Specs

Replace procedural CLAUDE.md instructions with Claude Code skills (custom slash commands) and hooks (automated enforcement) to make the plan workflow invocable, context-efficient, and enforceable.

Derive structured lifecycle timelines for plan entities from git history — attribute transitions, criteria progress, and resolution events — surfaced through the CLI and as a site rune.

Review and improve the @refrakt-md/plan package's conventions, folder structure, and onboarding experience. Three concerns drive this spec: the package is currently coupled to Claude Code as an AI tool, the folder naming convention uses singular nouns where plural would be more natural, and there is no path for users who already have planning documents to adopt the system.

Add a planning-only scaffold option to create-refrakt so users who want to adopt refrakt primarily as a plan management tool have a one-command entry point, matching the experience for sites and themes.

A single @refrakt-md/mcp package that exposes the refrakt CLI as a Model Context Protocol server, so AI agents can call refrakt operations through typed tools and structured I/O instead of shelling out and parsing text.

Tools and resources that close the agent's content-authoring inner loop. Today an AI agent writing refrakt content can read the rune reference, inspect a single rune in isolation, and run the build — but there's no way to render or validate an arbitrary Markdown snippet without the build, and no structured listing of the site's pages.

Sharpen the plan tools for the workflow CLAUDE.md actually prescribes — pulling context before starting work, checking off many criteria when finishing, and trusting that referenced IDs are real.