A spec-driven runtime for real AI coding work.
CafeKit installs skills, agents, hooks, state checks, and quality gates into your repo so an assistant can move from idea to verified implementation without drifting into ad-hoc prompting.
What CafeKit is#
CafeKit is a repo-local runtime bundle for AI coding assistants. The current package is @haposoft/cafekit, installed with npx @haposoft/cafekit.
It adds a structured workflow around Claude Code and OpenCode:
The docs should always point back to this contract loop.
Turn intent into scope, evidence, requirements, design, and task packets.
Implement one approved task packet, not the whole chat memory.
Run exact evidence commands and runtime checks.
Review for spec compliance, regressions, security, and drift.
Update task_registry and task markdown only after proof, or audit drift.
Start with SDD#
CafeKit is built around specification-driven development: an idea becomes code only after it becomes a source-backed contract. Start with the visual guide: Spec-driven development.
After that, the deeper path is Core workflow → Spec lifecycle → Skills.
What it installs#
CafeKit installs a working runtime, not just documentation snippets.
29 packaged workflows and domain capabilities
13 specialist agents for planning, code, QA, docs, git, and deploy
Privacy, scope, state, usage, docs sync, and spec-state gates
Runtime configuration, paths, locale, usage, and model hints
Claude Code hook wiring and statusline command
Git, context, usage, task, and agent statusline
The package currently ships 30 skills, 13 agents, installer support for Claude Code and OpenCode, and deterministic validators for specs and documentation bundles.
How to read these docs#
If you are new, read in this order:
- Installation
- Spec-driven development
- Quickstart
- Core workflow
- Runtime bundle
- Spec lifecycle
- Workflows
If you already know the workflow, jump to Skills, Agents, or Reference.
Core promise#
CafeKit does not try to make the assistant faster by removing process. It makes the assistant safer by forcing source-backed context, explicit task boundaries, fresh verification, adversarial review, and synchronized state.
Task files and scope_lock define what may change.
Exact commands and runtime proof must be fresh.
Critical findings block task completion.
spec.json and markdown state move together.