Worktree Toolkit

A streamlined workflow for developing in large Bazel + IntelliJ monorepos using Git worktrees.

Enables instant IntelliJ context switching between worktrees—no re-imports, no re-indexing—and scales to support parallel development by humans and AI agents alike.

Git worktrees let you work on multiple branches in parallel, but IntelliJ treats each worktree as a separate project, requiring expensive Bazel syncs and index rebuilds every time you switch to a new worktree.

This toolkit makes IntelliJ context switching instant by:

• Symlink trick: IntelliJ always opens the same path; switching worktrees looks like a branch checkout → incremental refresh in seconds, not minutes
• Metadata vault: IDE project metadata (.ijwb, .idea, .vscode, etc.) is stored externally and automatically installed into every new worktree—no manual IDE setup needed
• Safe worktree management: New branches are created from a freshly fetched base ref without ever touching the main checkout, plus cleanup of merged branches
• Parallel development at scale: Works for humans and AI agents alike

📊 See the presentation slides for a visual walkthrough.

# Install (interactive prompts for configuration)
./install.sh

# Reload shell
source ~/.zshrc

# Use
wt help

The installer will:

1. Copy the toolkit to ~/.wt/
2. Add sourcing to your shell rc file
3. Prompt for workspace paths (main repo, worktrees, metadata vault)
4. Create required directories
5. Optionally migrate existing repo to worktree structure
6. Optionally export project metadata to the vault
7. Optionally set up a nightly cron job to refresh Bazel IDE metadata

The directory structure expected (controlled by environment variables, can be overwritten):

~/Development/
├── java -> java-master          # Symlink (IntelliJ opens this)
├── java-master/                 # Main repository
├── java-worktrees/              # Worktrees go here
└── idea-project-files/          # Project metadata vault

                      ┌─────────────────────────────────────────────┐
                      │   External Project Metadata Vault           │
                      │  ~/Development/idea-project-files           │
                      │    (IDE configs: .ijwb, .idea, etc.)        │
                      └──────────▲───────────────┬──────────────────┘
                                 │               │
                                 │               │
           ┌──wt metadata-export─┘               └──wt metadata-import─┐
           │                                                           │
┌──────────┴───────────────────────┐                     ┌─────────────▼──────────────────────┐
│   Main Repository                │                     │    Worktrees                       │
│ ~/Development/java-master        │       wt add        │ ~/Development/java-worktrees/...   │
│  • master branch                 │ ──────────────────► │  • feature/foo                     │
│  • never touched by wt add       │(calls metadata-imp) │  • bugfix/bar                      │
│  • never removed                 │                     │  • agent-task-123                  │
└───────────────┬──────────────────┘                     └─────────┬──────────────────────────┘
                │                                                  │
             wt switch                                          wt remove
                │                                                  │
    ┌───────────▼──────────────────┐                     ┌─────────▼────────────┐
    │ Stable IntelliJ Project Dir  │                     │  Safe cleanup with   │
    │ ~/Development/java           │                     │  confirmation prompt │
    │ (symlink updated per switch) │                     └──────────────────────┘
    └───────────▲──────────────────┘
                │ 
          IntelliJ auto-refresh
                │
     ┌──────────▼───────────────────┐
     │ IntelliJ loads worktree      │
     │ instantly (no import needed) │
     └──────────────────────────────┘

# Existing branch
wt add feature/foo

# New branch (from latest master)
wt add -b feature/foo

When creating with -b, the script:

1. Fetches the latest base branch (git fetch origin <base>); a failed fetch warns and falls back to the last-known base ref
2. Creates branch + worktree from origin/<base> (or the local base branch when there is no remote), without ever touching the main repo checkout — no stash, no branch switch
3. Imports project metadata from vault
4. Copies configured seed files (WT_SEED_FILES, e.g. user.bazelrc) from the main repo

Each worktree gets its own Bazel output base (Bazel derives it from the worktree path), so builds in different worktrees never clobber each other. wt remove reclaims that disk space automatically.

Set WT_SKIP_PULL=1 to skip the git fetch step (useful offline or in scripts):

WT_SKIP_PULL=1 wt add -b feature/foo

# Interactive
wt switch

# Direct
wt switch ~/Development/java-worktrees/feature/foo

Updates the symlink so IntelliJ instantly loads the new worktree.

# Interactive cd
wt cd

# Direct cd
wt cd ~/Development/java-worktrees/feature/foo

wt list
wt list -v            # Adds dirty/ahead/behind indicators (slower; worktrees probed in parallel)
wt list --porcelain   # Machine-readable output (for scripts and agents)

Shows all worktrees with status indicators:

• * = Currently linked worktree
• [main] = Main repository root
• [linked] = Active symlink target
• [unadopted] = Worktree not adopted by wt (fix with wt adopt)
• [dirty] = Has uncommitted changes (with -v)
• [↑N] / [↓N] = Commits ahead/behind upstream (with -v)

--porcelain prints git worktree list --porcelain output augmented with extra lines per worktree: wt.active, wt.adopted, and (with -v) wt.dirty, wt.ahead N, wt.behind N.

