The site is a build, published at thankywal.github.io/muxel

[thankywal]

The console, the product page and the guide were served by an express process on a VPS, and the address every document sent a new owner to was app.muxel.site. That domain is going to another project, so both had to move.

The site is files

scripts/build-site.mjs writes the whole site, and GitHub Pages serves it:

/muxel/                      the console
/muxel/product/              what Muxel is
/muxel/docs/                 the guide, English
/muxel/docs/{my,th,ja,zh}/   the other four languages
/muxel/docs/{deploy-recovery,telegram-setup}/
/muxel/docs/media/           the pictures the READMEs show
404.html → the product page

The console is the site, not a page of it. It used to be chosen by hostname — app.muxel.site was the console and muxel.site was the product page — and the hostname that chose it is the one leaving. That is free: the console has no router, no history and no hash. Nothing in it reads the URL.

One record for the address

app.muxel.site was typed out in seventeen places — the Worker's own first screen, four languages of Telegram strings, the assistant's description of itself, the command line, two deploy scripts, the console's HTML — and in five READMEs besides.

CONSOLE_HOME now lives in @muxel/core, because the programs that name it cannot all import each other. The translations take it as a {console} variable rather than four copies of a sentence. The files that cannot import anything — the READMEs, the deploy scripts, the console's HTML — are held to the same value by a test.

The base path

A project site lives under /muxel/, a custom domain lives at /. The workflow reads that from GitHub (configure-pages → base_path) rather than deciding it here. Every root-absolute URL is moved onto that base on the way out, and what may be moved is the set of names the build actually emits — a link to something that is not there cannot be quietly repointed, and one nobody taught the build about fails it.

Two bugs the new test found

• An open bracket in the URL pattern ate text.replace(/docs/g, x) — a regular expression in JavaScript starts exactly the way a path does. It was one root name away from corrupting app.js.
• A link to the top of the site has no path segment to be recognised by, so href="/" was left pointing at thankywal.github.io/ rather than at the site. A lone slash is every split and every join, so only an attribute counts.

One site, two places it can be put

server.mjs now serves what the build wrote instead of building on demand, so the copy it hands a browser and the copy Pages hands a browser are the same bytes. It answers exactly as Pages does: /docs → 301 → /docs/, unknown paths → 404.html.

Verified

• 933 tests (83 files), build, typecheck, lint, worker bundle — all pass
• Built at both bases and driven in headless Chrome: every page 200, no console errors, no failed requests, no broken images
• Every internal link the built site makes — all 19 — resolves 200
• server.mjs against the root build: /, /product/, /docs, /docs/, /docs/my/, /app.js, /notice.json correct; /nope → 404

Also

.dev.vars.example still offered "live web search, reading a document as data" — both removed at revision 20.

---

Merging this turns GitHub Pages on (enablement: true) and publishes the site at https://thankywal.github.io/muxel/. Nothing is removed from the VPS by this PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoYEEHLJc4CifTLHjKpcDF