workbench

workbench-kit

A bench for shapeshifters tired of spinning up one more repo and one more AGENTS.md.

English · 한국어

---

Every new idea, another repo. Another AGENTS.md. Stuck on where to even start — because a shapeshifter can be anything, so it starts as nothing, every time.

Hand that setup to the bench. Bring anything and just work; when it's done, the one piece worth keeping stays and the rest is swept out. Need a new repo? Just one more thing on the bench.

🚧 Early stage. The framework runs, but APIs and docs are still settling. Expect rough edges and read the known gaps.

What you get

workbench-kit is a marketplace of two plugins that work together. Install both.

Plugin	Role	You use it to…
workbench	Engine. Worktree/task isolation, incremental knowledge harvest, and a skill-driven task lifecycle.	run day-to-day work: start a task, submit it, harvest what's worth keeping.
workbench-kit	Bootstrap. Interviews you for your conventions, then generates a minimal personalized workbench repo.	set up a new workbench once, tuned to your taste.

It's tool-neutral: the same skills/ source installs on both Claude Code and Codex. "Bring your own rules" — the mechanism is fixed, your conventions are not.

Install

Claude Code

/plugin marketplace add YOOGOMJA/workbench-kit
/plugin install workbench@workbench-kit
/plugin install workbench-kit@workbench-kit

Codex

codex plugin marketplace add https://github.com/YOOGOMJA/workbench-kit
codex plugin add workbench
codex plugin add workbench-kit

Requirements: Claude Code or Codex, git (with worktree support), and the gh CLI for issue/PR plumbing.

Quickstart

1. Create your workbench (once). The bootstrap plugin runs a short interview, then scaffolds a minimal repo personalized to your conventions:

/workbench-kit:interview-for-personalizing   # draws out your persona
/workbench-kit:generate-workbench            # scaffolds the repo, then discards the draft

What it produces is minimal by design — only your stuff. The engine (skills, utils/, bin/) stays in the installed plugin, never copied in:

my-workbench/
├── AGENTS.md          # composed: framework core + your persona (don't edit directly)
├── AGENTS.overlay.md  # your rules — edit here, then recompose
├── codebases.yaml     # your target repos
├── docs/              # knowledge wiki, starts empty, fills as you work
│   ├── decisions/  lessons/  runbooks/   # anchor tables ("did we already decide?")
│   └── index.md  log.md
├── templates/         # harvest entry + task-AGENTS formats (yours to tune)
└── .github/           # issue (4-element) + PR templates

2. Work, from inside your new workbench. The engine plugin drives the task lifecycle — each task lives on its own branch and worktree, and only the refined increment survives to main:

/workbench:task-start <issue>     # branch + worktree + task/ workspace
# … do the work …
/workbench:task-submit            # clean up task/, open a squash PR
/workbench:task-done <issue>      # sweep the workspace after merge

Other entry points: /workbench:ticket-incubate (idea → issue), /workbench:task-status · /workbench:task-tickets (read-only overviews), and the docs-query · docs-ingest · docs-lint skills for the knowledge wiki.

How it works

• Work is disposable, knowledge accumulates. Everything happens on a task branch; at the end task/ is cleaned up and the branch is squash-merged, so main keeps one refined increment. The exploration is swept out.
• Two layers. Judgment (what to extract, the prose) is the agent's; plumbing (git state transitions) is utils/'s. The agent's entry point is always a skill.
• AGENTS.core + AGENTS.overlay → AGENTS.md. The framework core is fixed and English (upgrades overwrite it). Your persona — language, slug style, issue/PR shape, labels, gates — lives in the overlay. The bootstrap composes the two.

Background and design rationale live in framework-docs/workbench-knowledge-ecosystem.md; the design decisions themselves are recorded as ADRs under framework-docs/decisions/.

Repo layout