@taghost/site-engine

Shared AI website-generation engine: brief → research → architect → copywriter → critic → design → render → QA. Pure, dependency-free, environment-agnostic; runs on Cloudflare Workers. Consumed by TagHost (Studio "Build from scratch") and SMEBundle (GoLive), so both products generate un-sloppy sites from one core.

The page is authored by a frontier gateway model when one is configured: authorSite() hands the model the brief, copy, research dossier and generated image URLs, and it writes the complete HTML/CSS page itself. Every authored page passes sanitizePageHtml() (no scripts, frames, event handlers, unsafe URLs or dead forms) and the static QA gate before shipping. When no gateway is configured, or authoring fails, the deterministic composition engine renders a fallback page instead, so a build never fails on a model outage. Workers AI is never in the authoring chain by default.

The engine never touches a database or a billing system. The host owns persistence, auth, quotas and the publish path, and passes:

• env — Worker bindings (env.AI for Workers AI, or an OpenAI-compatible key).
• store — a SiteEngineStore adapter (below).

Usage

import { runDeliberation, renderSite, runStaticQa } from '@taghost/site-engine';

const { generated } = await runDeliberation(env, { project, brief, spec, store });
const { files, templateId } = renderSite({ content: generated.content, brief, domain, spec });
const qa = runStaticQa({ files, content: generated.content, brief, domain });
// host: publish `files` (R2/KV) and open the editor

generateContentModel(env, brief, { provider }) does just the copy step; chatComplete / pickModelChain / providerInfo expose the provider+router layer directly (e.g. for the brainstorm tool).

SiteEngineStore (host-implemented)

interface SiteEngineStore {
  getHouseStyle?(): Promise<string>;
  latestSteering?(projectId: string): Promise<string>;
  saveDeliberation?(turn: {
    id: string; projectId: string; buildId: string | null; round: number;
    role: string; message: string; metadata: unknown; createdAt: string;
  }): Promise<void>;
  getDeliberations?(args: { projectId: string; buildId?: string | null; limit?: number }): Promise<unknown[]>;
  getExemplars?(sector: string): Promise<unknown[]>;
  getFeedbackHints?(): Promise<string[]>;
}

All methods are optional: with no store, generation still runs (it just does not persist a transcript or read exemplars/house style).

Providers

• Workers AI (default) — env.AI, edge, no external key.
• OpenCode Go / Zen or any OpenAI-compatible gateway — set SITE_BUILDER_PROVIDER (opencode-go | zen) and the key. The router picks a fast, low-reasoning model per task and blocks free/stealth models (which may train on inputs) from customer data.

Motion layer

Every authored page must move. The model writes only declarative hooks (data-motion="kinetic|reveal|stagger|parallax|pin|horizontal|marquee|counter|magnetic|cursor|stack|progress|glass"), picks 3–5 per page with at least one signature move (kinetic|pin|horizontal), and sets an intensity level (premium or restrained, sector-adaptive).

injectMotion() appends the engine-owned layer: baseline CSS, a tiny visibility guard and four asset files (GSAP core + ScrollTrigger + SplitText + motion.js). These are the only scripts a generated bundle may contain — QA matches external script src exactly and inline script content byte-for-byte. Content stays visible without JavaScript, the guard failsafe reveals everything after 3s, and prefers-reduced-motion disables motion entirely. GSAP is bundled under its free-for-commercial-use license (https://gsap.com/standard-license; regenerate with node scripts/embed-motion.mjs).

Composition axes

content.design is validated against catalog.js; anything unknown falls back to the family preset, about + services are always present and contact is always last:

• hero: centered | split | bleed | statement | panel
• services: cards | rows | bento | numbered
• about: plain | split | panel
• contact: center | split | band
• sections: ordered subset of trust | about | services | gallery | faq

Families (golive-a…golive-e) stay as the aesthetic skin (surface, radius, header tone); a host spec still works as a per-axis fallback and for backward compatibility.

What is NOT here

Build state machine (site_builds claim/lease), auth, billing, quotas and the publish path. Those are host-specific and live in each app. Imagery generation is here (generateSubjectImages); storing the bytes and referencing their URLs is the host's job.

Consumers

• TagHost apps (control-plane site-engine-routes.ts, Studio BuildFromScratchModal) import it via the pnpm workspace ("@taghost/site-engine": "workspace:*").
• SMEBundle (separate repo) consumes the published package. The pure modules in functions/_shared/site/* are export * from '@taghost/site-engine' re-exports; host code (state machine, publish, learning/exemplar DB) stays there. While developing, SMEBundle may use a local file:../TagHost/packages/site-engine link — but a deploy guard blocks file: dependencies from shipping, so publish before deploying.

Publishing

publish-site-engine.yml publishes to public npm on a site-engine-v* tag (needs the NPM_TOKEN secret). Then SMEBundle pins "^<version>".

Status

Shipped: the portable core, TagHost "Build from scratch" (schema, routes, tier quotas, preview mapping, Studio entry), SMEBundle consumption via re-exports, the composition engine (block library + design spec, v0.4), host-wired sector imagery (Flux → deployment R2 → content.images), model-authored full pages through OpenCode Go with sanitizer + QA and a composition fallback (v0.5), and the engine-owned GSAP motion layer with declarative hooks and a vision gate (v0.6).