They don't mention a twinkle that many task runners seem keen to omit: how do you handle things where there are human steps involved and not everything is automated? How do you track what has worked and what is still left to do if things go sidesways?

I built baker (https://github.com/coezbek/baker) for this some time ago (pre-LLM mostly). It uses markdown with embedded bash and ruby commands to give you a checklist which is run both automated for commands or with human in the loop for things which aren't automated (like login to some admin panel and generate that key, copy it here).

The checklist gets checked off both by human actions (you confirm that you did it) and automated e.g. success bash command runs. So you keep a markdown artifact on where you are in your project and can continue later.

You can wrap commands to run via SSH (of course clunkier than what scotty here does, but you can select a port for SSH).