2 - I generally run everything through openrouter. I've tried local models, but with subscription pricing it doesn't make sense for me to use local most of the time - though I do use local specialty models for Whisper and Chatterbox
3 - If I'm using Codex or Claude directly, I'm using the desktop apps. The apps are MUCH more useful than doing everything on the command line now.
But mainly, I'm switching to my own harnesses via my own agent factory.
4 - My own, called Workforce, which include general agent tasks + a lot of focus around code agents. (Workforce has agents that build and deploy Workforce, for example.). It also provides containers to agents for some bit of containment.
5 - My own... the same Workforce.
My main use cases are code and running a personal assistant I'm building called Clara. Its very similar to what a Hermes would do, but its all built exactly the way I want.
And that's my advice to everyone - don't use off the shelf agents and try to understand how they work - build your own and make them work how YOU want. There is nothing magical about OpenClaw and Hermes, and you can bootstrap one pretty quickly with Opus 5.5 now...
Edit: I should probably have said, my use case is software development - we're a consultancy that builds mainly back-office systems for customers - so I'm "coding" or having agents code 8-10 hours a day, and then my own projects after work. I just about make it under the $200 Claude plan limits though as I add more and more agents, that's going to get uncomfortable.