Flake fragments should prioritise the top-most attribute.

[L-as]

Describe the bug

Doing nix build x#y, nix eval x#y or similar, will cause Nix to evaluate packages.<system>.y before y.
This is problematic in my case, because packages uses IFD, whereas y does not.
I want to be able to evaluate y, which does not use IFD, with --no-allow-import-from-derivation.

Steps To Reproduce

1. Have packages use IFD.
2. Define top-level attribute y.
3. Do nix --no-allow-import-from-derivation eval .#y. This should fail.

Expected behavior

Nix should evaluate the top-most attribute before trying the one in packages and similar.

nix-env --version output
nix (Nix) 2.5pre20211007_844dd90

[lilyball]

Copying #5657 (comment) here as it really applies to the issue, not the implementation:

---

This is going to make nixpkgs#lib refer to the top-level lib in the flake instead of the lib in the package set. At the very least this means any overlays that update lib wouldn't apply to nixpkgs#lib after this change; I don't know if there are any other practical differences in this case besides the fact that the top-level lib has a nixosSystem function. But it's an example of a potentially surprising outcome of this change.

More generally, if nixpkgs#hello resolves to nixpkgs#legacyPackages.aarch64-darwin.hello then nixpkgs#foo should resolve to nixpkgs#legacyPackages.aarch64-darwin.foo. I shouldn't have to know the complete set of outputs of a flake to predict where arbitrary fragments resolve to. Also as a reproducibility issue, a flake adding a new top-level output shouldn't break existing flake references (e.g. nixpkgs adding a top-level hello output shouldn't break nixpkgs#hello).

I feel like perhaps a better approach is to have a separate syntax for "don't apply any prefixes to this fragment". For example, something like nixpkgs##lib.

[thufschmitt]

More generally, if nixpkgs#hello resolves to nixpkgs#legacyPackages.aarch64-darwin.hello then nixpkgs#foo should resolve to nixpkgs#legacyPackages.aarch64-darwin.foo. I shouldn't have to know the complete set of outputs of a flake to predict where arbitrary fragments resolve to. Also as a reproducibility issue, a flake adding a new top-level output shouldn't break existing flake references (e.g. nixpkgs adding a top-level hello output shouldn't break nixpkgs#hello).

That’s true, but it also play the other way: if nixpkgs#foo resolves to nixpkgs#foo, then nixpkgs#hello should resolve to nixpkgs#hello. Which isn’t the case.

An “interesting” side-effect of the current behavior is that if my flake has both a toplevel foo attribute and a packages.{mySystem}.foo attribute, then there’s no way (to my knowledge) to nix eval the toplevel foo attribute

[lilyball]

@regnat Despite my attempt to move the discussion here, it has continued in the PR (see #5657 (comment)).

[stale]

I marked this as stale due to inactivity. → More info

[tomberek]

I feel like perhaps a better approach is to have a separate syntax for "don't apply any prefixes to this fragment". For example, something like nixpkgs##lib.

I'm looking at an approach like this: master...flox:nix:disable_attrpath_resolution

If this is preferable I can add tests/docs/etc. It's a bit odd because we're using an invalid URL, but it is also for the unique situations when one wants to express "don't perform automatic attrPath resolution".

[roberth]

Why would eval implement any rules if it doesn't come with its own flake attribute? eval --package, sure, but this is weird.

URL

We need to stop pretending that they're universal.

[ysndr]

Something perhaps more compatible might be

flake#.hello
      ^

It's not a valid nix attrpath but it allows for the attrpath to be parsed as ["", "hello"] using the existing separator.
An attrpath starting with "" could then be considered a "rooted" attrpath enforcing fully qualified access while interpreted access is still possible with the paths we know

[ysndr]

In fact that already works by adding

"" = self;

to the flake output, which can be injected by nix.
I doubt that "" is commonly used as a flake output...

[tomberek]

Edit: even newer: master...flox:nix:disable_attrpath_resolution

Supports nix eval .#. --apply builtins.attrNames