This won't support foo = { config,...}: { marker_b = "bar"; }.

Type checking is done before merging, but after resolving mkIf (etc) wrappers.

Type checking is used to determine which type to use for merging.

(
  { options, ... }:
  {
    freeformType = attrsOf (
      if options.marker_a.isDefined && options.marker_b.isDefined then
        throw "both `${options.marker_a}' and `${options.marker_b}' are defined"
      else if options.marker_a.isDefined then
        a
      else if options.marker_b.isDefined then
        b
      else
        throw "neither `${options.marker_a}' or `${options.marker_b}' are defined"
    );
  }
)

I suspect this works and then gets rejected by some kind of a final type check: error: The option `marker_a' was accessed but has no value defined. Try setting the option

(my attempt: https://gist.github.com/SomeoneSerge/29a9e6e10cdcd3596f865c41c7042610)

That's not a type check, that's an "empty value" stub. When an option has no definitions, its value is set to a stub that throws that error.

You're probably reading the value somewhere, e.g. using config.marker_a instead of options.marker_a, or not using removeAttrs when creating __result?

On line 45 you're defining apply which is meaningless in a module, it's a feature of options.

If you're evaluating these directly with lib.evalModules instead of using a submodule option, you can do the apply manually:

(lib.evalModules { /* ... */ }).config.__result

On line 45 you're defining apply which is meaningless in a module, it's a feature of options.

Yeah I was wondering if I misunderstood the "freeformType is an option"

Here's what I meant by freeformType is an option: https://github.com/NixOS/nixpkgs/blob/ec56f7ae8cdc39eebb8a3b4aa7a3a6365c214fb5/lib/modules.nix#L213-L227

You can set it explicitly in a submodule by defining e.g. config._module.freeformType = attrsOf str, but the module system allows a "shorthand" syntax where freeformType can be defined as a top-level attr in a module and it'll do the conversion to assigning the option for you.

Here's what I meant by freeformType is an option:

Was my second guess. Thanks for walking me through these!

Awesome!

I'd highly recommend checking the marker_a.isDefined && marker_b.isDefined case too, which I assume should be an error?

The else case where neither are defined may need some special handling too: is it an error? Is it still an error if cfg is empty?