builtins.fetchTree behaves differently if uri or attrs are parsed

[domenkozar]

While implementing builtins.fetchTree for https://snix.dev/ I've noticed revCount is missing when it shouldn't:

nixpkgs on  master [$] took 3s 
❯ /nix/store/pkw2fif96sahwnnkfxsmdzs0dl1nr5ln-nix-2.26.3/bin/nix-instantiate -E "builtins.fetchTree git+https://github.com/cachix/devenv" --eval
{ lastModified = 1743521698; lastModifiedDate = "20250401153458"; narHash = "sha256-bvl7YmNZjGEZhyH86mGiJdhZGMaRot1TMM6cjwrwH6k="; outPath = "/nix/store/b9319a8ffx2h1vd1bv7ip2ris0sf2sq2-source"; rev = "4917f16d045f12d938d0a29fb1cbfb9e24ae6d0d"; revCount = 3463; shortRev = "4917f16"; submodules = false; }

nixpkgs on  master [$] 
❯ /nix/store/pkw2fif96sahwnnkfxsmdzs0dl1nr5ln-nix-2.26.3/bin/nix-instantiate -E "builtins.fetchTree (builtins.parseFlakeRef git+https://github.com/cachix/devenv)" --eval
{ lastModified = 1743521698; lastModifiedDate = "20250401153458"; narHash = "sha256-bvl7YmNZjGEZhyH86mGiJdhZGMaRot1TMM6cjwrwH6k="; outPath = "/nix/store/b9319a8ffx2h1vd1bv7ip2ris0sf2sq2-source"; rev = "4917f16d045f12d938d0a29fb1cbfb9e24ae6d0d"; shortRev = "4917f16"; submodules = false; }

Property builtins.fetchTree input == builtins.fetchTree (builtins.parseFlakeRef input) should hold.

[flokli]

FWIW, revCount is present on Lix 2.91.1.

[fzakaria]

@domenkozar I did some spelunking.

I found

nix/src/libexpr/primops/fetchTree.cc

Line 150 in ab10fdd

if (type == "git" && !params.isFetchGit && !attrs.contains("shallow")) {

which is executed when builtins.parseFlakeRef is used which came from 358c26f

You can see that the commit explicitly tests fetchTree when an attribute set is passed in but not just a URL which has a different code path.

@tomberek or @roberth do you have some advice on how to handle this?
We could

• do nothing and note that they are two different ways with different defaults, albeit not great documented
• make the URL code path also do a shallow clone -- this might make sense since it's also seen to be optimal?
• revert the old PR

I'm not sure how much of this is a 'standard' that Lix vs. Snix should be identical too.

[GoldsteinE]

Would be really nice if fetchTree at least returned shallow = true; when it did a shallow fetch. This way fetchTree (fetchTree x)) uses cache on the second fetch.

I do think that fetchTree x and fetchTree (parseFlakeRef x) should definitely do the same thing though.