DEV Community

Recca Tsai
Recca Tsai

Posted on • Originally published at recca0120.github.io

git worktree: Multiple Working Directories Per Repo, and the Key to Parallel AI Agents

#git #worktree #productivity #agents

Originally published at recca0120.github.io

You're mid-feature when PM says prod is on fire. The reflex move: git stash, check out master, fix, come back, stash pop. But then the dev server restarts, the IDE re-indexes, and stash silently ate your untracked build artifacts.

git worktree fixes this cleanly: one repo, many checked-out branches in separate directories, each with its own HEAD, index, and untracked files — all sharing the same object database. The feature stays exactly where you left it; the hotfix happens next door.

vs stash, vs multiple clones

Approach Switch cost Disk Object sync
git stash + checkout High (dev server restarts, IDE reindex) Cheap Single repo
Multiple clones Low Wasteful (GBs on large repos) Each fetches separately
git worktree Low Cheap (shared objects) One fetch updates all

Worktree is the synthesis — filesystem isolation with a unified object store.

Core Commands 

# Add — last path segment becomes the new branch name
git worktree add ../hotfix                   # branch "hotfix" from HEAD
git worktree add ../review pr-123            # checkout existing branch
git worktree add -b feat-x ../feat-x main    # new branch from main
git worktree add -d ../throwaway             # detached HEAD, no branch

git worktree list                            # --porcelain for scripts
git worktree remove ../hotfix                # clean only; -f for dirty
git worktree prune                           # clean metadata for manually deleted dirs
git worktree lock ../usb-drive --reason "removable drive"
git worktree unlock ../usb-drive
git worktree move ../old ../new              # won't move worktrees w/ submodules
git worktree repair                          # fix links after moving the main repo
Enter fullscreen mode Exit fullscreen mode

Short aliases make it muscle memory:

git config --global alias.wta 'worktree add'
git config --global alias.wtl 'worktree list'
git config --global alias.wtr 'worktree remove'
Enter fullscreen mode Exit fullscreen mode

Real Workflows

1. Hotfix without disturbing feature work

git worktree add ../hotfix-prod origin/main
cd ../hotfix-prod
# fix, commit, push, open PR
cd ../myrepo
git worktree remove ../hotfix-prod
Enter fullscreen mode Exit fullscreen mode

The feature's dev server never stopped. node_modules untouched. IDE didn't re-index.

2. Long test run + keep coding

pytest / cargo test takes 10 minutes. Open a worktree to run it there while you keep editing the next commit. Zero interference.

git worktree add ../ci-run branch-a
cd ../ci-run && pytest --slow &
cd -   # back to main worktree, keep coding
Enter fullscreen mode Exit fullscreen mode

3. Code review without polluting your state

git worktree add ../review-456 pr-456-branch
cd ../review-456
# run it, read code, leave comments
cd - && git worktree remove ../review-456
Enter fullscreen mode Exit fullscreen mode

4. Multiple AI agents in parallel (the most underrated 2025 use case)

Claude Code / Cursor / Aider only work in one directory at a time — running two in the same folder means they overwrite each other's files. One branch per worktree and you can run three agents on three features simultaneously, each with its own dev server port and node_modules, no collisions:

git worktree add ../agent-a -b feat-a
git worktree add ../agent-b -b feat-b
git worktree add ../agent-c -b feat-c

# three tmux panes / three terminal windows, one claude each
Enter fullscreen mode Exit fullscreen mode

Claude Code's Agent tool even has a built-in isolation: "worktree" option — sub-agents automatically run in a worktree and merge back when done.

Three Directory Layouts

Sibling dirs (simplest)

~/code/myrepo
~/code/myrepo-hotfix
~/code/myrepo-review-123
Enter fullscreen mode Exit fullscreen mode

Plays well with editors' "one folder, one project" mental model.

