schema: spec-driven

context: |
  Project: ASW — Agentic Semantic Web
  Purpose: A CSS framework built for agents, not humans. Semantic HTML + data-* attributes.

  Core idea: LLMs produce semantic HTML naturally. ASW makes that HTML look correct
  without the agent touching CSS. No classes, no inline styles, no presentation vocabulary.

  Stack: Pure CSS framework. PostCSS build (postcss-import, postcss-custom-media, cssnano).
  Foundation: Open Props provides primitive token scales; ASW aliases them to semantic names.

  Engine-agnostic: ASW ships as CSS + packs. Packs integrate with Hugo, Apache, Caddy,
  Flask, Nginx, Pandoc, Python. The framework's own website currently uses the Hugo pack.

  Repo structure (target):
    src/           — framework CSS source (layers/) and lab experiments (lab/)
    dist/          — built output: asw.css, asw.min.css
    vendor/        — explicit dependencies (Open Props with version tracking)
    packs/         — engine integration bundles
    docs/          — framework documentation (markdown only, engine-agnostic)
    examples/      — static HTML showcases of framework capabilities
    site/          — the ASW website (currently Hugo, consumes docs/ as content)
    openspec/      — specs and changes

  Key constraint: layers 02+ must only reference semantic aliases from 01-tokens.css,
  never Open Props primitives directly.

  Lineage: Pico CSS (component patterns, absorbed), Open Props (token scales, foundation),
  Charts.css (data-* attribute pattern, inspiration).

  Forgejo: git.trentuna.com/trentuna/asw
  Legacy repo: git.trentuna.com/trentuna/agentic-semantic-web (to be archived after import)

rules:
  proposal:
    - Keep scope tight — this is a CSS framework, not a JS library
    - Name the files that will be touched
    - Consider impact on existing deployed pages (trentuna.com)
  specs:
    - Use Given/When/Then for behavior scenarios
    - Specs should be testable by visual inspection in a browser
    - Map 1:1 to CSS layer files where applicable
  tasks:
    - Each task should be a standalone unit of work for a single agent session
    - Include acceptance criteria
    - No task should touch more than 3 files
    - Build check after every task