dwarves-kit adds a spec-driven workflow to Claude Code: you set the goal and the gates, and agents build and fix inside them. A read-only verifier decides when work is done, so no agent grades its own homework. Safety is mechanical: bash hooks block a force push, a secret read, or a ship with no proof, and a fix loop stops after two tries and asks you.
Figure: one change moves through eight phases; two gates block, the rest advise, and each retro feeds the next think.
/plugin marketplace add dwarvesf/dwarves-kit
/plugin install kit@dwarves-marketplace
Open your repo and run /kit:onboard. It adopts the repo and picks modules, and it previews every write; a decline changes nothing.
Then run /kit:start at the top of each session. It reads where the repo stands and names the next command.
No Claude Code plugin system, or you use Codex? See Install options and modules for the bash installer, the Codex adapter and the requirements.
| You want to | Read |
|---|---|
| See the whole loop once, start to finish | docs/QUICKSTART.md |
| Try the kit on a repo, and know what it writes | docs/guides/adopt.md |
| Ship one change and see what happens when | docs/guides/spine.md |
| Get a change reviewed before it ships | docs/guides/review.md |
| Run several pieces of work at once | docs/guides/mega-goal.md |
| Prove an artifact works for a cold reader (docs, a runbook, a spec, an API) | docs/guides/gauntlet.md |
| Land a finished session and clean up | docs/guides/wrap.md |
| Fix a push that was blocked | docs/guides/gates-and-proof.md |
| Hunt a bug | docs/guides/debug.md |
| Contribute to the kit | CONTRIBUTING.md |
Every guide: docs/guides/README.md.
You rarely type a command. Say what you want, such as "add a --version flag to the CLI", and the kit sizes the work, writes the spec, builds, verifies, and stops only where it needs your call.
- Commands. The
/kit:*actions for each phase. Every command is documented in the command reference. - Agents. Read-only verifiers and reviewers, plus a fix-agent and domain builders. The agents table says who dispatches each.
- Hooks. Bash scripts on Claude Code events. A small spine always runs; the rest is opt-in. The full inventory is
docs/reference/hooks.md. - Skills. Claude loads these when your intent matches, for example the audit loops. See the skills roster.
- The bash engine. Each subsystem is a standalone shell command under
bin/, for examplebash bin/board --helpandbash bin/precedent find "<query>". It runs without Claude Code.
A lead who hands implementation to contractors on one shared spec, and anyone who spends long days in Claude Code and wants guardrails and a record of what ran.
It is bash and markdown, with no UI and no server. It targets one engineer or one lead with helpers, not teams that need cross-machine orchestration; docs/reference/PHILOSOPHY.md says where that line sits.
| Doc | What it holds |
|---|---|
docs/MANUAL.md |
The command reference: every command, agent and recovery path |
docs/reference/lifecycle.md |
The loop, the gates and the five stages |
docs/reference/install.md |
Install layers, the module table, plugin, Codex and bash install |
docs/reference/skills.md |
The skills roster |
docs/reference/project-structure.md |
The directory layout |
docs/reference/operations.md |
Testing, CI, the weekly scheduler and external tools |
docs/reference/hooks.md |
Which hook prints which message, and the knobs that tune it |
docs/reference/architecture.md |
How the parts fit and how data moves |
docs/SYSTEM-DESIGN.md |
The design model: use cases, components, flows, state machines |
docs/reference/FEATURES.md |
Generated one-row-per-feature registry |
docs/reference/PHILOSOPHY.md |
Why the kit is shaped this way, and what it rejects |
CHANGELOG.md |
Release history |
CONTRIBUTING.md, SECURITY.md, LICENSE |
Contribute, report a vulnerability, MIT terms |