What changed

• New package apps/web-start (@phaseo/web-start). It is the whole website rebuilt on TanStack Start 1.169 and running on Cloudflare Workers. apps/web is untouched, and nothing in this PR changes production.
• Every route is ported: 347 route files, covering all pages, layouts, API route handlers, OG images, robots.txt and sitemap.xml.
  • The preview runs at https://phaseo-web-start-preview.danielbutler500.workers.dev (workers.dev only, no custom domain).
• New CI matrix entry web-start, which runs typecheck and build for the package.

Commit 9d16fb8b9d copies apps/web/src verbatim. To review only the port, use git diff 9d16fb8b9d HEAD -- apps/web-start/src (793 files, +13.8K/−3.3K), plus the files outside src.

Why

• Cost. We want to know whether the site can leave Vercel for Workers at a lower cost without losing speed, and which framework makes that cheapest.

• Earlier measurements on 2026-10-30 (same shared components, same live data, mean CPU per render):

  Framework	1× page	4× page
  vinext	37 ms	128 ms
  OpenNext	22 ms	46 ms
  TanStack Start	9 ms	22 ms

• The alternatives didn't hold up on the full app.

  • vinext still needed 220–280 ms CPU per page after optimising.
  • OpenNext could not deploy at all: 113.7 MiB against the 64 MiB Worker limit.

• This PR answers the question with the real app: the measured cost and speed are in Before and after below.

How

• Next.js compatibility layer (src/compat). Vite aliases point next/link, next/navigation, next/image, next/dynamic, next/headers, next/server, next/og, next-intl, next-intl/server, flags/next and others at small shims. The copied components therefore mostly run unchanged.
• Data loading. The 252 async server components became plain SSR:
  • Data comes through createServerFn in *.functions.ts, read with useSuspenseQuery.
  • setupRouterSsrQueryIntegration streams the results to the browser.
  • Route loaders prefetch, and handle redirects and not-founds before render.
• Generated routes (scripts/codemods/generate-routes.mjs) wrap each Next page or layout and run its generateMetadata in a server function.
  • Catch-all segments become $.tsx files. A $/index.tsx splat matches an empty tail and shadowed the model page.
  • Redirect-only pages redirect in beforeLoad.
• Request middleware (src/lib/start/requestMiddleware.ts, session.ts) replaces next.config.mjs redirects/headers and src/proxy.ts. It covers:
  • locale negotiation;
  • the Supabase session refresh;
  • proxies for /api/_web|account|chat|internal, /docs and /ingest;
  • markdown negotiation and the blog preview gate.
• Server actions. "use server" modules become POST stubs to /api/server-actions. The endpoint checks the Origin and only calls registered exports.
• Build plugins (build/phaseo-plugins.ts):
  • blog MDX is compiled at build time;
  • src/content is served from static assets through an fs shim;
  • the Geist font for OG images is inlined.
• Worker memory. A Worker isolate holds every uploaded module's source within its 128 MB, and any file with one non-Latin1 character costs 2 bytes per character. The first deploy needed about 81 MB for source alone and took about 2 s to first byte. The fixes:
  • translations load from static assets;
  • Shiki and the docs search index stay out of the server bundle;
  • server chunks escape non-Latin1 text where that is smaller.
• Trade-off: we return real 404s for unknown models and providers. Next streamed those as 200 with a not-found page.

Before and after

Method: 300 requests replaying the production page and locale mix at 1 request/s (seed 9). CPU comes from wrangler tail, and first byte is measured client-side from London. vinext was measured the same way on 2026-10-30, and the Vercel figure comes from that same session.

Median CPU by page type: models 90 ms, rankings 153 ms, tools 285 ms, home 58 ms, api-providers 47 ms, benchmarks 31 ms, compare 28 ms, chat 17 ms.

Verification

• pnpm --filter @phaseo/web-start typecheck: clean.
• vite build: passes with and without the local env file, as CI will run it.
• Parity against https://phaseo.app (scripts/parity-check.mjs, Playwright: status, title, visible text, console and page errors, error screen):
  • 50 of 58 sitemap route shapes, plus 11 of 12 extra locale checks (zh-Hans, ar-SA, ja, hi, fr-FR, de-DE, es-ES, pt-BR), match.
  • The remaining differences are all explained:
    • /pricing (and its locale versions) is missing the Enterprise section. It is behind a Statsig gate, and the Statsig server key is not set on the preview Worker.
    • Three methodology pages log missing-message errors. Prod has the same bug: phaseo.app shows raw translation keys as section headings.
    • /chat/image has a " | Phaseo" title suffix that prod omits because of a quirk in Next's title templates.
    • Unknown models and providers return a real 404 instead of 200.
• Spot checks with curl: the redirects match prod, including retired model sub-paths, /docs, /en-GB/*, the social links, /chat/audio and /gateway/usage?tab=. sitemap.xml lists the same 2,745 URLs as prod.
• Not verified:
  • Signed-in flows (settings, billing, server actions with a session) on the preview, since its workers.dev address is not an allowed auth origin.
  • Admin and internal pages.
  • Real browsers other than Playwright Chromium.
  • The web Jest suite, which is not configured for apps/web-start.
• Review: sol-loop, 2 rounds with GPT-6.1 Sol. Round 1 found one blocker: unscoped account query keys would have made the session listener clear the cache and reload the page. It is fixed in 39ae0203c8, and round 2 approved.
• No changeset: @phaseo/web-start is private and not published.

Reviewed with sol-loop: 2 rounds with gpt-6.1-sol on high. Sol verdict: approved

Model: Claude Opus 5.5 (effort unknown) · Harness: Claude Code in T3 Code

🤖 Generated with Claude Code