git-cow-worktree is a drop-in replacement for git worktree add.

It uses copy-on-write for the worktree, reducing disk usage.

It is entirely vibe-coded, but there's also not that much to it. Use at your own risk.

Everything below is LLM-written.

---

Install

go install github.com/josharian/git-cow-worktree@latest

Git can dispatch git cow-worktree ... when the installed git-cow-worktree binary is on your PATH.

Usage

git cow-worktree add [git-worktree-add flags] [--from <path>] [-v] <path> [<commit-ish>]

Examples:

git cow-worktree add ../repo-feature feature
git cow-worktree add -b topic ../repo-topic main
git cow-worktree add -v --from ../repo-main ../repo-topic topic

Most flags are inherited from git worktree add.

Added by git-cow-worktree:

• --from <path>: use a specific source worktree instead of auto-selecting one.
• -v, --verbose: print the chosen source, clone counts, and timings.

Special case:

• --no-checkout: passed through to git worktree add; no reflinking is attempted.

Filesystem Support

Reflinking is supported on APFS on macOS and on Linux filesystems with FICLONE support, such as btrfs, XFS, and bcachefs. On unsupported filesystems or across devices, git-cow-worktree falls back to Git's normal checkout behavior.

Scope

git-cow-worktree only replaces git worktree add. Other git worktree subcommands should be run with Git directly.

How it works

It first runs git worktree add --no-checkout, so Git creates the new worktree metadata but leaves the working tree empty.

Then it picks a source worktree. Candidates are the current worktree, the main worktree, and a few recently modified other worktrees. It ignores the target worktree and unmaterialized worktrees. Candidates are scored by git rev-list --left-right --count <source>...<target>; fewer commits ahead/behind is better.

It runs git ls-tree -r -t on the source and target commits and plans the cheapest way to materialize each path. A directory whose tree SHA is the same on both sides has identical tracked contents, so where the filesystem supports it (APFS does; Linux reflinks are file-only) the whole directory is cloned in a single syscall. Otherwise it descends and reflinks matching files one by one. Paths that are dirty, untracked, ignored, or submodules in the source are skipped, along with symlinks and mode mismatches.

Every cloned file is then hashed and compared against the blob it is supposed to be, and the ones that check out are written to the new worktree's index with their stat data. This is the step that preserves the sharing: Git decides whether to rewrite a working-tree file by comparing it against its index entry, and a worktree with no index — which is what --no-checkout leaves behind — makes checkout rewrite every file, undoing every clone. Hashing in parallel is also substantially cheaper than letting Git refresh the index itself, which is single-threaded and uses a collision-detecting SHA-1.

Finally it runs Git checkout in the new worktree. That fills in files that were not cloned, overwrites anything stale, and produces the same index a plain checkout would. Every step before it is best-effort: if reflinks are unsupported, if the source can't be inspected, or if the index can't be written, the checkout still runs and just becomes a normal one.