Hi,

The report below was investigated and drafted by Claude in the process of building a branch-protection tool; I've reviewed it and verified the reproduction before sending. Happy to answer questions or test patches.

---

githooks(5) says the reference-transaction hook "is invoked by any Git command that performs reference updates". Some branch copy and rename operations update a ref without invoking it.

Observed on git version 2.55.0.windows.5, with the hook registered both in .git/hooks and as a config-defined hook (hook.<name>.event).

1. `git branch -C <src> <dst>` where <dst> exists and is not checked out
   anywhere: <dst> is overwritten with <src>'s value and the hook is not
   invoked in any phase. This happens on both the files and the reftable
   backend. With every documented hook event registered to a logging
   script, none of them fired, and a GIT_TRACE2_EVENT log showed no child
   process.

2. `git branch -c <src> <dst>` where <dst> does not exist: <dst> is
   created and the hook sees no line for it.

3. `git branch -m/-M <src> <dst>`:
   - files backend: the hook sees the deletion of <src> and a
     "0000... 0000... refs/heads/<dst>" line, but no line carrying
     <dst>'s new value;
   - reftable backend: no line names <dst> at all, and renaming a branch
     away deletes it without a line for it.

4. `git reflog delete --updateref --rewrite <branch>@{0}` rewinds the
   branch without invoking the hook, on both backends.

Reproduction (any POSIX shell):

  git init -q -b main t && cd t
  git commit -q --allow-empty -m one
  git branch other
  git commit -q --allow-empty -m two
  git switch -q other
  printf '#!/bin/sh\necho "hook $1" >>"$PWD/hook.log"\ncat

  git config hook.log.event reference-transaction
  git config hook.log.command "sh $(cd ..; pwd)/h.sh"
  : >hook.log
  git branch -C other main     # main now points at "one"
  cat hook.log                 # empty

Expected: the hook runs with a line updating refs/heads/main, as it does
for `git branch -f main other`. Tools that enforce policy through this
hook cannot otherwise see these updates.

Note that git already refuses -C/-M onto a branch checked out in any worktree, so the gap is limited to branches that are not checked out.

On 2026-10-02 at 04:12:55, Никита Поникаров wrote:

Hi,

I think there was a recent thread about this at https://lore.kernel.org/git/CACQ=SRHCOCcmVCgHqd+sjMsZ9LCdSHuXdCo0gkwxXwYgF7iwig@mail.gmail.com/

There were some patches, but they appeared AI generated and were not picked up. Perhaps you or someone else could send in some higher-quality patches not using AI (see Documentation/SubmittingPatches) that could fix the issue.

I will say that using reference transactions to solve this problem would be very desirable from a variety of perspectives, especially since it would probably go a long way to unlocking way better performance for `git remote rename` with many remote-tracking branches as well.