refactor(stats): serve opencode.ai/data straight from the stats worker (#62192)

infra/stats.ts

@@ -1,5 +1,6 @@
 import { EMAILOCTOPUS_API_KEY } from "./app"
-import { domain } from "./stage"
+import { LOCALES } from "../packages/stats/app/src/lib/language"
+import { domain, zoneID } from "./stage"
 
 ////////////////
 // DATABASE
@@ -62,16 +63,42 @@ new sst.x.DevCommand("StatsStudio", {
 // APP
 ////////////////
 
-export const app = new sst.cloudflare.x.SolidStart("Stats", {
+export const publicUrl = `https://${domain}/data`
+
+const app = new sst.cloudflare.x.SolidStart("Stats", {
   path: "packages/stats/app",
   buildCommand: "bun run build",
-  domain: `stats.${domain}`,
   link: [database, EMAILOCTOPUS_API_KEY],
   environment: {
-    PUBLIC_URL: `https://${domain}/data`,
+    PUBLIC_URL: publicUrl,
+  },
+  transform: {
+    server: {
+      url: false,
+    },
   },
 })
 
+// The site serves itself under /data and /<locale>/data, and these routes are its only way in: it has no custom domain
+// or workers.dev URL. Route patterns can't wildcard the middle of a path, hence one pair per locale. `sst dev` deploys
+// no Worker.
+const appServer = app.nodes.server
+if (appServer) {
+  for (const prefix of ["data", ...LOCALES.map((locale) => `${locale}/data`)]) {
+    const name = prefix.replaceAll("/", "")
+    for (const [suffix, pattern] of [
+      ["", `${domain}/${prefix}`],
+      ["Path", `${domain}/${prefix}/*`],
+    ]) {
+      new cloudflare.WorkersRoute(`StatsRoute${name}${suffix}`, {
+        zoneId: zoneID,
+        pattern,
+        script: appServer.nodes.worker.scriptName,
+      })
+    }
+  }
+}
+
 ////////////////
 // SERVICES
 ////////////////

packages/stats/README.md

@@ -40,6 +40,6 @@ To run one sync pass on your machine against a stage's resources, run `bun sst s
 
 ## Public page caching
 
-Stats HTML, JSON, and Markdown use a one-minute browser cache and a five-minute shared cache. Expired responses must revalidate instead of serving stale snapshots. The console `/data` proxy inherits this policy from the stats app.
+Stats HTML, JSON, and Markdown use a one-minute browser cache and a five-minute shared cache. Expired responses must revalidate instead of serving stale snapshots.
 
-Cloudflare must respect these origin headers. The live `Stats pages: respect short origin cache TTL` cache rule applies to `/data` and `/data/*` on `opencode.ai` and `stats.opencode.ai`: browser TTL respects origin, edge TTL uses cache-control headers and bypasses cache when absent, and serving stale content while updating is disabled. This dashboard-managed rule overrides the zone's four-hour browser-cache minimum; preserve it when changing Cloudflare configuration. Purge both stats URL prefixes after changing the policy to clear existing cached responses.
+The Worker has no custom domain. `infra/stats.ts` routes `/data`, `/data/*`, `/<locale>/data` and `/<locale>/data/*` on the main host straight to it, so its responses skip the Cloudflare cache and reach browsers with these headers unchanged. Static assets are uploaded under `/data`, and the build moves Nitro's `_headers` to the assets root so they stay immutable.

packages/stats/app/src/context/language.tsx

@@ -49,6 +49,8 @@ export const { use: useLanguage, provider: LanguageProvider } = createSimpleCont
       document.documentElement.lang = tag(store.locale)
       document.documentElement.dir = dir(store.locale)
       document.documentElement.dataset.locale = store.locale
+      // Remember a language picked by URL, so unprefixed /data links redirect back to it.
+      if (fromPathname(window.location.pathname)) document.cookie = cookie(store.locale)
     })
 
     return {

packages/stats/app/src/middleware.ts

@@ -9,7 +9,18 @@ import {
   modelJson,
   modelMarkdown,
 } from "./lib/agent-formats"
-import { fromPathname, LOCALE_HEADER, pageUrl, parseLocale, prefersMarkdown, route, strip } from "./lib/language"
+import {
+  basePath,
+  cookie,
+  fromPathname,
+  LOCALE_HEADER,
+  localeFromRequest,
+  pageUrl,
+  parseLocale,
+  prefersMarkdown,
+  route,
+  strip,
+} from "./lib/language"
 import { loadComparePage, loadHomePage, loadLabPage, loadModelPage } from "./lib/page-data"
 import { canonicalModelEntry, catalogModelPath, findModelCatalogLab, loadModelCatalog } from "./routes/model-catalog"
 import { statsPageCacheControl } from "./routes/stats-cache"
@@ -41,6 +52,8 @@ export default createMiddleware({
     const request = event.request
     if (request.method !== "GET" && request.method !== "HEAD") return
     const url = new URL(request.url)
+    const localized = localizedRedirect(request, url)
+    if (localized) return localized
     const path = strip(url.pathname)
     const parsed = parsePath(path)
     if (!parsed) return
@@ -69,6 +82,43 @@ export default createMiddleware({
   },
 })
 
