Revert auto-channel detection that produced irrecoverable edge-cases

[blaggacao]

In more advanced usage scenarios, the guarantees of this code path
are never strong enough and the error produced is non-recoverable.

Sorry to lovers of "automagic".

[Pacman99]

Can you provide an example of such an irrecoverable error, to strengthen the argument for this change.

[blaggacao]

Can you provide an example of such an irrecoverable error, to strengthen the argument for this change.

filterAttrs (_: value: value ? legacyPackages && value.legacyPackages.x86_64-linux ? nix) inputs;

any item that unintentionally falls throgh that predicate will suffice.

With the current implementation, any repo that bears legacyPackages.x86_64-linux.nix and is not nixpkgs is such a repo.

But pushing this formally incorrect code further to the edges, would do little more than just modifying the predicate for that failure scenario.

I'm in a situation a.t.m. where I can't accept looming regressions for my own "safety/sanity". Other people might not be in that situation and hence be a little more forgiving.

I think my best argument — other than "it actually breaks" — is that if we want to promote more serious and higher stake adoption of fup we should aim at a formally correct, easily auditable, feature limited version of fup.

EDIT: oh and the irrecoverable breakage of course happens when fup tries to import such repo with nixpkgs' import signature.

It's just a mismatch between asked and offered guarantees within the same code-base.

[blaggacao]

Here is another good one:

The expressivity to be able to canonically encode the intent to not use a nixpkgs type input also as a channel must be preserved for a low level lib.

Use case: don't encourage your users to use a channel you never intended to expose.

[Pacman99]

I'm fine with this, I see the edge case problems. I figure that it would just be better to not infer and just do inputs.${channelName} so that would just trigger a nix error saying that the attribute doesn't exist in inputs immediately guiding the user to the problem. I think the current inference methods are cool but extra.

I also personally don't think this kind of magic should be here anyways, but I'm alright with it if done right.

[mstrangfeld]

Just wanted to add that this currently happens when having the current version of agenix in the inputs. Was pretty hard to figure out what was actually causing it... 😅

error: attribute 'x86_64-linux' missing

       at /nix/store/i5ipw8rvj1fvdaj01bqzc0dpw9irrb4h-source/lib/mkFlake.nix:208:74:

          207|         # For some odd reason `devshell` contains `legacyPackages` out put as well
          208|         channelFlakes = filterAttrs (_: value: value ? legacyPackages && value.legacyPackages.x86_64-linux ? nix) inputs;
             |                                                                          ^
          209|         channelsFromFlakes = mapAttrs (name: input: { inherit input; }) channelFlakes;