worktree

Create Worktree Workspace

Create an isolated workspace with git worktrees for each repo, so a Claude Code instance can work on a feature branch in parallel without affecting the main checkout.

Input

Branch name: $ARGUMENTS

If no argument provided, ask the user for the branch name.

Step 1: Read config

Read .claude/worktree.json from the project root.

If the file does not exist:

1. Scan the current directory for subdirectories that are git repos ([ -d "$dir/.git" ])
2. Auto-detect install/start commands by reading package.json, pyproject.toml, Makefile, etc. in each repo. Also detect cross-service relationships (e.g., a frontend env var pointing to a backend API).
3. Create .claude/worktree.json populated with the discovered repos. Use what you detected — only leave <FILL IN> placeholders for fields you truly cannot determine:

{
  "repos": [
    {
      "name": "<detected-dir>",
      "port": 8000,
      "install": "uv sync",
      "start": "uv run uvicorn app.main:app --port {port}",
      "env_files": [".env"],
      "env_patches": {
        "NEXT_PUBLIC_API_URL": "{backend.url}"
      },
      "cache_dirs": [".venv"]
    }
  ]
}

Important: For env_patches that reference other repos, always use {<repo_name>.url} placeholders. These resolve correctly in both portless and fallback modes. Never hardcode http://localhost:{<repo_name>.port} — use {<repo_name>.url} instead.

Note on start commands: The start field is the raw server command with a {port} placeholder. Do not include portless in the start command — portless wrapping is applied automatically when generating workspace scripts (Step 11).

4. If any <FILL IN> placeholders remain, print:

Created .claude/worktree.json — some fields need manual input.
Please fill in the <FILL IN> placeholders, then run /worktree again.

Stop here. Do not create any worktrees.

5. If all fields are filled (no <FILL IN> placeholders), print the config as a table and proceed to Step 2 immediately.

---

If the file does exist, proceed with the steps below.

Step 2: Derive the slug

SLUG=$(echo "$ARGUMENTS" | tr '/' '-')
WORKSPACE="../$(basename $(pwd))-wt-${SLUG}"

Step 3: Create the workspace root

mkdir -p "$WORKSPACE"

Step 4: Detect portless and assign service endpoints

Check if portless is installed:

command -v portless &>/dev/null

If portless is installed → Portless mode

For each repo, derive a portless service name:

SERVICE_NAME="<repo.name>-${SLUG}"

Example: repo backend-parsanai on branch feature/auth → backend-parsanai-feature-auth

Build a URL map for env patching:

repo.name → http://<SERVICE_NAME>.localhost:1355

No port assignment, offset counting, or conflict resolution needed. Skip the fallback section below.

If portless is NOT installed → Fallback mode

For each repo in the config, calculate its port:

OFFSET=$(cd <repo.name> && git worktree list | wc -l | tr -d ' ')

The port for each repo is repo.port + OFFSET.

If any assigned port is busy (lsof -i :<port>), increment all ports by 1 and check again.

Build a URL map for env patching:

repo.name → http://localhost:<assigned_port>

Step 5: Create worktrees

For each repo in the config:

cd <repo.name>
git worktree add "${WORKSPACE}/<repo.name>" -b "$ARGUMENTS" 2>/dev/null \
  || git worktree add "${WORKSPACE}/<repo.name>" "$ARGUMENTS"

If creation fails because the branch exists remotely but not locally:

git fetch origin "$ARGUMENTS"

Then retry.

Step 6: Copy root config

Copy the root CLAUDE.md and .claude/ directory so the workspace inherits shared commands, skills, and rules:

cp CLAUDE.md "$WORKSPACE/" 2>/dev/null || true
cp -r .claude "$WORKSPACE/" 2>/dev/null || true

Step 7: Generate worktree rules

Create ${WORKSPACE}/.claude/rules/worktree.md so the agent automatically knows its environment. This file is auto-loaded into context.

Make sure the .claude/rules/ directory exists:

mkdir -p "${WORKSPACE}/.claude/rules"

Generate the file with the following content (replace all placeholders with actual values):

Portless mode

# Worktree Workspace

You are working in an isolated worktree workspace on branch `<branch_name>`.

## Servers

Start all servers:
\`\`\`bash
./start.sh
\`\`\`

Stop all servers:
\`\`\`bash
./stop.sh
\`\`\`

<for each repo>
### <repo.name>
- Path: <repo.name>/
- URL: http://<service_name>.localhost:1355
- Start individually: cd <repo.name> && portless <service_name> <repo.start with {port} replaced by \$PORT>
<end for>

## Rules

- **No migrations.** Never run migration tools (alembic, prisma migrate, knex migrate, etc.) or any command that auto-runs migrations on startup. The database is shared across all worktrees.
- **Schema changes = blocker.** If your feature requires database schema changes, STOP and report back to the user.
- **Don't push** to remote unless explicitly asked.
- **Commit freely** to your feature branch — commits stay local until pushed.

## Cleanup

When done, the user can run `./cleanup.sh` from the workspace root to stop servers, remove worktrees, and delete this workspace.

Fallback mode

# Worktree Workspace

You are working in an isolated worktree workspace on branch `<branch_name>`.

## Servers

Start all servers:
\`\`\`bash
./start.sh
\`\`\`

Stop all servers:
\`\`\`bash
./stop.sh
\`\`\`

<for each repo>
### <repo.name>
- Path: <repo.name>/
- Port: <assigned_port>
- Start individually: cd <repo.name> && <repo.start with {port} replaced>
<end for>

## Rules

- **No migrations.** Never run migration tools (alembic, prisma migrate, knex migrate, etc.) or any command that auto-runs migrations on startup. The database is shared across all worktrees.
- **Schema changes = blocker.** If your feature requires database schema changes, STOP and report back to the user.
- **Don't push** to remote unless explicitly asked.
- **Commit freely** to your feature branch — commits stay local until pushed.

