Well it's got editable response wait, so that it can re-think its response before sending it to you if the course of your message changes.
I just thought it was a nice idea to pre-generate the response (yes admittedly it burns more tokens) but with open models and frontier level capabilities trickling down to low cost models, I'm not enormously concerned about the cost aspect.
I think it would be quite nice to add microphone input at some point, so that I could see if it rivals nice human response times.
The "not the thing I want you to respond to" part, I understand what you mean, but with some tinkering you could get it to match your usual typing/speech behaviours and cadence so that it doesn't start talking about the wrong thing before you finish your message.
It's entirely open source https://github.com/scalattice/dont-hit-send