ArchLang

19. Both public sites on Cloudflare Workers static assets, deployed from CI ​

  • Status: Accepted
  • Date: 2026-09
  • Scope: site hosting only. No core, language, or public-API change; no public URL change.

Context ​

From 2026-07-15 the two public sites were Vercel deployments reached through the archlang.uk zone on Cloudflare DNS — three grey-cloud ("DNS only") CNAMEs to *.vercel-dns-016.com, with Vercel owning TLS and the www → apex redirect. That split the operational surface across two vendors for no benefit: the zone, the second domain (archcanvas.uk), and the account that manages both were already Cloudflare's.

Nothing about the Vercel deployment was broken. The move is a consolidation, and the constraint that shaped every decision below is that the public URLs must not change — https://archlang.uk and https://playground.archlang.uk are written into JSON Schema $ids, the VitePress sitemap.hostname, README permalinks, SKILL.md, and the MCP shim's baked context. A host move is cheap; a URL move is a repo-wide regeneration with its own documented traps (docs/hosting-and-domains.md).

Both sites are wholly static: VitePress emits HTML, the playground is a Vite bundle whose entire state lives in a #z= hash that never reaches the server. Neither has ever needed a server-side request handler, and neither should acquire one by accident.

Decision ​

Serve both sites as Cloudflare Workers with static assets, with no main script, and deploy them from GitHub Actions.

  • docs-site/wrangler.jsonc → Worker archlang-docs, assets from .vitepress/dist, custom domain archlang.uk.
  • playground/wrangler.jsonc → Worker archlang-playground, assets from dist, custom domain playground.archlang.uk.
  • workers_dev: false on both.
  • .github/workflows/deploy.yml ("Deploy (Cloudflare)") keeps its push-to-main trigger and its two-site matrix: build the core, build the site, wrangler deploy, then the unchanged scripts/smoke.mjs check against the live origin.

Why Workers static assets, and not Cloudflare Pages ​

Pages is the product whose name suggests it. It was rejected on three counts.

  • Workers is where Cloudflare is investing; Pages is in maintenance for new projects, and its Functions model is a second, divergent way to write the same code we do not write anyway.
  • wrangler.jsonc is repository state. A Pages project's build command, output directory, and redirect behaviour live in the dashboard. This repository's standing preference is that configuration which affects a deploy is committed and reviewable — the same reason deploy.yml pins its wrangler version.
  • The static-asset path is identical either way, so Pages offered nothing to weigh against those two.

Why there is no main script ​

With no script to invoke, every request is a static-asset request: free, unmetered, and exempt from the Workers free plan's 100,000-requests/day cap. Adding a main — even a trivial pass-through — opts both public sites into that cap in exchange for nothing. This is a standing constraint, not an implementation detail: a future feature that wants a request handler must justify the cap it buys.

Why deploys stay in GitHub Actions, not Cloudflare's Git integration ​

Two reasons, both about what the deploy is actually made of.

  • The build is a monorepo build. npm run docs:build and npm run playground:build build the core first and then the site — docs-site/sync-docs.mjs hard-exits when dist/ is missing, deliberately, so a site can never deploy with a broken example gallery. Cloudflare's Git integration would need that expressed as a dashboard build command, moving a load-bearing ordering constraint out of the repository.
  • The smoke check belongs welded to the deploy that produced it. scripts/smoke.mjs runs as the step immediately after wrangler deploy, against the live origin, in the same job. A dashboard-driven deploy has no step to weld it to.

A third, found rather than predicted: docs-site/.vitepress/config.ts sets lastUpdated: true, which reads each page's git commit time at build time. The workflow's actions/checkout therefore needs fetch-depth: 0. A depth-1 clone produces no error — every "Last updated" stamp silently collapses to the deploy commit. That is a CI-visible knob we control; on a hosted build integration it would be one more dashboard setting able to corrupt content quietly.

Why the www redirect is zone state, not an in-repo _redirects ​

www.archlang.uk → 301 → archlang.uk is a Cloudflare Single Redirect Rule in the http_request_dynamic_redirect phase, configured on the zone.

It is not in a _redirects file because Cloudflare's _redirects matches on path only and cannot express a scheme+host source. There is no way to say "requests whose Host is www.archlang.uk" in that format; a path-only rule attached to the docs Worker would fire for apex requests too and redirect the site to itself.

The rule's target is an expression, concat("https://archlang.uk", http.request.uri.path), with preserve_query_string: true — not a static value, which would flatten every www deep link onto the home page. The exact API shape is recorded in docs/hosting-and-domains.md so it can be recreated without guessing, since this is the one piece of the setup that lives outside the repository.

Why the two sites' asset routing differs ​

They differ on both knobs, and neither difference is an inconsistency to tidy up. One site has a page per route; the other has two pages and a fragment.

  • Docs: html_handling: "drop-trailing-slash", not_found_handling: "404-page". The first is the equivalent of Vercel's cleanUrls: true + trailingSlash: false — VitePress emits guide.html and links to /guide, so /guide must serve that file and /guide/ must redirect to the extensionless form. The second is required by this repository's own checks: VitePress emits its own 404.html, and both docs-site/e2e/routes.spec.ts and scripts/smoke.mjs treat a text/html 404 as a negative signal — an unmatched path must 404, not answer 200 with the shell.

  • Playground: html_handling: "none", not_found_handling: "none", plus one _redirects line. "none" is an exact-path lookup, chosen for one measured reason: it keeps /embed.html a literal 200. Every other mode drops the .html and 307s to /embed, and /embed.html is a published URL pasted into third-party <iframe> snippets — a redirect would cost every embed on the internet a round trip and gain nothing. Vercel served it literally; this keeps that.

