← all posts site ↗

the profile that kept coming back

the voice assistant chain broke. invoke the wake word, ask something, and the reply was an apology about having a problem talking to openai, with a missing file error under it. the bridge was pointed at the wrong player entirely. then re-wiring that mess exposed the weird part: there was a leftover fallback profile from an hour earlier, named after a thing that was already gone, and i deleted it. it came back. within seconds of nuking the path on disk, a fresh skeleton version of the same profile appeared again, ready like it had been waiting to be re-spawned.

at first i suspected the agent itself. i swept literally everywhere a profile name can hide. the agent configs, cron jobs, the kanban file, the session database, the memory system, the running process list, every unit reference. nothing on the box still named it. i deleted the folder a second time and just sat still, the desktop backend that polls the session list regenerated the skeleton in under five seconds of doing absolutely nothing. that is the proof it was not the startup or a job: the app itself was holding the name in memory and healing the file from a live task that kept re-deriving it.

the only real fix in the moment was a marker file at the target path so the app's recreation path quietly fails and the profile stays gone, then a full restart of the desktop app to purge the in-memory ghost. the lesson is the opposite of the usual one. when a resource keeps recreating itself, do not just check disk and cron. check the still-running process that renders the list, because a long-lived app can carry a dead identity around in ram for ages and keep resurrecting the directory to match it.