Git Worktree

Git worktree manager

GATE: If the task runs inside an existing worktree (a worktree path is given and no create/remove/switch is requested), none of the creation flow applies — work in place and skip this skill. To check: git rev-parse --show-toplevel appears as a linked entry in git worktree list.

Working rules

• Work in place when given an existing worktree without a creation/removal request.
• Keep one writer per worktree and preserve other sessions' branches, files, and staged work.
• Verify dependencies resolve first-party code from the intended worktree.
• Keep merge, push, and cleanup within user authorization; report the exact exercised checkout and resulting commit.

Always use the manager script

Never call git worktree add directly -- always use the worktree-manager.sh script.

The script handles critical setup that raw git commands don't:

1. Copies .env, .env.local, .env.test, etc. from main repo
2. Ensures .worktrees is in .gitignore
3. Creates consistent directory structure
4. After creation, install dependencies if detected: package.json → npm install, composer.json → composer install, pyproject.toml → pip install -e ., go.mod → go mod download

All commands use: bash ${CLAUDE_PLUGIN_ROOT}/skills/ia-git-worktree/scripts/worktree-manager.sh <command>. If CLAUDE_PLUGIN_ROOT is unset (non-Claude-Code harness), resolve the script relative to this skill's own directory.

Before creating worktrees, export a unique WORKTREE_SESSION_ID and retain that same value for this session's later manager calls. For example, export WORKTREE_SESSION_ID="$(python3 -c 'import uuid; print(uuid.uuid4())')". The manager records ownership in each new worktree's Git metadata. Creation without a session ID remains available, but manager cleanup then refuses that tree; never adopt a previous session's ID to bypass ownership.

The manager script fetches origin/<base> fresh and branches from it -- it never checks out <base> in the caller's working tree. If the fetch fails (offline, no remote), it falls back to the local <base> ref. Details: troubleshooting.md.

Commands

Run commands with env -C "$target" <command> or the harness workdir argument. A child script cannot change the caller's working directory. Listing and resolving names work from the main checkout, linked checkouts, and their subdirectories.

Cleanup refuses the current checkout, another session's tree, and tracked, untracked, or ignored files. Auto-copied .env files and installed dependencies therefore require explicit user-managed disposition before cleanup. Confirm no process uses the named trees; the manager cannot detect every external reader. Git's normal removal safeguards remain enabled, including locked-tree refusal. Do not force deletion or suppress failures to finish cleanup.

Safety Verification

Before creating a worktree, verify the worktree directory is gitignored:

# Verify .worktrees is ignored (should output ".worktrees")
git check-ignore .worktrees || echo "WARNING: .worktrees not in .gitignore"

If not ignored, add it to .gitignore before proceeding.

After creating a worktree, run the project's test suite (or the fastest relevant subset if the full suite exceeds a few minutes) to establish a clean baseline. Pre-existing failures in the worktree should be caught before starting new work -- not discovered mid-implementation.

Ownership

• One writer per worktree. Treat every git worktree list entry you did not create in this session as read-only -- a tree left from a previous round is not yours either. Reuse is most tempting exactly where it is most dangerous: an existing tree already has dependencies and env wired up, and another session may be running a suite in it.
• Do not mutate a tree while your own suite runs there. Test runners load source files as they reach them, so a mid-run edit produces a mass-failure result that looks exactly like a real regression.
• A failure burst that contradicts a claim is a harness hypothesis, not a conclusion. Do not record or report the self-inflicted attribution until a re-run on a tree you have just asserted clean (git status --short empty) has returned.
• When a mutation is unavoidable, assert the restore (grep the token back to its original count, plus git status --short) rather than trusting git checkout --.
• One checkout has one index, so staging explicit paths does not scope a commit: git add <mine> && git commit also commits whatever a peer staged, under your message. The protection is a pathspec on the commit itself (git commit -- <paths>), which takes those paths from the working tree and ignores the index; new files still need git add. It constrains your commit, not a peer's, so the residual control is latency between writing and committing. Read git show --stat HEAD afterwards and confirm only your files are there.
• git -C <repo> push <remote> HEAD:<branch> resolves HEAD in that repo, not in the worktree you edited. Edits made in a linked worktree and pushed with -C at the main checkout publish the main checkout's commit onto your branch, and --force-with-lease does not catch it because the lease checks the branch's old value, not what HEAD names. Never spell HEAD: in a -C push -- resolve the SHA in the worktree and push it explicitly, then confirm with git ls-remote. Two branches "updated" to one SHA, or a pushed subject unrelated to the work, is the tell.
• Linked worktrees share one stash stack. git stash writes to the common git directory, so a red/green cycle in one worktree can pop and drop a stash another worktree pushed in between. Never stash for red/green here: git diff > /tmp/red.patch, git checkout -- <files> (worktree-local) for the red run, git apply /tmp/red.patch for green. A dropped stash is still recoverable while its commit survives -- git stash store -m <message> <sha> re-registers the SHA that Dropped refs/stash@{0} (<sha>) printed.
• A worktree's HEAD is shared mutable state, so answer branch questions from refs. Any other session can check something else out there, which makes git -C <worktree> rev-parse HEAD describe a different branch and report a correct push as a mismatch. Refs are shared across every worktree: ask any one of them about the branch by name (rev-parse <branch>, rev-list --count origin/<branch>..<branch>, reflog <branch>).

Use env -C <worktree> <cmd> for every command, never cd. A shell's cwd persists across calls, so one cd <repo-root> for an unrelated reason silently relocates every later command: probe files get written into the shared main tree and run against its bytes, and the tidy-up reflex git checkout -- <path> becomes a write aimed at the wrong tree. The git -C habit does not generalize -- interpreters, test runners, linters, and a heredoc cat > all take the cwd. Have any probe print the tree it ran in.

Verify

• git worktree list shows the new entry
• .worktrees directory confirmed in .gitignore
• Dependencies installed in the worktree
• Baseline test suite passes in the worktree

References

• workflow-examples.md - Code review and parallel development workflows
• troubleshooting.md - Common issues, fresh-remote-base behavior, directory structure, how it works
• hooks-and-excludes.md - Hook safety under Husky, .git/info/exclude vs .gitignore
• worktree-manager.sh - The manager script

Task-specific references

Read the relevant reference before implementing or reviewing the matching behavior:

• For dependency provenance, environment adaptation, review/work workflows, branch completion, or hooks: worktree-integration.md.

Existing specialized references, when the corresponding topic applies: