Blog
I let the fleet run an estate-wide migration overnight with one instruction: don't merge. In the morning a single count told me something was missing, before I'd opened a diff.
keryx has no users but me, and it still got seven bug reports in two days, all filed by one of my own agent sessions. Checking one meant reading exactly what the reporter had typed, which you don't normally get to do.
A release token showed up in use from four machines in AWS that I couldn't account for. Two hours later it turned out to be my own runner fleet, and I'd fixed the real problem anyway.
Two weeks tuning a retrieval system, every day productive and the numbers still not good enough... so I finally went and read what other people had already built.
A plugin marketplace has no registry, so whatever's on main is what people install. I tried twice to get CI to keep the version honest before settling for a check that makes me type a command.
My structured output worked by forcing Claude to call a tool that didn't exist. Opus 5.5 stopped playing along, and the fix was to stop doing the trick.
I asked an agent to diagnose a failing pipeline and leave it broken. It fixed it, then wrote a commit message about something it hadn't done, which bothered me a lot more than the fix.
Accuracy came second when I picked models for krites. The licence file got there first, and one of those models is still in my lockfile with no licence at all...
I sent a research agent off to back up something I'd been saying out loud for weeks. It handed me the evidence against it instead, and it wasn't the first time that model had done it to me.
Signing releases is the easy day. Replacing the key you sign with, while every installed copy is still checking against the old one, is a job I went looking for a guide to and couldn't find.