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#

• Preserve other sessions' branches, files, and staged work (see Ownership).
• 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 the main checkout, only when the base is the default branch or a local branch that already contains the fetched commit (a branch that exists only on origin gets none); it never overwrites or follows an existing file or symlink in the new tree
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

Untrusted bases (PR heads, fork or origin-only branches, third-party code): create with --no-env, never copy-env into the tree, and keep credentials out of its commands. Skip installs; if needed, disable lifecycle scripts and plugins ( npm ci --ignore-scripts, composer install --no-scripts --no-plugins) and never pip install -e ., which runs the build backend.

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. Example: export WORKTREE_SESSION_ID="$(python3 -I -c 'import uuid; print(uuid.uuid4())')". The manager records ownership in each new worktree's Git metadata. Creation without a session ID works, but manager cleanup then refuses that tree; never adopt a previous session's ID to bypass ownership.

The manager script fetches <base> from origin into an invocation-private ref, resolves that commit, and creates the branch from its SHA, even when origin/<base> is absent or stale or another checkout fetches concurrently. It removes only its private ref and never checks out <base> in the caller's working tree. If the fetch fails (offline, no remote), it resolves the local <base> instead. PR review: pass the fetched PR head as <base>, then verify its SHA ( workflow-examples.md). Details: troubleshooting.md.

Commands#

CommandDescriptionExamplecreate [--no-env] <branch> [from]Create worktree + branch (default: from main)...worktree-manager.sh create feature-login``list / lsList all worktrees with status...worktree-manager.sh list``switch <name> / goPrint the registered worktree's absolute path; the caller applies it as workdirtarget=$(...worktree-manager.sh switch feature-login)``copy-env <name>Copy .env files to existing worktree...worktree-manager.sh copy-env feature-login``cleanup <name> [...] / cleanConfirm removal of named, session-owned, clean worktrees...worktree-manager.sh cleanup feature-login

Run commands with env -C "$target" <command> or the harness workdir argument. A child script cannot change the caller's working directory. Listing and name resolution work from any main or linked checkout, including 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:

bash

# 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.

Before any baseline suite or migration, apply ia-verification-before-completion's effective-target preflight. Resolve worktree environment/configuration overlays; verify disposable database and service targets, never the dev database. Copied .env files do not prove isolation. Without verified disposable targets, stop before tests. Then run the suite or its fastest relevant subset to establish a clean baseline before implementation.

Ownership#

• One writer per worktree. Treat every git worktree list entry not created in the current session as read-only; a tree left from a previous round is not the current session's 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 the current session's 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 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: a bare commit includes unrelated staged entries, while git commit -- <paths> records complete working-tree files, including caller or peer hunks. Use a pathspec commit only after establishing whole-file ownership of every selected change. For mixed hunks, commit the attributable patch in a clean session-owned checkout while preserving the caller's HEAD, index, and files. Inspect the exact proposed and committed patches, not only filenames or --stat; follow commit-ownership.md.
• Never use HEAD: in a git -C push. Resolve the SHA in the edited worktree, push it explicitly, and confirm with git ls-remote; a lease cannot catch selection of another checkout's HEAD. Details: commit-ownership.md.
• Never git stash for red/green: linked worktrees share one stash stack. Use the patch-file cycle and stash recovery in commit-ownership.md.
• Answer branch questions from refs ( rev-parse <branch>), not a worktree's HEAD, which other sessions can move; see commit-ownership.md.

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 (trusted bases; untrusted bases follow the rules above)
• 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
• commit-ownership.md - Whole-file commits and isolated staging for mixed hunks
• 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: