fix: correct tool shapes, installer hang, and the half-built non-Stitch source paths

[gabelul]

Three review passes plus an exhaustive schema comparison and an actual install test. Everything here would have shipped broken in 1.12.0.

Would have failed every call

apply_design_system documented selectedScreenIds: ["abc"]; the live API takes selectedScreenInstances: [{id, sourceScreen}]. Both new DESIGN.md skills route into it, so the release's headline feature dead-ended.

update_design_system was structurally wrong — name nested inside designSystem, projectId absent entirely. Live takes all three as siblings, all required.

Found by comparing all 15 tools against live rather than only the reported ones:

• generate_variants — prompt and variantOptions documented optional, actually required
• get_screen — name missing entirely; it's the current identifier and takes the full resource path, while projectId/screenId are deprecated but still mandatory

All 15 schemas now match live on properties and required fields, checked programmatically.

Installer died in any non-TTY context

npx @booplex/stitch-kit under CI, Docker, or piped input printed the API-key prompt then exited with Detected unsettled top-level await. readline's callback never fires on EOF, so the promise never settled — a silently failed install.

Verified fixed against a scratch HOME seeded with all seven client markers: exits 0, agent + 36 skills to Codex and OpenCode, 36 skills to Crush, agent-only to Claude Code (skills ship via plugin), MCP-only for Cursor and VS Code. That also makes the README's "Works with" table a measured fact rather than an unverified claim.

The non-Stitch routes were a door with no room

PR #21 made six conversion skills accept local HTML and URLs, but only edited Step 1:

• Token extraction assumed a Stitch-emitted tailwind.config. Arbitrary HTML has none, so the new routes couldn't resolve colors, fonts, radii or dark mode at all. Now a four-step chain — tailwind.config → CSS custom properties → linked/inline stylesheet → infer from dominant computed values and say what was inferred. Plus an explicit note that the URL route fetches only the single HTML response.
• RN and SwiftUI demanded a Stitch design in five places each, including telling users to "regenerate with deviceType: MOBILE" — meaningless to someone who supplied a local file. The constraint was always a mobile layout. Stitch screens still check deviceType; local files and URLs check viewport meta, mobile-first media queries, narrow root widths, touch targets. Mobile-only requirement unchanged.

Smaller corrections

• assetId goes bare to apply_design_system but prefixed to generate_screen_from_text and update_design_system — three skills taught the wrong form
• generate_screen_from_text was missing its designSystem param. Here the skills were right and our schema was stale — that param is what makes DESIGN.md → design system → generation hold together
• Input payloads sent font, backgroundLight, backgroundDark, description, designTokens, styleGuidelines — none exist on the live input. Output examples showing font stay; it's returned, just not settable
• DesignTheme had required: null despite requiring five fields
• ROBOTO offered as a font value; not in the enum (ROBOTO_FLEX is)
• README claimed every target ships TypeScript and ARIA — false for the HTML target and SwiftUI
• create_design_system_from_design_md.json had $defs outside the arguments fragment, so its $ref didn't resolve standalone
• Skill counts and the two new wrappers were missing from AGENTS.md and the agent definition, so the shipped agent didn't know its own new skills existed

Validator: Passed with 30 warning(s) — no errors.