.worktrees/ subdir

myrepo/
├── src/
└── .worktrees/
    ├── feat-x/
    └── hotfix/
Enter fullscreen mode Exit fullscreen mode

Everything co-located. Add .worktrees/ to global gitignore. Downside: some tools (ESLint, tsc) recurse into it.

Bare repo pattern (best for heavy multi-branch work)

mkdir myproj && cd myproj
git clone --bare git@github.com:org/repo.git .bare
echo "gitdir: ./.bare" > .git
git worktree add main
git worktree add feat-x
Enter fullscreen mode Exit fullscreen mode

Result:

myproj/
├── .bare/           # actual object store
├── .git             # file, pointing to .bare
├── main/            # main branch checkout
└── feat-x/          # feat-x branch checkout
Enter fullscreen mode Exit fullscreen mode

No "primary working copy" — every branch is a worktree, cd is the branch switch. Pairs beautifully with tmux and AI agents.

Gotchas

.git in a linked worktree is a file, not a directory. Contents are gitdir: /path/to/main/.git/worktrees/<name>. Tools that read .git/ as a directory will break.

You can't check out the same branch in two worktrees. By design — prevents index divergence. Override with --force or use detached HEAD.

Submodules are painful. worktree move refuses; remove needs --force. .git/modules/ is shared, so switching submodule commits in one worktree affects the others. Heavy-submodule repos need care.

node_modules, venv, target/ are per-worktree. Disk gets hungry. Mitigations:

  • pnpm's content-addressable store — reinstalling across worktrees uses almost no extra space
  • Rust: CARGO_TARGET_DIR=~/.cache/cargo-target shares target dirs
  • uv, poetry caches are shareable too

.env doesn't get copied. Use direnv — drop a .envrc in each worktree and it auto-loads on cd.

Hooks are shared (.git/hooks/ is in the common dir). If a hook assumes repo root via a hardcoded path, it breaks in linked worktrees. Always use git rev-parse --show-toplevel for the current worktree's root.

IDE indexes run per worktree. VS Code / JetBrains index each one independently — duplicate CPU and disk. Structural limit, no way around it.

Editor Integration

VS Code: use multi-root workspace (File → Add Folder to Workspace) to open multiple worktrees at once, or just open each in its own window. Since 2024, GitHub.vscode-pull-request-github checks out PRs as worktrees.

JetBrains: native worktree UI in the Git tool window since 2023.2 — create, switch, remove from the GUI.

fzf quick-switch:

wtcd() {
  cd "$(git worktree list --porcelain | awk '/^worktree /{print $2}' | fzf)"
}
Enter fullscreen mode Exit fullscreen mode

Type wtcd, fuzzy-search worktree paths, enter to cd.

What's New in 2024–2026

Git 2.44 (2024/2): git worktree add --orphan — creates a worktree with an unborn branch. Handy for gh-pages-style split deploy branches.

Git 2.46 (2024/7): worktree.useRelativePaths config + --relative-paths flag — internal links use relative paths. The main repo (or the whole dir tree) can move without breaking worktrees. Huge for Dropbox/iCloud sync and containerized dev.

Git 2.48 (2025/1): git worktree repair auto-fixes absolute/relative path mismatches.

Git 2.50 (mid-2025): relative paths and porcelain output have stabilized.

Ecosystem: "one worktree per AI agent branch" went mainstream in 2025. git-town, ghq, and the gh CLI all picked up first-class worktree support.

Getting Started

  1. In any repo, try git worktree add ../test-wt -b test-branch and poke around
  2. Back in the main worktree, notice git branch sees it but git status is untouched
  3. git worktree remove ../test-wt — main worktree is completely unchanged
  4. Add wta / wtl / wtr aliases, build muscle memory
  5. Next hotfix, use worktree instead of stash — feel the difference

For anyone using AI coding agents, worktree isn't a bonus — it's required. The "only one agent per repo" limitation is the exact thing worktree removes.

References

Top comments (0)