Worktrees created outside wt (e.g. with plain git worktree add) show as [unadopted] in wt list. Adoption imports project metadata from the vault and marks the worktree as managed by wt:

wt adopt                          # Adopt the worktree at the current directory
wt adopt <worktree-path|branch>   # Adopt a specific worktree
wt adopt --redo                   # Re-run adoption on an adopted worktree
wt adopt --force                  # Skip conflict checks, overwrite without prompting

The main repository cannot be adopted (only worktrees).

# Interactive
wt remove

# Direct (with confirmation)
wt remove ~/Development/java-worktrees/feature/foo

# Skip confirmation (unless uncommitted changes exist)
wt remove -y ~/Development/java-worktrees/feature/foo

# Also delete the git branch
wt remove -b feature/foo

# Remove all worktrees with branches merged into base branch
wt remove --merged

# Remove merged without prompts, skipping worktrees with uncommitted changes
wt remove --merged -y --on-dirty=skip

Flags:

• -y/--yes = Skip confirmation prompts (dirty worktrees still prompt unless --on-dirty says otherwise)
• -b/--branch = Also delete the git branch after removing the worktree
• --merged = Remove all worktrees whose branches are merged (regular + squash)
• --on-dirty=MODE = What to do with uncommitted changes: warn (default, prompts), skip, or remove

Safety features:

• Warns if the worktree is currently linked (symlink will be switched to main repo)
• Warns if there are uncommitted changes (shows summary)
• Always prompts for confirmation if uncommitted changes exist, even with -y (use --on-dirty=skip or --on-dirty=remove to override)
• --merged mode: automatically finds and removes all worktrees whose branches are merged

After a successful removal, wt remove also reaps the worktree's dedicated Bazel output base (best-effort): it locates the output base named after the md5 hash of the worktree path under the known Bazel output user roots (~/Library/Caches/bazel/_bazel_$USER, /private/var/tmp/_bazel_$USER, ~/.cache/bazel/_bazel_$USER), verifies it contains execroot/, deletes it, and reports the disk space freed. If no matching output base exists, nothing is printed; cleanup failures only warn and never fail the removal.

wt can manage worktrees for multiple repositories. Each repository is a named context whose configuration lives in ~/.wt/repos/<name>.conf:

wt context                       # Interactive: pick context to switch to
wt context <name>                # Switch to named context
wt context --list                # List all available contexts
wt context add                   # Add a new repository context (interactive)
wt context add <path>            # Add context for repository at path
wt context add <name> <path>     # Add context with specific name and path
wt context remove [name]         # Remove a context and clean up all wt config

# Export metadata from main repo to vault (run after setting up new IDE projects)
wt metadata-export

# Import metadata into a worktree (interactive selection if target omitted)
wt metadata-import
wt metadata-import ~/Development/java-worktrees/feature/foo

# Skip confirmation prompts (useful in scripts)
wt metadata-export -y
wt metadata-import -y ~/Development/java-worktrees/feature/foo

When most development work is done in worktrees, the Bazel IDE directories (.ijwb, .aswb, .clwb) in the main repository can become stale (targets files don't reflect new Bazel targets).

The bin/wt-metadata-refresh script is designed to run as a cron job to keep metadata current.

Note: When IntelliJ has derive_targets_from_directories: true in .bazelproject (the default), it queries Bazel fresh on every sync. The targets-* file serves as a cache for initial project imports and may improve import speed.

Note: The installer (install.sh) offers to set up this cron job automatically (default: yes).

To set it up manually:

# Create log directory
mkdir -p ~/.wt/logs

# Edit crontab
crontab -e

# Add this line to run nightly at 2am (uses login shell for full PATH):
0 2 * * * /bin/zsh -lc '~/.wt/bin/wt-metadata-refresh' >> ~/.wt/logs/metadata-refresh.log 2>&1

You can also run the script manually:

# Refresh all Bazel IDE directories and re-export to vault
~/.wt/bin/wt-metadata-refresh

# Preview what would be refreshed (dry run)
~/.wt/bin/wt-metadata-refresh --dry-run

# Refresh targets files only (skip re-export step)
~/.wt/bin/wt-metadata-refresh --no-export

The refresh script:

• Uses bazel query to regenerate targets/targets-* files in each Bazel IDE directory
• Supports all Bazel patterns configured in WT_METADATA_PATTERNS (.ijwb, .aswb, .clwb)
• Parses .bazelproject to determine which directories to include in the query
• Preserves existing targets file hashes (IntelliJ may reference them)
• Re-exports all metadata to the vault (including non-Bazel patterns)
• Logs timestamped output for monitoring
• Returns exit codes: 0=success, 1=error, 2=partial success

The scripts rely on a few environment variables to know where your main repository, worktrees, and IntelliJ metadata live.

