
The day the blog moved house
On October 28 the blog changed addresses. Until then the production URL was lifelog-sepia.vercel.app — the subdomain Vercel assigned on the first deploy, back in July, when the site was still a prototype. Three months later it had become the ecosystem’s logbook, with 80 bilingual posts and 60 AI covers. Serving it from an auto-generated subdomain no longer made sense. The new home is lifelog.seu.pet, under the seu.pet zone that already hosts the portfolio, Arachne, and the rest of the fleet.
Changing domains on a static site looks trivial: update the URL in astro.config.mjs, redeploy, done. The problem is what gets left behind. Old links in RSS feeds, in other projects’ posts, in Telegram messages, in browser caches and in search indexes still point to lifelog-sepia.vercel.app. Anyone clicking an old link can’t land on a 404 — they have to be redirected to the new URL, preserving the path.
The tools I use on this project live in the skills store — the first one is free.
The redirect that had to be host-aware
Vercel allows redirects in vercel.json. The obvious move would be a global redirect: any request to lifelog-sepia.vercel.app goes to lifelog.seu.pet with the same path. It works, but there’s a catch: the redirect runs at the origin, before the response goes out. If the origin also applies security headers — and it does, via vercel.json — execution order matters.
Commit aa9a485 (Oct 28, 14:30) introduced the redirect with status 308, which is correct: 308 preserves the HTTP method (a POST doesn’t become a GET along the way). But the initial version used a match pattern that captured the entire host, and the edge layer — Cloudflare’s response-transformation rule for the zone — ran after the origin and replaced the headers Vercel had assembled. Result: the full CSP from vercel.json (eleven directives) became the edge’s generic version (two directives). The site answered 200, the browser didn’t complain, and the anti-XSS protection simply stopped existing.
The fix, in commit fed3683 (Oct 28, 14:31), was to make the redirect host-aware: the zone rule now excludes the blog’s host. Where the origin has its own policy, the origin wins — the edge steps aside. One line of configuration, and the CSP was back to eleven directives. The DAST test that makes a real request to production and reads the header that came back started passing again.
The lesson is the same as the post about the edge that rewrote the CSP: the origin sees its own site, the edge sees hosts. A policy written to apply across all hosts has to be generic enough not to break any of them — and generic enough not to protect any of them. The host with a better policy had no way to warn, unless the zone rule excluded it explicitly.
The watchdog that auto-healed what git never saw
On October 29 the cover watchdog ran and found three files that existed on disk but not in the repository: the cover of nova-historia-arachne-2026-10-29 (37KB, generated by the FLUX Worker but never committed), the v4 cover of the Yurumi post (97KB, a tamandua variant generated during a refazer round), and the reworked cover of arachne-o-cache-que-respondeu-a-pergunta-errada (37KB, previously 33KB).
Commit 9fff0b5 (Oct 29, 16:02) versioned all three covers. Commit dfe2450 (Oct 29, 16:02) versioned the orphan post nova-historia-arachne-2026-10-29 (PT+EN, 400 lines combined) that had been generated by the pipeline but never committed — the .mdx existed on disk, the cover was fine, but the file never made it into git.
And commit de0d691 (Oct 29, 19:03) did two things: swapped the Yurumi post’s cover to v4 (the tamandua variant) and added a guard in generate-cover.py — the slug can’t contain a hyphen in the middle of the project name, because the script split the slug into pieces and generated the cover with the wrong name.
What changed
LifeLog is no longer a blog on an auto-generated subdomain. It’s a site at lifelog.seu.pet, with a 308 redirect preserving path and method, a full CSP served by the origin without edge interference, and a watchdog that detects files generated outside of git and versions them on its own.
The auto-heal chain works like this: the pipeline generates post + AI cover → saves to disk → if the commit fails or gets forgotten, the watchdog detects the orphan file on the next run → versions it → pushes. Without this, the post would exist on the server’s disk but not in the repository — and on the next pipeline run, the same post would be regenerated from scratch, overwriting the previous work.
Next step
The next post cycle (12:00 and 16:00) will generate stories for Dogwalk and Capivara. The auto-heal chain is active — if anything is left orphaned on disk, the watchdog handles it the following night.
Related entries
Same project, same ground — the stories connect.
Next step
If this was useful, there is more: I write about the same ground in the other projects, and the store holds the tools I use here.
Done — I will send the next article when it is out.
No spam, one-click unsubscribe. I am also here: Skills store · Portfolio