Site Studio

Site Studio is a Cloudflare-native AI website builder for academics and researchers. Users describe a site or change in natural language, the agent works directly in a sandboxed Dynamic Worker, and the app updates preview and published output from R2-backed project storage.

Current Stack

• Frontend: SvelteKit 5, Tailwind CSS v4, CodeMirror 6
• App runtime: Cloudflare Workers + Hono
• Agent runtime: Cloudflare Agents SDK + @cloudflare/ai-chat
• Sandbox execution: Dynamic Worker Loader + @cloudflare/codemode
• Model access: CAIL model proxy → Cloudflare Workers AI (@cf/zai-org/glm-5.2 default)
• Storage: Cloudflare R2 + KV + Durable Objects

Repo Layout

site-studio/
├── packages/
│   ├── app/                # Cloudflare Worker app + AI chat agent
│   │   ├── src/
│   │   │   ├── agents/     # SiteBuilderAgent
│   │   │   ├── routes/     # Projects, files, preview, publish, agents
│   │   │   ├── storage/    # R2-backed project storage
│   │   │   └── lib/        # Auth, paths, templates, HTTP helpers
│   │   └── wrangler.jsonc
│   ├── frontend/           # SvelteKit dashboard/editor
│   ├── serving-core/       # Shared published/preview HTTP behavior
│   └── worker/             # Legacy-domain published-site Worker

What Works

• Project CRUD against existing R2 data
• Legacy anonymous-session recovery for returning users
• Live preview at /preview/:projectId/*
• Public published sites at /u/:handle/:slug/* (legacy /sites/:userId/:slug/* serves or 301s)
• Cloudflare Agents chat transport with persisted messages
• Dynamic Worker sandbox execution for multi-step project edits
• Clarification questions when a request is materially ambiguous

Local Development

bun install
bun run dev

This starts:

• App: http://localhost:8792
• Frontend: http://localhost:5173

All packages share one Bun workspace and lockfile. To use different local ports, run the package commands directly and pass Vite/Wrangler --port flags.

Environment

Local Worker secrets live in:

• packages/app/.dev.vars

Local identity verification material is ops-managed (see cail-gateway docs/INTEGRATION.md):

CAIL_IDENTITY_JWKS={"keys":[...]}

Site Studio accepts identity only in X-CAIL-Identity-JWT and verifies it as RS256 against CAIL_IDENTITY_JWKS, with audience cail:site-studio and the canonical/staging issuer allowlist. A presented token rejects when the JWKS is missing or malformed or verification fails. CAIL_REQUIRE_IDENTITY=true rejects requests that do not carry a verified identity.

Identity enforcement also requires a bounded legacy-account import window:

CAIL_SSO_SWITCHED_AT=2026-07-13T14:00:00Z
CAIL_ACCOUNT_IMPORT_UNTIL=2026-07-27T14:00:00Z

Both values must be ISO 8601 instants with an explicit UTC offset. The import window is half-open (CAIL_SSO_SWITCHED_AT <= now < CAIL_ACCOUNT_IMPORT_UNTIL), the end must not precede the start, and the duration cannot exceed 30 days. When CAIL_REQUIRE_IDENTITY=true, missing or invalid values fail protected requests with 500 invalid_account_import_configuration. While identity is optional, bad or absent window configuration disables legacy import.

Site Studio holds no provider API keys — model calls go through the CAIL model proxy, which attaches credentials itself.

The Worker also reads these vars from packages/app/wrangler.jsonc:

• APP_PUBLIC_DOMAIN
• PUBLISHED_BASE_URL
• CAIL_API_BASE
• CAIL_MODEL (Workers AI @cf/... id only — CAIL policy is Cloudflare models only)
• CAIL_REQUIRE_IDENTITY
• CAIL_SSO_SWITCHED_AT
• CAIL_ACCOUNT_IMPORT_UNTIL

For production, configure secrets with Wrangler / Cloudflare, not by committing env files.

The app stores CSRF tokens under the csrf/ prefix in the existing private R2 bucket. Match their 30-day session lifetime with a bucket lifecycle rule:

bunx wrangler r2 bucket lifecycle add site-studio delete-expired-csrf csrf/ --expire-days 30 --force

Main Routes

• GET /api/health
• GET /api/projects
• POST /api/projects
• GET /api/projects/:id/files
• POST /api/projects/:id/file
• POST /api/projects/:id/upload
• GET /preview/:id/*
• POST /api/projects/:id/publish (409 handle_required until the user claims a handle)
• POST /api/projects/:id/unpublish
• GET /api/handle, GET /api/handle/check?handle=…, POST /api/handle
• ALL /api/agents/site-builder/:projectId
• GET /u/:handle/:slug/* (canonical published sites)
• GET /sites/:userId/:slug/* (legacy: 301s to /u/… when the owner has a handle, else serves)

Notes

• The new app is a static-file site builder. Runtime build tools are out of scope.
• The normal chat path is execute-first, not approve-first.
• The built-in gallery includes blank, CV, course, portfolio, publication, event, photo, resource, timeline, and data-visualization templates.
• Canonical published URLs are /u/:handle/:slug/, keyed by a user-chosen handle so the owner id never appears in a public URL. Old published sites remain readable permanently from the same R2 bucket via the legacy /sites/:userId/:slug/* shape, which 301s to the /u/… equivalent once the owner has a handle and otherwise serves content directly. This is a permanent compatibility exception: temporary account-import cleanup must retain /sites routing, direct serving, redirects, and .migrated.json forwarding-pointer resolution.

See docs/legacy-account-import-removal.md for the temporary import telemetry and deletion follow-up.

License

MIT