Git Integration & Worktrees

Parallel work without git stash hell

Claude Code is git-native. Every file edit creates a checkpoint; /rewind or Esc Esc rolls back. The agent commits, branches, rebases, drafts PRs through plain natural language. The bigger pattern: worktrees turn a single repo into N parallel workspaces, one Claude session per worktree, no file collisions.

Worktrees solve the actual problem: you have three half-finished branches, you don't want to git stash through context, and you definitely don't want them sharing a working directory. git worktree add creates a sibling directory pinned to a different branch. Open a Claude session in each. They share commits, they don't share files.

Claude Code's --worktree <name> automates the dance: creates .claude/worktrees/<name>/, branches off origin/HEAD, drops you in. /batch takes it further — decomposes a big refactor into independent units, spawns a background agent per unit in its own worktree, each opens a PR. You review each in isolation.

Code

Parallel sessions, no stash·bash

# Three worktrees in three terminals
claude --worktree feature-auth     "implement OAuth2 login"
claude --worktree bugfix-payment   "fix the null in checkout"
claude --worktree perf-search      "profile and fix slow search"

# All three running simultaneously
# Each touches its own files, all share git history

# Manual equivalent
git worktree add -b feat/x ../x origin/main
cd ../x && claude

# Cleanup when shipped
git worktree remove ../x

/batch — fleet of agents in one shot·bash

# Inside a session
/batch migrate every Express 4 route in src/routes/ to Express 5

# Claude:
# 1. Lists matching files (decomposition)
# 2. Shows you a plan with one unit per file
# 3. On approval: spawns one background agent per unit
# 4. Each agent in its own worktree, implements, runs tests
# 5. Each unit opens its own PR
# You review and merge independently

Git Integration & Worktrees

Parallel work without git stash hell

Code

External links

Exercise

Progress

Comments 0