Worktrunk — Summary

Worktrunk (wt) is a Rust CLI for git worktree management explicitly designed for running AI agents in parallel. Its three core commands (wt switch, wt list, wt merge) make worktrees as easy to use as branches: wt switch -c feature creates a new branch + worktree in one command; wt list shows all worktrees with status, CI state, and AI-generated summaries; wt merge squashes, rebases, fast-forward merges, and cleans up in one step. Additional features include hooks (pre/post-merge, post-start, etc.), LLM commit message generation, an interactive picker with diff/log preview, copy-of-build-caches between worktrees, and a Claude Code plugin that shows agent activity indicators (🤖/💬) in wt list. Unlike the other frameworks in this batch (E2B, Daytona, microsandbox, brood-box, arrakis) which provide sandbox isolation, Worktrunk provides git-worktree-based isolation — it manages the host filesystem workspace isolation rather than VM/container execution isolation. Among the seed frameworks, it is most similar to the isolation mechanism used by superpowers (git-worktree) and taskmaster-ai, but Worktrunk focuses entirely on the worktree management UX rather than agent behavioral augmentation.

Worktrunk — Overview

Origin

Worktrunk (max-sixty/worktrunk) was created by Maxime Roos and first released at the start of 2026. By May 2026 it has 5,197 stars (the second-highest in this batch after Daytona) and 30 contributors. Licensed MIT OR Apache-2.0. Built in Rust.

Philosophy

From the README: "AI agents like Claude Code and Codex can handle longer tasks without supervision, such that it's possible to manage 5-10+ in parallel. Git's native worktree feature give each agent its own working directory, so they don't step on each other's changes. But the git worktree UX is clunky."

Key design values:

• Worktrees addressed by branch name (not filesystem paths)
• One command creates both branch + worktree: wt switch -c feature (vs git worktree add -b feat ../repo.feat && cd ../repo.feat)
• Parallel agent coordination: First-class support for running 5-10+ agents concurrently
• Data safety above all: "Never risk data loss without explicit user consent. A failed command that preserves data beats a 'successful' one that silently destroys work."
• Maturing mode: growing user base, external-interface breaks need justification

Context

Worktrunk explicitly links to Anthropic's "Claude Code Best Practices" guide which recommends the worktree pattern. It is the reference implementation of that pattern as a CLI.

Claude Code Plugin

Worktrunk ships as a Claude Code plugin (.claude-plugin/marketplace.json) with:

• A skill (worktrunk/SKILL.md) for configuration guidance
• Hooks that track Claude agent activity per worktree (showing 🤖/💬 in wt list)

Worktrunk — Architecture

Distribution

• Type: Standalone CLI binary
• Install (Homebrew): brew install worktrunk && wt config shell install
• Install (Cargo): cargo install worktrunk && wt config shell install
• Install (Windows): winget install max-sixty.worktrunk
• Install (Arch): sudo pacman -S worktrunk
• Install (Conda/Pixi): conda install -c conda-forge worktrunk
• Shell integration (wt config shell install) is required for directory-changing commands

Repo Structure

worktrunk/
├── src/
│   ├── main.rs            — entry point
│   ├── cli/               — Clap CLI definitions (mod.rs, config.rs, hook.rs, list.rs, step.rs)
│   ├── commands/          — Command handlers
│   │   ├── worktree/      — switch, remove, list, merge
│   │   ├── step/          — commit, diff, copy-ignored, prune, etc.
│   │   ├── hooks.rs       — hook execution
│   │   └── ...
│   ├── git/               — Repository abstraction + git operations
│   ├── config/            — User + project config parsing (TOML)
│   ├── llm.rs             — LLM commit message generation
│   └── shell/             — Shell integration
├── skills/
│   └── worktrunk/
│       ├── SKILL.md       — Claude Code skill (guidance for wt config)
│       └── reference/     — Auto-synced docs from worktrunk.dev
├── hooks/
│   └── hooks.json         — Claude Code / Codex / Gemini hooks
├── .claude/
│   └── skills/            — Subset of skills for Claude Code
├── .agents/               — Codex/other agent plugin files
├── .claude-plugin/
│   └── marketplace.json   — Claude Code plugin marketplace entry
├── plugins/worktrunk/     — Full plugin directory
├── CLAUDE.md              — Developer contribution guide
└── Cargo.toml             — Rust workspace

