Vibecoding 025 — go.short: a custom URL shortener, admin UI, and linktree

I’ve been pasting long blog post URLs into my Mastodon bio, notes, and slide decks for months. The obvious fix was a URL shortener. The less obvious part was how much Cloudflare’s routing layers can fight each other when you’re building one.

Linktree: mr-dinesh-links.pages.dev · Short links: s.mrdee.in · Admin: go-admin-76o.pages.dev

What it does #

A slug like s.mrdee.in/vc1 hits a Cloudflare Worker, does a KV lookup, and returns a 301 to the destination URL. That’s it. The interesting parts are everything around it.

The admin UI is a single HTML file deployed to Cloudflare Pages. Password authentication via Authorization: Bearer header checked against a Worker secret. List all slugs, add new ones, delete existing ones — all against a JSON API served by the same Worker.

The linktree (s.mrdee.in/links) is a static page deployed to Cloudflare Pages. It groups all 37 short links into three sections — live tools, blog & notes, and all 25 vibecoding posts — so there’s one URL to share that maps the whole project.

Architecture #

The Worker handles four routes:

• GET /<slug> — KV lookup → 301 redirect or branded 404 HTML
• GET /api/links — list all slugs (auth required)
• POST /api/links — create/overwrite a slug with {slug, url} (auth required)
• DELETE /api/links/<slug> — remove a slug (auth required)

Slug sanitisation strips everything outside [a-z0-9_-] on write. Destinations without a scheme get https:// prepended automatically.

The DNS rabbit hole #

Getting s.mrdee.in to invoke the Worker was not straightforward.

The initial domain was go.mrdee.in. Every request returned a 301 back to itself. wrangler tail showed zero Worker invocations — the request never reached the Worker. Something upstream was swallowing it.

The culprit was a Cloudflare Redirect Rule in the zone settings, set to apply to all incoming requests:

wildcard_replace(http.request.full_uri, "http://*", "https://${1}")

The intent was HTTP → HTTPS. The bug: http://* doesn’t match https://..., so wildcard_replace returns the original URL unchanged, and Cloudflare issues a 301 to the same HTTPS URL. A self-redirect loop for every single request, before the Worker ever fires.

Redirect Rules fire before Workers in Cloudflare’s processing pipeline. Deleting the rule fixed the loop immediately.

Why go.mrdee.in stayed broken #

Even after fixing the redirect rule, go.mrdee.in was stuck. The DNS A records for it were Pages-managed and locked — created when an earlier project used *.mrdee.in as a wildcard custom domain. Locked records can’t be edited or deleted through the dashboard or API.

The cleanest fix was to move to a new subdomain: s.mrdee.in. A Workers Custom Domain on s.mrdee.in created the right AAAA 100:: DNS record (Cloudflare’s null-routing record, carrying internal worker routing metadata in meta.origin_worker_id) and the Worker started receiving traffic.

Restoring the blog #

During cleanup, the mrdee.in apex temporarily lost its DNS records. Pages custom domains at the apex use a AAAA 100:: record with internal routing metadata — not the usual kind you’d hand-edit. When we manually created the record, we got an empty meta: {} which Cloudflare’s routing layer doesn’t know how to handle:

{ "type": "AAAA", "content": "100::", "meta": {} }

The fix: delete the bare record, delete the stale Pages domain entry, then let Pages re-add the domain itself — which creates the AAAA 100:: with the correct routing metadata populated. For the apex, Pages also needs a CNAME mrdee.in → curl-notes-blog.pages.dev to pass its own domain verification check (Cloudflare flattens it at the edge for external queries). Once both were in place, the domain went active within 30 seconds.

One wrangler kv note #

KV is exactly the right storage for this — slug lookups are read-heavy, values are small, and eventual consistency on writes is fine for a personal shortener. The only limitation worth knowing: wrangler kv key list returns all keys but not values, so populating the admin UI requires a GET per key on first load. Fine at 37 slugs, would need rethinking at thousands.

Slugs #

All 37 short links are at s.mrdee.in/<slug>. The linktree at s.mrdee.in/links has the full list grouped and labelled. The admin UI at go-admin-76o.pages.dev lets me add or remove slugs without touching the code.

The ones I actually use most: s.mrdee.in/blog for the vibecoding series, s.mrdee.in/argus and s.mrdee.in/juicesec in talks, and s.mrdee.in/now in the Mastodon bio.