🛠️ Letting the Schema Grow Up

I just merged in support for versioned plan schemas, including the debut of v2. This version is way more complex than v1. It's clearly meant for machine consumption, not human clarity. It’s full of token budgets, token caps, cross-referenced JSON files I haven’t even verified exist. A lot of it sounds reasonable, but I can’t tell how much real value it adds.

Still, this isn’t a product layer thing — it’s infra. GPT-5 wanted it, and I’m treating this as one of those “let the LLM own its domain” moments. It seems like it did the work. So I’m shipping it and moving on.

I also find this an exercise in exercising the trust needed to make this whole LLM-driven coding flow work. Nothing about it the idea makes sense if I can't let the LLM make a million decisions--large and small--about how to implement the code, so if the LLM wants more rigor in the plans, the LLM gets more rigor in the plans.

This lands in the middle of the minimality block, which is actually perfect. If the new schema can survive a block dedicated to eliminating unnecessary code, that’s a good initial test. That said, minimality’s an abstract target, and I probably won’t know if this schema direction holds water until I’ve run it through a few more blocks.

Right now, I’m in observation mode. The structure’s in. Let’s see if it earns its weight.