I also grew to understand the value of people digging deeper into the underlying issue, instead of just answering "how do you do X in Y". The usual reaction was "I don't want to explain to you why I want to do it like this. Just tell me how to do this!"
I toyed with the idea of (on open source projects) having the human assign any PR-bot submissions to their own bot (cheapest one available will do) with the explicit instructions to cause as much rework as possible.
Sorta like a tarpit. Could be cheaper if the rejection is generated from a markov chain as that's going to be cheaper than even a cheap LLM.