Looks really cool! I'm not sure which architecture choices ruled out agent subscription plans, but you should reconsider (unless enterprise is your target).

There have been a few very good harnesses - like Amp, which is excellent - which made this same choice and just haven't had the adoption they deserve.

Is there a middle ground , like where the "main" agent uses an API, but delegates using provider mechanisms (claude -p et al)? I don't mind paying for intelligence, but coding agents are expensive and at this time I am nearly exclusively using subscriptions for them, until the math (inevitably) changes.