+// Unprefixed pages redirect to the visitor's language from their cookie or Accept-Language. /en/data stays put so
+// English remains reachable for visitors whose browser prefers another language.
+function localizedRedirect(request: Request, url: URL) {
+  if (fromPathname(url.pathname)) return undefined
+  if (url.pathname !== basePath && !url.pathname.startsWith(`${basePath}/`)) return undefined
+  if (isDataBypassPath(url.pathname)) return undefined
+  const accept = request.headers.get("accept")
+  // Agents asking for Markdown get the English page in place.
+  if (prefersMarkdown(accept)) return undefined
+  if (accept && !accept.includes("text/html") && !accept.includes("*/*")) return undefined
+  const locale = localeFromRequest(request)
+  if (locale === "en") return undefined
+  const headers = new Headers({
+    // route() normalizes /data to /data/, so keep the visitor's own trailing slash.
+    Location: `${url.pathname === basePath ? `/${locale}${basePath}` : route(locale, url.pathname)}${url.search}`,
+    "Cache-Control": "private, no-store",
+    Vary: "Accept, Accept-Language, Cookie",
+  })
+  headers.append("Set-Cookie", cookie(locale))
+  return new Response(null, { status: 308, headers })
+}
+
+function isDataBypassPath(pathname: string) {
+  return (
+    pathname.startsWith(`${basePath}/_build/`) ||
+    pathname.startsWith(`${basePath}/api/`) ||
+    pathname.startsWith(`${basePath}/stats/`) ||
+    pathname.startsWith(`${basePath}/_server`) ||
+    pathname === `${basePath}/banner.jpg` ||
+    pathname === `${basePath}/banner.png` ||
+    pathname === `${basePath}/sitemap.xml` ||
+    pathname.endsWith("/llms.txt") ||
+    pathname.endsWith(".md") ||
+    pathname.endsWith(".json")
+  )
+}
+
 function parsePath(path: string) {
   const segments = path.split("/").filter(Boolean).slice(1)
   if (segments.some((segment) => !/^[\w.-]+$/.test(segment))) return undefined

packages/stats/app/vite.config.ts

@@ -1,3 +1,5 @@
+import { readFile, rm, writeFile } from "node:fs/promises"
+import path from "node:path"
 import { solidStart } from "@solidjs/start/config"
 import { nitro } from "nitro/vite"
 import { defineConfig, type PluginOption } from "vite"
@@ -7,12 +9,30 @@ export default defineConfig({
   plugins: [
     solidStart({ middleware: "./src/middleware.ts" }) as PluginOption,
     nitro({
+      // The Worker owns opencode.ai/data/* directly, so static assets are uploaded under /data like the pages that
+      // reference them. Nitro's baseURL would do this too, but its router would then stop serving /<locale>/data.
+      output: { publicDir: "{{ output.dir }}/public/data" },
       compatibilityDate: "2024-09-19",
       preset: "cloudflare-module",
       cloudflare: {
         nodeCompat: true,
       },
     }),
+    {
+      // Cloudflare only reads _headers at the assets root, but Nitro writes it into publicDir with unprefixed paths.
+      name: "stats:headers",
+      buildApp: {
+        order: "post",
+        async handler() {
+          const source = path.join(import.meta.dirname, ".output/public/data/_headers")
+          await writeFile(
+            path.join(import.meta.dirname, ".output/public/_headers"),
+            (await readFile(source, "utf8")).replace(/^\//gm, "/data/"),
+          )
+          await rm(source)
+        },
+      },
+    },
   ],
   server: {
     allowedHosts: true,

sst.config.ts

@@ -43,7 +43,7 @@ export default $config({
     await import("./infra/share.js")
 
     return {
-      ...(stats ? { StatsUrl: stats.app.url } : {}),
+      ...(stats ? { StatsUrl: stats.publicUrl } : {}),
       AwsStage: stage.awsStage,
     }
   },