executable manifest.

The Zig language seems to have learned the lesson the other way though: https://ziglang.org/learn/build-system/

Isn't build.zig.zon the equivalent of the manifest in that case though?

Sure, but a zon file is really just an anonymous zig struct.

With the disclaimer that I know nothing about zig, I will say that Nix's flake files follow an interesting pattern where the output section is evaluated under the regular Nix interpreter and can do anything, whereas the input section is evaluated under a far stricter mini-interpreter that prohibits everything other than basic strings and attrsets (dictionaries).

A flake.nix is certainly not packaging metadata, but I do know those input section restrictions do cause occasional confusion with users who think "ah, Nix code" and only realise afterward that it's really json in Nix syntax."

All that to say, if a thing is not evaluate-able, it's probably best to represent it in a different enough form that no one could mistake it for a thing that can be evaluated.

Why not use a function to add dependencies, e.g. `add_dependency "dependency@version"` or `add_dependencies { ...table-here... }`. This is, more or less, how premake works [1]. For `lx add ...` you'd append the function call to the script.

[1] https://premake.github.io/docs/Your-First-Script

That doesn't seem very ergonomic - especially for the use cases we have in mind.

Load the table. Modify. Serialize to file.

Not too hard. Emacs does it with the .emacs file, mixing generated and manual content.

> Load the table

If it comes from an impure function, you don't know if you'll get the same result each time you evaluate it.

> Modify. Serialize to file.

And potentially lose information.