These variables are normally read from the per-context config file (~/.wt/repos/<name>.conf, written by wt context add). If set in your shell configuration, they take precedence over the context config (except in wt.enabled repos, where git local config has highest priority).

wt commands require a config source to actually load: if neither a context .conf nor a repo's wt.* git local config is found, or if WT_MAIN_REPO_ROOT does not point at an existing git repository, commands exit with an error directing you to wt context add. The built-in defaults below exist only so sourcing wt.sh never breaks an unconfigured shell.

Path to your primary git repository clone.

Default: ~/.wt/repos/repo/base

export WT_MAIN_REPO_ROOT="$HOME/Development/java-master"

Used by:

• wt-add (base branch operations)
• wt-choose (listing worktrees)
• wt-switch (default symlink target)
• wt-remove (safety check to prevent removing main repo)

Directory where new worktrees are created by default.

Default: ~/.wt/repos/repo/worktrees

export WT_WORKTREES_BASE="$HOME/Development/java-worktrees"

Canonical metadata vault storing project metadata (IDE configs, etc.).

Default: ~/.wt/repos/repo/idea-files

export WT_IDEA_FILES_BASE="$HOME/Development/idea-project-files"

Used by:

• wt-metadata-import
• wt-metadata-export
• wt-metadata-refresh
• wt-add (when installing metadata)

Symlink path that points to the currently active worktree. This is where IntelliJ should open the project.

Default: ~/Development/java

export WT_ACTIVE_WORKTREE="$HOME/Development/java"

Used by:

• wt-switch (updates this symlink)
• wt-remove (warns if removing the linked worktree)

Name of the mainline branch to branch from.

Default: master

export WT_BASE_BRANCH="master"

Space-separated, repo-root-relative names of files to copy (cp -p) from the main repo into every new/adopted worktree. Useful for gitignored per-machine files like user.bazelrc or .bazelversion overrides that a fresh checkout would otherwise lack.

Seeding is best-effort: files missing in the main repo are skipped, files already present in the worktree are never overwritten, and copy failures only warn.

The context setup wizard (wt context add / install.sh) prompts for this and preserves the previously configured value when a context is replaced.

Default: (empty)

export WT_SEED_FILES="user.bazelrc .bazelversion"

Used by:

• wt-add (via adoption treatment)
• wt-adopt

Space-separated list of project metadata patterns to preserve across worktrees.

Default: empty (context setup typically sets .ijwb)

export WT_METADATA_PATTERNS=".ijwb .idea .vscode"

Used by:

• wt-metadata-import / wt-metadata-export
• wt-metadata-refresh
• wt-add and wt-adopt (when installing metadata)

A 10-minute overview presentation is available in the presentation/ directory:

• slides.md — Marp markdown source
• slides.pdf — Generated PDF

To regenerate the PDF from the markdown:

npx @marp-team/marp-cli presentation/slides.md -o presentation/slides.pdf

wt/
├── wt.sh                    # Entry point (source this)
├── presentation/            # Overview slides
├── bin/                     # Executable commands
│   ├── wt-add
│   ├── wt-adopt
│   ├── wt-cd
│   ├── wt-context
│   ├── wt-list
│   ├── wt-remove
│   ├── wt-switch
│   ├── wt-metadata-import
│   ├── wt-metadata-export
│   └── wt-metadata-refresh  # Cron script to refresh Bazel IDE metadata
├── lib/                     # Shared libraries
│   ├── wt-common            # Configuration and helpers
│   ├── wt-adopt             # Worktree adoption helpers
│   ├── wt-choose            # Interactive worktree selection
│   ├── wt-context           # Multi-repo context management
│   ├── wt-context-setup     # Context creation (wt context add)
│   └── wt-help              # Help text for wt command
├── completion/              # Shell completions for wt-* scripts
│   ├── wt.zsh
│   └── wt.bash
├── install.sh
└── README.md

You can also run the underlying scripts directly:

wt-add, wt-adopt, wt-switch, wt-remove, wt-list, wt-cd, wt-context, wt-metadata-export, wt-metadata-import

These are located in bin/ and work identically to the wt subcommands.

The bin/ directory is the subcommand registry: wt <name> dispatches to any executable bin/wt-<name>, and shell completion derives the top-level command list from the same directory. Only cd, remove, and context have dedicated dispatch branches (they need in-shell behavior), so adding a new subcommand is just dropping an executable bin/wt-<name> script.

The bin/wt-metadata-refresh script is designed for cron jobs; run it directly or as wt metadata-refresh.