A parent branch can match its remote exactly while its child still contains an earlier version of that parent. In a stack of GitLab merge requests, that can make the child diff show changes that belong to the parent. This showed up in a local test for GitLab CLI issue #8283. The useful clue was the commit graph hiding behind a perfectly reasonable git status.

Reproduce it without a GitLab account

This creates a disposable repository and a local bare remote. The lease-protected push below goes to that local directory.

demo=$(mktemp -d) git init -q --bare "$demo/origin.git" git init -q -b main "$demo/work" cd "$demo/work" git remote add origin ../origin.git git config user.name "Stack Demo" git config user.email stack-demo@example.invalid printf "mode=base\n" > settings.txt git add settings.txt git commit -qm "Initial setting" git switch -qc parent printf "mode=original\n" > settings.txt git commit -qam "Set the default mode" old_parent=$(git rev-parse HEAD) git switch -qc child printf "consume mode\n" > app.txt git add app.txt git commit -qm "Read the default mode" git push -qu origin main parent child git switch -q parent printf "mode=reviewed\n" > settings.txt git commit -qa --amend --no-edit git push -q --force-with-lease origin parent 

The amend changes the parent commit ID while preserving its subject. The child still points to the old commit. Moving origin/parent along with parent makes Git report that the parent is up to date without changing the child at all.

Ask Git about the parent and child

git status --short --branch git merge-base --is-ancestor parent child git log --format=%s parent..child 

The status line is ## parent...origin/parent, with no ahead or behind count. The ancestry check exits with status 1. The log lists Read the default mode and Set the default mode. That second entry is the old parent commit, even though the new parent has the same subject. This is the difference between a branch agreeing with its upstream and two branches agreeing on their history.

Replay the child from the boundary you know

In this disposable repo, old_parent is the exact boundary captured when the child was created. First check that it is an ancestor of the child. Then keep a reference to the pre-rebase child and replay only the commits after that boundary.

git merge-base --is-ancestor "$old_parent" child && git branch backup/child-before-restack child && git rebase --onto parent "$old_parent" child git merge-base --is-ancestor parent child git log --format=%s parent..child git show child:settings.txt git diff --stat parent..child 

Now the ancestry check succeeds. The log lists only Read the default mode, settings.txt says mode=reviewed, and the diff contains only app.txt. There is no publication step in this second block. On a shared branch, fetch and inspect the child’s remote tip before rewriting history, especially if a teammate has pushed since you last looked.

The CLI regression behind this example checks three outcomes with a local bare remote: a local-only sync keeps the remote child unchanged, a later regular sync publishes the rebased child, and another sync keeps the same commit IDs. That tests the result users care about: each merge request should still show its own layer after the parent changes.

Reference: Git’s merge-base documentation.

AI disclosure: This article was drafted by an AI agent. The Git example and CLI regression were run in disposable local repositories.