using-git-worktrees · ganyuanran — скилл для Claude Code: рабочие деревья Git

---
name: using-git-worktrees
description: "Use when a coding task needs a concurrent checkout, unrelated dirty state blocks safe branch switching, or the user or repository explicitly requires a worktree."
---

# Using Git Worktrees

## Overview

A worktree is an exception for simultaneously checking out another Git history,
not a general safety badge. Prefer the current workspace and reuse an existing
suitable worktree. A branch decision and a worktree decision are separate.

**Core principle:** create only with necessity and ownership evidence; pair every
created worktree with a cleanup path.

**Announce at start:** explain which concrete checkout conflict requires a
worktree.

## Step 0: Environment Detection

Before mutation, record the repository root, `HEAD`, branch/detached state,
staged/unstaged/untracked paths, active Git operations, and:

```bash
git worktree list --porcelain
```

Classify the current host surface's task/chat-bound managed workspace semantics
as `managed`, `non-managed`, or `unknown`. Use trusted host/session context, an
explicit host contract, or a host-native lifecycle result; do not infer the
classification from a product name, directory prefix, command-level `workdir`,
missing tool, or absent metadata. `non-managed` requires positive evidence that
the current surface lacks task-bound managed semantics. When the host exposes
workspace fields, record the task workspace and the command's default `cwd`
without a per-command directory override.

Stop on unresolved conflicts, an active merge/rebase/cherry-pick/revert/bisect,
or an unexplained detached HEAD. A detached HEAD is acceptable only when
trusted host binding evidence identifies it as the intended state of the
current managed worktree; Codex-managed worktrees, for example, start detached
by default. Preserve user state: no automatic stash, reset, clean, broad
staging, or commit.

Before deciding necessity or placement, read the smallest relevant project
`AGENTS.md`, `CLAUDE.md`, current authority, and existing worktree convention.
Repository policy may require or forbid isolation and owns any approved
in-repo location; do not discover that policy after creating the checkout.

## Step 1: Necessity Decision

Create at most one task-owned worktree only when one of these is true:

- another branch must remain checked out while this task proceeds;
- unrelated dirty state prevents a safe checkout in the current workspace;
- explicit user or current repository authority requires it.

Task complexity, planning, TDD, use of subagents, or a `main`/`master` branch
name is not sufficient. Same-task subagents share one workspace; do not create
per-subagent or nested worktrees. If independent history is unnecessary, do not
create a branch either.

Record: reason, branch, exact path, creating coordinator, and expected cleanup
event. If the owner or cleanup condition is unknown, stop before creation.

## Step 2: Safe Placement

Use an existing repository convention only when the target directory is already
ignored (`git check-ignore`). Never edit or commit `.gitignore` solely to host a
worktree. If no safe in-repo convention exists, use a task-specific external
user-level temporary directory outside the repository. Resolve and verify the
exact path before creation; do not target a broad home, workspace, or repo root.

## Step 3: Create the Minimum Checkout

Prefer reusing an existing branch carrying the same goal. Create a branch only
when the approved Git lifecycle requires independent history.

### Managed-host binding gate

When the host surface declares task/chat-bound managed workspaces, use its
native Worktree/Handoff lifecycle instead of shell creation. Reuse the current
task only when trusted host evidence says it is already bound to the intended
workspace. If the native operation creates or moves work into another task,
continue only in that bound task.

After the native lifecycle operation, or immediately when entering/reusing an
already-bound managed workspace, and before the first task content or Git
history write, read back one joint postcondition:

- trusted task workspace;
- default command `cwd` without a command-level directory override;
- intended Git worktree root;
- intended `HEAD` and branch/detached state.

All four must describe the same execution environment. `git worktree list`, a
managed-looking path, or a per-command `workdir` proves only part of that
postcondition. If native creation/handoff is unavailable, only UI-driven, or
cannot be verified, stop before shell worktree creation or task content/history
writes and give the user the host-native re-entry action. Do not silently fall
back on `git worktree add`.
Compare resolved path identity using host-appropriate semantics, not raw path
strings. The default `cwd` may equal the worktree root or be inside it; running
Git from that default `cwd` must resolve to the intended worktree root.
An intended host-managed detached HEAD satisfies the Git-state component; do
not create a branch merely to make that state look like a generic checkout.

If an unbound manual worktree already exists, preserve it and report its exact
path, `HEAD`, branch, status, and ownership evidence. Dirty, untracked, or
ownership-unknown state is not auto-migrated, deleted, stashed, reset, or
overwritten. A later commit/patch transfer follows normal Git authorization and
must be verified in the new bound task before cleanup is offered.

### Generic Git fallback

Use the shell path below only when trusted evidence positively classifies the
current host/CLI surface as `non-managed`. An `unknown` classification fails
closed; an absent native tool or missing binding metadata is not
generic-fallback evidence.

```bash
# Existing branch
git worktree add <exact-path> <branch>

# New independently justified branch
git worktree add -b <branch> <exact-path> <start-point>
```

Read back `git worktree list --porcelain`, the new worktree's `HEAD`/branch, and
its status. Do not use force flags. A command-level `workdir` may target this
checkout on a generic host, but it never proves managed task binding.

## Step 4: Authority-Led Setup and Baseline

Read the project's install docs, lockfiles, and existing setup scripts. Run
only setup the project explicitly requires. Do not infer `npm install`,
`cargo build`, `pip install`, `poetry install`, or dependency downloads merely
from manifest presence.

Run the smallest relevant baseline verification. If setup or tests fail, report
the exact failure and preserve the worktree; do not hide it by mutating unrelated
files.

## Step 5: Handoff and Cleanup Ownership

Report the exact path, branch, start `HEAD`, baseline result, owner, and cleanup
trigger. The creating coordinator remains responsible until it proves cleanup
or explicitly hands ownership off.

Use `finishing-a-development-branch` after the task commit or integration
decision. A worktree may usually be removed while its unmerged branch remains;
an open PR does not require a permanent checkout.

## Red Flags

Never:

- create a worktree only because a task modifies code;
- modify `.gitignore` merely to make worktree creation possible;
- install dependencies blindly;
- create one worktree per subagent;
- treat Git readback, a path prefix, or command-level `workdir` as proof of
  managed task binding;
- use shell worktree creation merely because a host-native capability is
  deferred, UI-only, or absent from the initial tool list;
- run global prune or force cleanup as routine hygiene;
- remove dirty, untracked, locked, user-owned, or ownership-unknown resources.

## Integration

Called conditionally by planning/execution workflows only after the necessity
decision. Pairs with `finishing-a-development-branch` for evidence-led cleanup.