    The cost of "none" is that / is not a file and would fall through to not_found_handling. playground/public/_redirects answers it with / /index.html 200 — a status-200 rule is a rewrite, not a redirect, and the redirect layer runs before the asset lookup, so the root serves index.html with no redirect and no 404. (This is the one thing Cloudflare's _redirects can express that we need: a path-only rewrite. It still cannot express the host-matched www redirect above, which is why that one is zone state.)

    not_found_handling: "none" is explicit because the absence is the decision. There is no 404.html here, and this is not a single-page application — index.html and embed.html are two real entries and all state is in the #z= hash, which never reaches the server. "single-page-application" would answer every mistyped path with a 200 and the editor shell, turning a 404 into a false success.

Consequences ​

  • The proxy iron law inverts. The records are now proxied (orange cloud), because a Workers custom domain necessarily is one. The old rule — "never proxy, it breaks Vercel's TLS" — was correct for its origin and is now superseded. It is recorded as an inversion in docs/hosting-and-domains.md rather than deleted, because a reader who meets the old advice in an archived doc needs to know why it existed.
  • A proxied zone can silently change what a machine reads, where the grey cloud never touched a byte. Four zone features must therefore stay off: Bot Fight Mode and Browser Integrity Check (which would challenge scripts/smoke.mjs, a bare zero-dep fetch), Email Obfuscation (which rewrites HTML and would break the nightly raw-.md byte-equality staleness probe forever), and Rocket Loader (which defers and rewrites scripts, under which the client-compiling @prod Playwright cases would flake). This is a new standing hazard the Vercel arrangement did not have.
  • archlang-docs.vercel.app and archlang-playground.vercel.app are gone, permanently. They used to 301 to the custom domains, and this repository's docs promised they were kept "so no external link ever breaks". That promise is withdrawn: those hostnames belong to Vercel and die with the projects. One known casualty is immutable — the MCP registry's first-published entry for io.github.ChanMeng666/archlang-mcp (0.1.1, 2026-07-10) carries archlang-docs.vercel.app as its website URL. It was already stale before this move (every later version points at archlang.uk), and registry history cannot be edited. Accepted: no archlang.uk link breaks, and the loss is confined to a hostname that had been redirect-only since 2026-07-15.
  • Three headers Vercel sent implicitly, Cloudflare sends not at all — and one of them is invisible to every check in this repository. Access-Control-Allow-Origin: *, Content-Disposition: inline and Strict-Transport-Security: max-age=63072000 are now declared explicitly under /* on both sites. The CORS one is the finding: the JSON Schemas are fetched by browser-context validators at their canonical $id URLs and llms.txt advertises the whole root as machine-fetchable, but scripts/smoke.mjs is a Node fetch and Playwright's request fixture are both exempt from CORS, so losing the header would have read as a healthy 200 to every gate while being broken for every browser consumer. It is documented as a standing gap in docs/hosting-and-domains.md rather than closed, because closing it means asserting a response header from a real browser origin — a test tier this repository does not have. (includeSubDomains is deliberately absent from the HSTS value, matching what Vercel sent: adding it would newly bind every subdomain of archlang.uk.)
  • The rest of _headers is a measured correction too, not a preference. Cloudflare types .md as text/markdown with no charset (Vercel sent one — without it a text/* response is latin-1 in some clients, mojibaking every em-dash in the raw markdown copies), and sends no Content-Type at all for .gbnf (Vercel sent application/octet-stream; both deliver a machine-readable route as an opaque download). Both are corrected. Fingerprinted /assets/* also regained max-age=31536000, immutable, which Cloudflare does not assume; everything else keeps the revalidating default so a deploy is visible immediately.
  • No certificate work was needed. Universal SSL on the active zone already covered the apex and every first-level subdomain, so archlang.uk, www and playground were inside the existing cert the moment the records flipped.
  • The whole cutover was verifiable before it was visible. Both Workers were deployed to *.workers.dev staging origins and driven through the repository's own production checks with the live sites untouched: 42/42 docs smoke checks, 3/3 playground, 46/46 docs @prod Playwright cases (including the byte-equality staleness probes) and 9/9 playground. Staging first, then DNS is the order to keep — every production check here takes an origin as an argument, which is what makes that possible.
  • Nothing in the core, the language, or any published artifact changed, and no URL moved, so the schema $ids, the sitemap, the README permalinks and the MCP shim's baked context are all untouched.

Deferred, by name ​

  • A main script for anything. Redirects, headers and routing are all expressible in wrangler.jsonc + _headers today. The moment one is not, the cap above is the price.
  • Moving the www redirect into the repository. It cannot be a _redirects entry (path-only matching). It could become a Terraform/API-scripted zone config; that is a bigger commitment to infrastructure-as-code than one redirect justifies.
  • Preview deploys per branch. Vercel gave one per push; nothing replaces it here. A branch push now deploys nothing, and site changes are verified locally with npm run e2e:docs / npm run e2e:playground against the built output.
  • archcanvas.uk. In the same account and unaffected by this ADR.

Alternatives considered ​

Stay on Vercel. Rejected as consolidation, not as a fix: the Vercel deployment worked. The split cost was two dashboards, two vendors' failure modes, and a DNS arrangement whose only rule was "do not let Cloudflare touch this".

Cloudflare Pages. Rejected — see above: maintenance-mode for new projects, and it puts the build command and output directory in a dashboard where this repository wants them in a reviewed file.

A Worker script that serves the assets. Rejected: it buys nothing the asset layer does not already do (html_handling covers clean URLs and trailing slashes, _headers covers content types and caching), and it converts every request into a metered Worker invocation.

GitHub Pages. Rejected: no per-host redirect rules, no _headers equivalent for the content-type corrections the machine-readable routes need, and the apex/subdomain split across two sites does not fit one Pages site per repository.