Use case:
- Smoke test that mutates the host in a way that its effect can only be observed after an host restart.

In practice, this would require the bot:
- Interpret SIGTERM differently.
  - Have the bot not try to kill the task but instead "wrap up" and serialize its state
  - Upon restart, negotiate with the server taking back the task. Since there's no subprocess anymore, this is unclear how the workflow should be continued. 
- Tell the server that it is expected the bot to go MIA for N number of minutes, currently hardcoded to 5 minutes.

This is complicated enough that this will need a design doc.