GitHub CLI extension for managing repositories that use a bare-git worktree layout.
gh wt keeps one bare repository in .bare/ and checks branches out as sibling
worktree directories:
repo/
├── .bare/
├── main/
├── feature-a/
└── feature-b/
Install from GitHub:
gh extension install williamw/gh-wtFor local development, clone and install from the working copy:
git clone https://github.com/williamw/gh-wt.git
cd gh-wt
gh extension install .Clone a repository into the .bare/ layout and create the default branch worktree:
gh wt clone owner/repoclone also writes an executable starter worktree-setup.sh and a
worktree-config.toml into the new repo root, and configures the remote
fetch refspec so git fetch keeps origin/* remote-tracking refs up to
date. The starter setup script validates the new worktree folder and leaves
a commented slot for project-specific setup (copying .env files, installing
dependencies, and so on). Edit it to taste, or delete it if you do not want
the hook to run.
Set up an existing repo — either a bare-git layout that predates the config file, or a plain clone you want converted to the worktree layout:
gh wt initIn a bare-layout repo, init fixes the fetch refspec if it is missing and
scaffolds worktree-config.toml, asking for an optional branch prefix
(Enter for none). If it finds the legacy setup-worktree.sh, it offers to
rename it to worktree-setup.sh; either way the config records the name in
use. In a normal clone, init offers to convert it in place:
.git/ becomes .bare/, the checked-out branch moves into its own worktree
folder, and local branches, stashes, and reflog all survive. Conversion
requires a fully clean tree (no modified, staged, or untracked files).
Add a worktree for a branch:
gh wt add feature-branchBranch names with slashes keep the branch name intact and use the final path segment as the folder name:
gh wt add user/feature-branch
# folder: feature-branch
# branch: user/feature-branchUse --branch-name when the folder name should differ from the branch name:
gh wt add feature-branch --branch-name user/feature-branch
gh wt add feature-branch -b user/feature-branch--base-branch controls the source branch for new worktrees. It does not rename
the new branch. Use -B as the short form:
gh wt add feature-branch --base-branch main
gh wt add feature-branch -B main--local creates a branch locally without pushing to origin. Use -L as the
short form (-l now belongs to --linear):
gh wt add feature-branch --local
gh wt add feature-branch -L--linear names the branch from a Linear issue URL. Use -l as the short
form. The issue ID and title slug become the branch name, and a positional
argument acts as a branch prefix (with or without a trailing slash):
gh wt add billw/ -l https://linear.app/modularml/issue/MKT-176/add-redirect-for-mojo-package-submission-page
# branch: billw/MKT-176-add-redirect-for-mojo-package-submission-page
# folder: MKT-176-add-redirect-for-mojo-package-submission-page
gh wt add -l https://linear.app/modularml/issue/MKT-176/add-redirect-for-mojo-package-submission-page
# branch: MKT-176-add-redirect-for-mojo-package-submission-pageA URL without a title slug yields just the issue ID; query strings, fragments,
and trailing slashes are ignored. --linear cannot be combined with
--branch-name, since both control the branch name.
Pasting a Linear issue URL as the positional argument works without the flag —
it is detected automatically and behaves exactly like -l:
gh wt add https://linear.app/modularml/issue/MKT-176/add-redirect-for-mojo-package-submission-pageEach repo can have a worktree-config.toml in the repo root (next to the
setup script). Both keys are optional:
setup-script = "worktree-setup.sh"
branch-prefix = "billw"branch-prefix prepends <prefix>/ to bare branch names. gh wt add foo
first checks out origin/billw/foo if it exists, then origin/foo, and
otherwise creates a new billw/foo branch. The prefix never applies when the
name already contains a / (full branch names, including other people's
branches), when --branch-name is given (exact names stay exact), or when a
positional prefix is used with --linear (it wins over the config). Linear-
derived branch names do get the prefix.
setup-script names the setup hook, resolved relative to the repo root. When
set, the script must exist and be executable — a missing configured script is
an error, not a silent skip.
Without the key, gh wt add looks in the current directory for an executable
worktree-setup.sh, then the legacy setup-worktree.sh. Finding both is an
error, since the choice is ambiguous — set setup-script to pick one. Repos
still using the legacy name without a config get a one-line hint pointing at
gh wt init.
The hook runs after the worktree is created, with the new folder name as its argument:
./worktree-setup.sh feature-branchIf the setup script exits non-zero, gh wt add exits non-zero too. The created
worktree is left in place.
Remove a worktree:
gh wt rm feature-branchrm resolves the actual checked-out branch from the worktree before deleting
the local branch, so the folder name does not need to match the branch name. Add
-d or --delete-remote to delete the resolved branch from origin too. If
the resolved branch is already missing from origin, rm warns and continues:
gh wt rm feature-branch -dRemove all worktrees whose associated pull requests are merged or closed:
gh wt rm --mergedUse --force to bypass the local branch safety check when you know you want to
delete a worktree with unpushed or untracked work.
If Git refuses to remove a worktree because it contains submodules, gh wt rm
deinitializes submodules in that worktree and retries with --force before
deleting the branch.
gh wt rm <folder> can also remove Git-registered worktrees outside the repo
root, such as temporary worktrees under /private/tmp, by matching the folder
basename shown by gh wt status. If that worktree is detached, gh wt rm
removes the worktree only and skips local or remote branch deletion.
If the worktree folder was already deleted but Git still has stale worktree
metadata for that folder, gh wt rm <folder> automatically runs
git worktree prune -v to clear the stale record. This cleanup does not delete
local or remote branches because the checked-out branch can no longer be safely
resolved from the missing worktree.
