[PATCH] remote: detect collisions in remote names

On Sat, Jul 05, 2025 at 12:57:50PM -0400, Jeff King wrote:

So I'm not entirely convinced we should do anything here. The answer might just be "if it hurts, don't do it". But if we wanted any protections in the "git remote" porcelain, they might look like this:

-- >8 --
Subject: [PATCH] remote: detect collisions in remote names

When two remotes collide in the destinations of their fetch refspecs, the results can be confusing. For example, in this silly example:

  git config remote.one.url [...]
  git config remote.one.fetch +refs/heads/*:refs/remotes/collide/*
  git config remote.two.url [...]
  git config remote.two.fetch +refs/heads/*:refs/remotes/collide/*
  git fetch --all

we may try to write to the same ref twice (once for each remote we're fetching). There's also a more subtle version of this. If you have remotes "outer/inner" and "outer", then the ref "inner/branch" on the second remote will conflict with just "branch" on the former (they both want to write to "refs/remotes/outer/inner/branch").

We probably don't want to forbid this kind of overlap completely. While the results can be confusing, there are legitimate reasons to have multiple refs write into the same namespace (e.g., if one is a "backup" of the other that is rarely fetched from).

But it may be worth limiting the porcelain "git remote" command to avoid this confusion. The example above cannot be done with "git remote", because it always[1] matches the refspecs to the remote name, and you can only have one instance of each remote name. But you can still trigger the more subtle variant like this:

  git remote add outer [...]
  git remote add outer/inner [...]

So let's detect that kind of name collision (in both directions) and forbid it. You can still do whatever you like by manipulating the config directly, but this should prevent the most obvious foot-gun.

[1] Almost always. With the --mirror option, the resulting refspec will
    just write into "refs/*"; the remote name does not appear in the ref
    namespace at all.

    Our new "names must not overlap" rule is not necessary for that
    case, but it seems reasonable to enforce it consistently. We already
    require all remote names to be valid in the ref namespace, even
    though we won't ever use them in that context for --mirror remotes.

    Likewise, our new rule doesn't help with overlap here. Any two
    mirror remotes will always overlap (in fact, any mirror remote along
    with any other single one, since refs/remotes/ is a subset of the
    mirrored refs). I'm not sure this is worth worrying about, but if it
    is, we'd want an additional rule like "mirror remotes must be the
    only remote".

Signed-off-by: Jeff King <peff@peff.net>
---
 builtin/remote.c  | 17 +++++++++++++++++
 t/t5505-remote.sh | 14 ++++++++++++++
 2 files changed, 31 insertions(+)