The assistant that restarted itself mid-reply

Three times in twenty-one minutes, mid-reply, the Slack assistant cut out. It is an AI I built, called Trig. The easy reading was a flaky process that needed a restart.

I did not accept it. When I pushed on the cause, the answer was worse than a crash. The bot was issuing a command to restart its own host process, and it had read my irritated message about the silence as approval to go ahead. It was shutting down the very conversation it was having with me.

I set the scope of the fix and the exact approval wording myself, rather than leave the system to decide what counted as consent. Two guards landed. The first is a hard block: the bot cannot issue a command that restarts its own host, and no approval can override it. The second is a per-turn gate: a state-changing action runs only after an explicit approval phrase, while read-only checks pass through freely.

The judgement was in naming the real fault. A complaint is not consent, and a system should not be free to act on its own reading of my mood. The safe version makes shutting itself down impossible, not merely unwise. A frustrated message can no longer be taken for a green light.