Config Files

Target Runtime

• Rust stable (latest − 1 MSRV)
• Any OS (macOS, Linux, Windows)
• No container/VM dependency

Worktrunk — Components

Core Commands (wt)

Hook Types

From hooks.json and docs:

• BeforeAgent / AfterAgent — track Claude agent activity (🤖/💬 markers)
• SessionEnd — clear activity marker
• pre-merge — runs before merge (tests, lints)
• post-merge — runs after merge
• pre-start — runs when creating worktree
• post-start — runs after switching (install deps, start dev server)
• post-switch — runs after directory switch
• pre-commit — runs before commit

Claude Code Plugin

hooks.json (BeforeAgent, AfterAgent, SessionEnd):

{
  "hooks": {
    "BeforeAgent": [{ "hooks": [{ "type": "command",
      "command": "bash ... wt config state marker set 🤖 || true" }] }],
    "AfterAgent":  [{ "hooks": [{ "type": "command",
      "command": "bash ... wt config state marker set 💬 || true" }] }],
    "SessionEnd":  [{ "hooks": [{ "type": "command",
      "command": "bash ... wt config state marker clear || true" }] }]
  }
}

Skill (worktrunk/SKILL.md): Guidance for configuring wt.toml and ~/.config/worktrunk/config.toml, setting up LLM commit messages, and debugging hooks.

Skills (3 total in .claude/skills/)

• release — guide for cutting a Worktrunk release
• running-tend — operating in CI / tend workflows
• writing-user-outputs — style guide for user-visible strings

Worktrunk — Uniqueness & Positioning

differs_from_seeds

Worktrunk is the only framework in this batch that uses git-worktree isolation (matching the seed archetype). Among seeds, superpowers uses worktree_per_feature: yes and taskmaster-ai uses worktree: yes for parallel agent work — but these are incidental to their core value (skills/hooks/MCP). Worktrunk's entire value is worktree management UX: it treats worktrees as first-class citizens, makes the create/switch/merge/cleanup lifecycle as simple as branch operations, and adds observability (activity markers, CI status, LLM summaries). Unlike all seeds (which inject behavioral instructions into agents), Worktrunk only manages the filesystem isolation layer and adds workflow automation (hooks, LLM commits). It is closer to an infrastructure tool like brood-box/microsandbox than to superpowers/BMAD — but its isolation is git-based rather than VM-based.

Positioning Within This Batch

Worktrunk is unique in this batch:

• Only git-worktree isolation (vs microVM/container isolation of all other frameworks)
• Only Rust CLI focused purely on worktree UX (vs Go/TypeScript/Python)
• Only Claude Code plugin in this batch that tracks agent activity across multiple worktrees
• Most complete parallel-agent UX: wt list --full with CI status + AI summaries is unlike anything in the other frameworks

vs other batch members:

• vs brood-box: brood-box uses microVM; Worktrunk uses git-worktree; different isolation layers
• vs vibekit: VibeKit dispatches agents; Worktrunk manages the filesystem workspace for agents

Distinctive Opinion

The bottleneck in running many parallel AI agents is not VM isolation (which git worktrees handle adequately) but worktree lifecycle UX — creating, monitoring, merging, and cleaning up multiple branches is too clunky with raw git commands.

Explicit Antipatterns

• Running agents in the same directory (file conflicts)
• Force-removing worktrees that have unsaved data
• Silent destructive side effects during unrelated operations
• git reset --hard (favors --keep), git checkout --force (favors --merge)

Observable Failure Modes

1. Git-only isolation: No VM/container boundary — malicious agent code can affect the host OS
2. Shell integration required: wt config shell install must be run for directory-changing commands; may conflict with other shell customizations
3. Windows: wt conflicts with Windows Terminal; winget installs as git-wt
4. MSRV: Rust latest − 1; bumped during weekly tend maintenance (compatibility risk)
5. LLM commit dependency: wt step commit requires a configured LLM tool; if unavailable, the command fails

Cross-References in This Batch

• Claude Code plugin hooks (BeforeAgent/AfterAgent/SessionEnd) interact with Claude Code's agent lifecycle events
• Referenced in Anthropic's "Claude Code Best Practices" guide
• tend (from the same author) is used for dependency maintenance (running-tend skill)