As someone who maintains own coding harness, arguably more capable and powerful than Pi, I assure you the level of integration required to have something working really well nowadays goes against "this harness is yours" philosophy. "Codemode" completely changes the inter-workings of the harness, which spills all over everything else. And codemode is likely the inevitable future, as it's just much more efficient and powerful way for agents to do stuff. I'm planning to completely abandon non-codemode and make codemode the only "mode" in my harness as soon as open source models that I can run on my GPU can handle it well. I'm holding back as I don't want to deal with the complexity of having to support both.

> This just reads like Pi is becoming Armins and Marios opinions on what a good harness is instead of letting us build our own with the features we want and need.

Realistically it was always their harness. If you want "your own harness" you can always click the Fork button.

That seems a bit unfair. The choice isn't between opinionated vs unopinionated, it's "which layer do the options live at?". The harness has to be somewhat opinionated about the implementation of the agent loop, otherwise what is the harness exactly. I guess the question is if I don't want the new "code mode" changes in the loop can I trivially ignore it, but it seems it has to be implemented in the core.