In our case, we often have complex decision trees to diagnose issues before we get into the ops. Our playbooks are very thorough with details about all of these debugging techniques.

You can never completely automate the debugging step: it often requires human judgement to direct the investigation in ways that can't really be expressed in a script.

The friction of the shell comes up not in executing the ops, which we do in fact have heavy automation around, but in all the little pieces of debugging has to happen to tell you what ops need to be done.

(If you get paged and the solution is to skip the debugging step and run a script, then why didn't you automate the script so that you don't get paged in the first place?)