I recently made a new website for td (“timmy’s distro”): www.tdlinux.org. From my point of view it’s serverless–blob storage fronted by a CDN. I didn’t want a git forge to be the landing page for the distro, so it was nice to have a homepage and a place where iso images and software updates can be hosted.
Today I’m writing this blog post in td-editor on Guix, but I ordered a framework desktop machine that I’m planning to install td on so that I can dogfood the distro 100% of the time. I should have all my main scenarios covered with the recent addition of td-sandbox letting me use the freedesktop sdk flatpak layer like a chroot environment so that I can run any traditional applications that won’t work natively (e.g. claude code, codex). Firefox should already be packaged, and other than that, I don’t have too many needs–a terminal, mail, news, td-photo, td-pass, td-editor, and that might be it.
Along similar news, I migrated from stalwart to td-mta, a new MTA that’s basically vibe-coded from the beginning. I have no idea how to run it or set it up–it was done purely with AI agents. Likewise, when I need to upgrade, I just tell an agent to upgrade production. I had an outage a couple of days ago due to a bug, which was quick to fix and redeploy, but it caused some concern with the underlying design. Here’s a root cause summary:
The mail server kept an internal receipt for every incoming message so it could confirm delivery if the database response was lost. These receipts accumulated indefinitely, and once the production account reached its 1,024-receipt limit, the server rejected every new message for that account with a temporary SMTP 451 error.
Unfortunately this raises more questions than it answers. From the best I could tell, this was an artifact of rewriting the SQLite backend to Postgres and some of the things the agent did during the migration. There is a slightly interesting story here–I originally wanted to try having the agent own it’s own mail database (i.e. no dependencies on any library), but after a while I realized it was actually creating a database (not just maildir and some indices), so I switched to SQLite. I was nearing production readiniess with SQLite when I decided to rewrite into Postgres. That late decision came because I wanted a way to backup/replicate the data while the MTA was running, and I also wanted to separate the database more logically from the code.
If I was handwriting code, I probably wouldn’t have overengineered this sort of receipt mechanism, but it seems fairly common with this level of AI model that one of them makes a documentation/requirement update, and it grows into something else over time. I’ve consistently fought with these sorts of issues throughout development of td, starting with the original AGENTS.md document being too dense with legalese, then with the build system using Guix as an oracle and spending a lot of time getting away from it. I was previously using Mes bootstrap to bootstrap the distro from a very small seed, but got rid of that in favor of using some debian binaries for the initial glibc/C compiler to compile the kernel, etc. With AI being able to reverse engineer and audit anything, it didn’t make sense to pay the high cost of build times to bootstrap the world.
As I write this the current battle of the day is AI agents writing a bunch of documentation and regression tests with every change. It becomes a lot harder to refactor something when asking for a boolean change becomes hundreds of lines of changes to update documentation and unit tests. Then you start spending a lot of time running tests for every change. Maybe it’ll all be solved by having the next version rewrite the previous version, sort of how a lot of human-written code is being rewritten now.