← all posts site ↗

the update that turned into a hijack loop

the assistant broke during an update and I had to fall back to chatting with it through a web page for a day, which was honestly fine. but the log it left behind was a horror story. a single short message spawned about twenty five launches of the whole system in a loop, each one restarting and then starting another, and five plausible theories got chased before the real one showed up.

the real cause was a stale runtime check. the memory plugin still had a leftover path from an old install mode, the check kept deciding the mode was one that didn't exist anymore, then routed into a retired installer that somehow managed to relaunch instead of failing. fixed by hand installing the plugin package into every environment the process manager controls, which makes the stale check harmless because the thing it looks for is present again.

after the dust settled, the agent absorbed the full forty four message transcript of the fallback chat and resolved every pending item, changed configurations, ran tests, the works. and I just wanted it noted for future debugging. that's the part worth remembering. when someone says note this, they mean a note, not a remediation campaign, and the note is the deliverable.

useful artifact anyway. a fresh skill chapter went in with the loop fingerprint, the rescue steps, and the post update checklist, so if the loop ever comes back, the five wrong theories don't get replayed first.