worktrunk

The full skill.

Original instructions from the publisher’s SKILL.md

# Worktrunk

Help users work with Worktrunk, a CLI tool for managing git worktrees.

## Available documentation

Reference files are synced from [worktrunk.dev](https://worktrunk.dev) documentation:

- **reference/config.md**: User and project configuration (LLM, hooks, command defaults)
- **reference/hook.md**: Hook types, timing, and execution order
- **reference/switch.md**, **merge.md**, **list.md**, etc.: Command documentation
- **reference/extending.md**: Aliases, multi-step pipelines, custom subcommands, and template-expansion gotchas (two-pass `{% raw %}` deferral, for-each recipes)
- **reference/llm-commits.md**: LLM commit message generation
- **reference/tips-patterns.md**: Practical recipes — aliases, per-branch variables, dev server per worktree, parallel agent patterns
- **reference/shell-integration.md**: Shell integration debugging
- **reference/troubleshooting.md**: Troubleshooting for LLM and hooks (Claude-specific)

For command-specific options, run `wt <command> --help`. For configuration, follow the workflows below.

## Which worktree a command acts on

`wt` finds the *repository* from the working directory, and the *worktree* from the command's own arguments. Two rules cover every case:

1. **A command that names a branch already names its worktree.** Worktrees are addressed by branch name, so `wt switch <branch>`, `wt remove <branch>`, `wt step diff --branch <branch>`, and `wt config state marker set --branch <branch>` act on that branch's worktree no matter which worktree you run them from. Every such argument also accepts the worktree's own path, for the cases a branch cannot name — a second checkout of the same branch, or a detached worktree (which `marker` still rejects, since it keys state by branch name).
2. **`-C <path>` moves the working directory, not the worktree selection.** Reach for it when the repository lookup is what's wrong: a *different* repository; a command that acts on the current worktree and takes no branch argument (`wt merge`, `wt step rebase|squash|push` — their `[TARGET]` is the merge target, not a worktree); or a caller whose working directory isn't inside a repository at all, such as an agent hook the host pins elsewhere.

Layering `-C` on top of a branch argument names the same worktree twice. From inside the `alpha` worktree of a repo that also has `beta`:

```bash
wt step diff --branch beta                    # ✓ the branch argument selects the worktree
wt -C ../repo.beta step diff --branch beta    # ✗ says beta twice

wt switch --create beta                       # ✓ --base already defaults to the default branch
wt -C ../repo switch --create beta            # ✗ -C adds nothing; you are already in that repo
```

## Two types of configuration

Worktrunk uses two config files with different scopes and permission models:

**User config** (`~/.config/worktrunk/config.toml`, never checked into git) holds personal preferences: LLM integration, worktree path templates, command settings, user hooks. Treat it conservatively — propose changes and get consent before editing, never install tools on the user's behalf, and preserve the file's existing structure and comments. See `reference/config.md`.

**Project config** (`<repo>/.config/wt.toml`, checked into git) holds team-wide automation: hooks for the worktree lifecycle (pre-start, pre-merge, etc.). Edit proactively — changes are versioned and reversible via git. Comment why each hook exists, and warn the user before adding destructive commands (`rm -rf`, `DROP TABLE`), network fetches piped to shells, or `sudo`. See `reference/hook.md`.

Some requests span both: commit-message generation is user config, while the team's quality checks are project config.

## Core workflows

### Setting up commit message generation (user config)

Detect which tools are installed (`which claude codex llm aichat`); if none, recommend Claude Code. Take the exact command for the chosen tool from `reference/llm-commits.md`, propose the `[commit.generation]` change, and apply it after approval (`wt config create` first if no config exists). To verify, `wt step commit --dry-run` renders the prompt, runs the LLM, and prints the message without committing.

### Configuring project hooks

Pick the hook type by when the command should run and whether it may block — `reference/hook.md` maps all ten (5 events × pre/post) to their timing and typical uses.

Derive the commands from the project itself (`package.json` scripts, `Cargo.toml`, `pyproject.toml`) and verify they run before adding them.

When a new hook must wait for an existing one, convert the entry to a pipeline; independent commands in a named table run concurrently:

```toml
# Pipeline: install completes before migrate starts
[[pre-start]]
install = "npm install"

[[pre-start]]
migrate = "npm run db:migrate"

# Concurrent: independent commands in one table
[pre-start]
install = "npm install"
env = "cp .env.example .env"
```

Test with `wt switch --create test-hooks`.

