Commits from a sibling checkout

Use this skill when another checkout of the monorepo on this machine holds commits that this checkout needs, for example a reorganization that a new change builds on. The user keeps the checkouts separate, so a clone, a worktree or a shared directory is not an option.

Why not fetch

A git fetch /path/to/sibling walks and packs objects over the whole history. With a hundred upstream commits of difference it took more than 15 minutes here, and a second fetch of 12 new commits did not finish in 10 minutes. It also leaves a remote-tracking ref and a copy of every object behind. A patch carries only the change.

Procedure

1. Find the range in the sibling. The base is the last sibling commit this checkout already has; a commit taken before has a new hash here, so compare by subject or by patch id (git patch-id --stable), not by hash.

2. From the root of the receiving checkout run the script. It writes the patches under out/sibling-patches/, refuses when a patch touches a path that is dirty here, then applies and commits each patch in order:

   ./community/.agents/skills/sibling-checkout-commits/scripts/take-patches.sh ../idea-2 <base>..master
   

3. Read the ok lines. Each one names the new commit and confirms that its patch id equals the patch. A DIFF line stops the script; the commit on HEAD then holds something else, and the next patch is not applied.

4. On a weekend, the committer date is today: apply the weekend rule of the personal CLAUDE.md to the new commits.

What the script does, and does not do

• git apply --binary changes the working tree only. git commit --only -- <paths> builds the commit from HEAD plus those paths, so nothing that a concurrent session staged in the shared index lands in the commit.
• A path that a patch creates or renames is untracked after the apply, and --only skips an untracked path. The script marks every such path with git add -N first; an intent-to-add entry carries no content.
• A path that a patch deletes stays in the pathspec, so the commit records the deletion.
• The author, the date and the message come from the patch through git mailinfo.
• It does not use git am: git am commits the shared index, and a concurrent session's staged files would land in the commit.
• It does not fetch, and it creates no ref. Delete a ref of an earlier fetch with git update-ref -d refs/remotes/<name>/master.

When a patch does not apply

A patch of the range depends on the patches before it, so a --check of a later patch alone fails. The script applies in order. When a patch fails against this tree, the sibling and this checkout diverged on that path: stop, name the path and the commit in the report, and let the user decide. Do not resolve the conflict by hand inside a commit that carries the sibling author's name.