## Cleanup

When done, the user can run `./cleanup.sh` from the workspace root to stop servers, remove worktrees, and delete this workspace.

Step 8: Copy cached dependency directories

For each repo in the config, if repo.cache_dirs is defined and non-empty, clone each directory from the original repo into the worktree before installing. This avoids downloading everything from scratch.

On macOS (APFS), use cp -Rc for instant copy-on-write clones:

cp -Rc <repo.name>/<cache_dir> "${WORKSPACE}/<repo.name>/<cache_dir>"

On Linux, use hardlinks for speed:

cp -al <repo.name>/<cache_dir> "${WORKSPACE}/<repo.name>/<cache_dir>"

Skip any that don't exist in the original.

Hint for users configuring cache_dirs: APFS clones and hardlinks are fast for directories with few large files (e.g., .venv, vendor/), but slow for directories with many small files (e.g., node_modules). Package managers with global caches (pnpm, yarn berry, uv) often make a fresh install fast enough that cloning is unnecessary. When in doubt, leave cache_dirs empty and let the package manager handle it.

Step 9: Install dependencies

For each repo in the config:

cd "${WORKSPACE}/<repo.name>" && <repo.install>

This should be fast since cached directories were already copied.

Step 10: Copy and patch environment files

For each repo in the config:

1. Copy each file listed in repo.env_files from the original repo into the worktree:

   cp <repo.name>/<env_file> "${WORKSPACE}/<repo.name>/<env_file>" 2>/dev/null || true
   

2. Apply repo.env_patches: for each key-value pair, find and replace that key's value in the worktree's env file. Values support two kinds of placeholders:

   • {<repo_name>.url} — resolves to the full URL from the URL map built in Step 4. Works in both modes. Recommended.
   • {<repo_name>.port} — resolves to the assigned port number in fallback mode. In portless mode, if the value matches the pattern http://localhost:{<repo_name>.port}, replace the entire pattern with the portless URL from the URL map. This ensures backward compatibility with existing configs.

Step 11: Generate workspace scripts

Create executable scripts in the workspace root so agents (or users) can start/stop servers without knowing the details.

Portless mode

start.sh

#!/usr/bin/env bash
set -e
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"

<for each repo>
echo "Starting <repo.name> → http://<service_name>.localhost:1355"
(cd "$SCRIPT_DIR/<repo.name>" && portless <service_name> <repo.start with {port} replaced by \$PORT>) &
<end for>

echo ""
echo "All servers started."
echo "Run ./stop.sh to shut them down."
wait

stop.sh

#!/usr/bin/env bash

<for each repo>
echo "Stopping <service_name>..."
pkill -f "portless <service_name>" 2>/dev/null || true
<end for>

echo "All servers stopped."

Fallback mode

start.sh

#!/usr/bin/env bash
set -e
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"

# Start each repo's server in the background
<for each repo>
echo "Starting <repo.name> on port <assigned_port>..."
(cd "$SCRIPT_DIR/<repo.name>" && <repo.start with {port} replaced>) &
<end for>

echo ""
echo "All servers started. PIDs logged above."
echo "Run ./stop.sh to shut them down."
wait

stop.sh

#!/usr/bin/env bash

<for each repo>
echo "Stopping port <assigned_port>..."
lsof -ti :<assigned_port> | xargs kill 2>/dev/null || true
<end for>

echo "All servers stopped."

Both modes: cleanup.sh

#!/usr/bin/env bash
set -e
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
PARENT_DIR="$(dirname "$SCRIPT_DIR")"
WORKSPACE_NAME="$(basename "$SCRIPT_DIR")"

# Stop any running servers
"$SCRIPT_DIR/stop.sh" 2>/dev/null || true

# Remove worktrees
<for each repo>
echo "Removing worktree <repo.name>..."
git -C "$PARENT_DIR/<original_project>/<repo.name>" worktree remove "$SCRIPT_DIR/<repo.name>" 2>/dev/null || true
<end for>

# Remove workspace directory
echo "Removing workspace..."
rm -rf "$SCRIPT_DIR"
echo "Cleanup complete."

Make all scripts executable:

chmod +x "${WORKSPACE}/start.sh" "${WORKSPACE}/stop.sh" "${WORKSPACE}/cleanup.sh"

Step 12: Report

Portless mode

Workspace ready for branch: $ARGUMENTS
  Root: <WORKSPACE>/

<for each repo>
  <repo.name>:
    Path: <WORKSPACE>/<repo.name>
    URL:  http://<service_name>.localhost:1355
<end for>

Scripts:
  ./start.sh    — start all servers
  ./stop.sh     — stop all servers
  ./cleanup.sh  — stop servers, remove worktrees, delete workspace

Launch agent:
  cd <WORKSPACE> && claude --dangerously-skip-permissions

Fallback mode

Workspace ready for branch: $ARGUMENTS
  Root: <WORKSPACE>/

<for each repo>
  <repo.name>:
    Path: <WORKSPACE>/<repo.name>
    Port: <assigned_port>
<end for>

Scripts:
  ./start.sh    — start all servers
  ./stop.sh     — stop all servers
  ./cleanup.sh  — stop servers, remove worktrees, delete workspace

Tip: Install portless for stable named URLs → npm i -g portless
     https://github.com/vercel-labs/portless

Launch agent:
  cd <WORKSPACE> && claude --dangerously-skip-permissions