List worktrees:
gh wt listgh wt stack turns the worktree you are standing in into a stack of pull
requests, with one folder per layer so several people or agents can work
different layers at the same time. It is a converter, not a creator: start work
normally, and convert when the change outgrows one review.
gh wt add some-feature
cd some-feature
# ...work grows too large for one PR...
gh wt stack baseConversion renames the branch to billw/some-feature/01-base, moves the
worktree to some-feature/01-base/, and turns some-feature/ into a plain
container folder holding the layers. The name argument is optional: without it
layer 1 is just 01, and gh wt stack rename 01 01-base renames it later.
Conversion also enables git rerere in the shared .bare/config, which
gh stack init does not do despite its README.
Add layers, each in its own worktree:
cd 01-base
gh wt stack add api # -> some-feature/02-api, billw/some-feature/02-api
gh wt stack add ui # -> some-feature/03-ui, billw/some-feature/03-uiEach new layer is pushed to origin with an upstream, like gh wt add; pass
-L/--local to keep it local until gh stack submit. The number is added
for you; gh wt stack add 02-api also works, but a number that is not the next
layer's is refused. add, rename, rebase, and agent are reserved words,
so they cannot name layer 1. The older flag spellings
(--add, --rename, --rebase, --agent) still work.
some-feature/
├── 01-base/ billw/some-feature/01-base
├── 02-api/ billw/some-feature/02-api
└── 03-ui/ billw/some-feature/03-ui
Folder and branch always spell the same thing. The branch path under the
prefix is the folder path, which is why conversion has to rename layer 1: a ref
cannot be both a leaf and a directory, so billw/some-feature existing would
block billw/some-feature/02-api from ever being created.
Renaming layer 1 closes an open PR on it. GitHub closes any pull request
whose head branch is renamed. gh wt stack prompts before doing it, and a run
whose output is captured — an agent, a script — fails instead, so --force is
required to go ahead. gh wt stack rename is gated the same way.
The bottom layer is the host. Its worktree holds gh stack's metadata,
so it has to stay on layer 1's branch: gh stack view fails on a detached
HEAD. It is also the one layer gh-wt never detaches. gh stack add only works
at the top of a stack, so gh wt stack add briefly detaches the layer
worktrees above, climbs, adds, and puts them back. Uncommitted work in those
folders is preserved.
Run gh stack commands from the bottom layer's folder. It is the only
worktree carrying the metadata, so gh stack view, push, submit, merge,
and sync report "not part of a stack" anywhere else. The gh wt stack
subcommands find the host themselves and work from any layer folder or from
the container.
Move between layers with cd, e.g. cd ../02-api. Never use
gh stack up, down, switch, top, or bottom: they check a layer out in
the current worktree, and every other layer is already checked out in its own.
Cascade with gh wt stack rebase, never gh stack rebase. The latter
checks each layer out in turn and dies on any branch another worktree holds,
then rolls the whole cascade back. gh wt stack rebase detaches every layer
worktree first and restores it afterwards:
gh wt stack rebase # pull trunk and rebase every layer
gh wt stack rebase --no-trunk # cascade a lower-layer fix without pulling trunkEvery agent must be idle and committed when it runs. On a conflict the layers
are left detached and the message points you at the host folder; resolve there
and run gh wt stack rebase --continue.
gh wt stack refuses to run on the trunk worktree, on a detached HEAD, on a
worktree that already hosts a stack, and when the gh stack extension is
missing. Install it with gh extension install github/gh-stack.
If the adopted branch is still too large for one review (more than 3 commits
or 500 changed lines), gh wt stack says so. When its output is being captured
rather than shown in a terminal — as when an AI agent runs it — it also prints
a short block pointing at gh wt stack agent, which prints the full
layer-splitting procedure. Run gh wt stack agent yourself to see exactly
what agents are told.
gh wt status renders a stack once, under its container, rather than once per
layer worktree:
some-feature
Stack: #2331 - 3 layers
Status: Clean
Layers:
1 billw/some-feature/01-base #2326 (OPEN)
2 billw/some-feature/02-api #2327 (OPEN)
3 billw/some-feature/03-ui #2328 (OPEN)
A layer needing a rebase is flagged and makes status exit non-zero. If
gh stack view is unavailable, status falls back to the ordinary
single-branch display.
gh wt rm treats a stack as all-or-nothing. Remove it by its container name:
gh wt rm some-featureRemoving a single layer is refused, because deleting one layer's branch out of
a live stack strands the rest. gh wt rm --merged removes a stack only when
every layer's PR is merged or closed, and never sweeps layers one at a time as
their individual PRs land.
Known limitation: gh stack trunk and gh stack checkout <trunk> cannot work
in this layout, because trunk is already checked out in its own worktree.
A stack from start to finish:
gh wt add MCL-123-feature && cd MCL-123-feature
# ...work grows too large for one PR...
gh wt stack docs # -> MCL-123-feature/01-docs/
cd 01-docs && gh wt stack add api # -> ../02-api/, pushed
# fix something in 01-docs and commit it, then cascade it upward:
gh wt stack rebase --no-trunk && gh stack push
gh stack submit --auto --open # open every layer's PR, linked as a stack
# later, trunk moved:
gh wt stack rebase && gh stack push
# after review:
gh stack merge
gh wt rm MCL-123-featureShow status for all worktrees:
gh wt status- GitHub CLI (
gh) - Git
- Python 3.10+ (3.11+ when a
worktree-config.tomlis present, fortomllib) - Bare-git layout created by
gh wt cloneorgh wt init, or an existing repo with.bare/
Install dependencies:
pixi installRun tests:
pixi run testInstall the local checkout as the active gh wt extension:
gh extension remove gh-wt 2>/dev/null || true
gh extension install .
gh wt --help