## Common tasks reference

### User config tasks
- Set up commit message generation → `reference/llm-commits.md`
- Customize worktree paths → `reference/config.md#worktree-path-template`
- Custom commit templates → `reference/llm-commits.md#prompt-templates`
- Configure command defaults → `reference/config.md#command-config`
- Set up personal hooks → `reference/config.md#user-hooks`

### Project config tasks
- Set up hooks for new project → `reference/hook.md`
- Add hook to existing config → `reference/hook.md#hook-forms`
- Use template variables → `reference/hook.md#template-variables`
- Add dev server URL to list → `reference/config.md#dev-server-url`

### Aliases & multi-worktree tasks
- Create a `wt` alias → `reference/extending.md#aliases`
- Run a command in every worktree → `reference/step.md#wt-step-for-each`
- Rebase every worktree (up-style) → `reference/extending.md#recipe-rebase-every-worktree-onto-its-upstream`
- Pass a template to a nested `wt` command → `reference/extending.md#nesting-templates`

## Key commands

```bash
# View all configuration
wt config show

# Create initial user config (LLM/commit setup: see reference/llm-commits.md)
wt config create

# Full config reference (subcommands, templates, env vars)
wt config --help
```

## Hook approvals in non-interactive sessions

Worktrunk never runs a project's hooks or aliases until the user has explicitly approved them. The commands in `.config/wt.toml` are arbitrary shell code shipped in a repository the user may have just cloned, so on first run Worktrunk shows each command and waits for the user to approve it — an untrusted `.config/wt.toml` cannot silently execute anything. Approvals are stored per-project in `~/.config/worktrunk/approvals.toml` and re-prompted whenever a command template changes, so a hook can't be swapped for a different command after it was approved.

Agents running `wt merge`, `wt switch`, or other commands that trigger hooks will hit an error like:

```
▲ cargo-difftest needs approval to execute 1 command:
○ post-merge install:
  cargo install --path .
✗ Cannot prompt for approval in non-interactive environment
↳ To skip prompts in CI/CD, add --yes; to pre-approve commands, run wt config approvals add --yes
```

The resolution is for the user to make the trust decision themselves:

- **`wt config approvals add`** — interactive prompt where the user reviews each command before it is stored to `~/.config/worktrunk/approvals.toml`. Run once per project; the approval persists across invocations until the command template changes or the project moves. This is the path to recommend — the user reviews and consents to exactly the commands that will run.

**When invoked as an agent, stop and escalate to the user.** Approving a project's hooks is a security decision about whether this repository should be trusted to run arbitrary commands on the user's machine — that decision belongs to the user, not the agent. Tell the user to run `wt config approvals add` and let them review the commands. Do not reach for either `--yes` on the user's behalf: on the blocked command it skips the gate for that invocation, and `wt config approvals add --yes` records every command the project declares with nobody reading them. Both exist for CI/CD pipelines and containers that already control their own hook contents; neither is a shortcut for an interactive agent to silence an approval prompt.

## Advanced: agent handoffs

When the user requests spawning a worktree with an agent in a background session ("spawn a worktree for...", "hand off to another agent"), use the tmux or Zellij command from `reference/tips-patterns.md#agent-handoffs` with the CLI you are running as in place of `claude`, following that section's note on where a subcommand such as OpenCode's `run` goes.

**Requirements** (all must be true):
- User explicitly requests spawning/handoff
- User is in a supported multiplexer (check `$TMUX` / `$ZELLIJ`)
- The user's project instructions (`CLAUDE.md` or `AGENTS.md`) or an explicit prompt authorize this pattern

**Do not use this pattern** for normal worktree operations.

### Parallel sub-Agents (single Claude Code session)

To spawn multiple sub-Agents that each work in their own worktree from one Claude Code session — no terminal multiplexer, no human in the other pane — pre-start each worktree from the parent and pass the path into the sub-Agent prompt:

```bash
wt switch --create <branch> --no-cd
```

Then call the `Agent` tool **without** `isolation: "worktree"`, naming the path in the prompt:

```
You are working in `/abs/path/to/myproject.<branch>` on branch `<branch>`.
All edits must stay in that worktree.
```

`--no-cd` skips the shell-integration cd script the parent can't consume. Add `--no-hooks` only when no user or project hooks provision the worktree and each sub-Agent does its own build/test step (e.g. `cargo run -- hook pre-merge --yes`) — a `pre-start` hook that installs dependencies or links a gitignored build environment hands the sub-Agent a worktree it can't build in when it's skipped. Leaving hooks on needs the project's hook commands already approved: the parent session can't prompt, so an unapproved command in `.config/wt.toml` aborts the create — see **Hook approvals in non-interactive sessions** above.

**Do not** use `Agent { isolation: "worktree" }` for this. Claude Code passes its internal agent ID as `name` to the `WorktreeCreate` hook, so `wt` creates the worktree as `myproject.agent-<id>` on a throwaway branch. If the sub-Agent then creates a feature branch on top, you end up with non-canonical paths, orphan branches, and post-start hooks fired against the wrong branch. Pre-creating with `wt switch --create` keeps path, branch, and hook target aligned.