I ran this a couple of times against terminal bench with different models. I saw that some models aren't as good with code mode, but overall codemode doesn't slow them down. I was usually at about 70% on tbench 2.1 with Sol.

The issue I have with existing harnesses is that they try to solve all issues for all people. I have workflows I use with fork, and a history tree view is another important thing I needed. One of the issues they don't solve for me is easy visibility and good recap (Claude is okay ish).

Additionally, I loved Amp Codes UI/UX and was incredibly inspired by iy