v3.8.0-alpha.8 — Design Intelligence v2 — FEAT-009 ships

Closes the long-blocked FEAT-009 tail and adds the missing vision + design-system-persistence layers.
Driven by the 2026-07-16 deep-research pass (Anthropic frontend-design skill, Agent SDK vision loop,
Google Stitch MCP, superdesign). Full design: docs/architecture/LLD-DESIGN-INTELLIGENCE.md.

Added

• skills/design-vision-loop — the missing "look at it" step. Renders the running UI via the Playwright
  MCP, screenshots it across 3 viewports + dark mode, runs a deterministic tier-1 gate (the new
  design-conformance hook + console errors) before spending a vision call, then critiques the shots
  against the design system and iterates (max 3). Two-tier by design — Anthropic ranks rules-based feedback
  above LLM-as-judge. Ships with references/critique-rubric.md + references/viewport-matrix.md.
• hooks/design-conformance.cjs (PostToolUse Write|Edit) — deterministic design gate that finally
  enforces what theme-consistency / design-system-usage / motion-design only described in prose:
  hardcoded hex/rgb, raw px/numeric spacing, one file importing 2+ component libraries, and
  animation/transition with no prefers-reduced-motion guard. Warnings only, fail-open, skips token/theme
  files. 27 unit tests; RED caught two real regex bugs before merge.
• rules/agent/design-system-persistence.md — .claude/design/design-system.md as the durable design
  source of truth (schema + producers/consumers), so tokens/type/library survive across sessions instead of
  being re-picked each time (pattern from Stitch design.md / superdesign .superdesign/design-system.md).
• Google Stitch MCP — opt-in remote server in .mcp.json (stitch, http transport, STITCH_API_KEY),
  disabled by default like the other opt-ins.

Changed

• skills/stitch-design rewritten — dropped the stale "Stitch has no API" claim; now prefers the MCP
  (generate/edit screens + design-system tools, Flash-for-drafts quota discipline) and keeps the manual
  copy-paste flow as a graceful fallback. Seeds .claude/design/design-system.md + saves screen PNGs as
  vision-loop targets.
• skills/frontend-aesthetics → v2 — added Anthropic's two-pass process (compact design plan →
  self-critique against the brief before coding), named the specific AI-default clusters to avoid
  (cream+serif+terracotta, dark+acid-green, hairline-broadsheet), and a screenshot self-critique directive.
• skills/design-tokens, skills/design-expert, agents/frontend — now write/read the design SoT file.
• Figma tool-mismatch fixed — prompts referenced Dev-Mode-MCP tools (get_variable_defs/
  get_code_connect_map) that the installed figma-developer-mcp doesn't expose; corrected to
  get_figma_data / download_figma_images.
• Counts synced — skills 59→60, rules 71→72, hooks 48→49, MCP 10→11 across CLAUDE.md / plugin.json /
  README / stats.json.

Note

Stitch endpoint + tool names came from research that couldn't be fully adversarially verified (session
limit); the skill instructs verifying via list tools on first connect and degrades to the manual path if
the server is unreachable. Tracked as STORY-0033 open risk in docs/ROADMAP.md.

---