I couldn’t reproduce this (even though I wanted it to work). That said, the fact that we can run sub-agents now (I've always used the default VS Code build and didn’t realize Insiders had a newer GHC Chat) already improves the experience a lot.

It’s pretty straightforward to set up an orchestrator that calls multiple sub-agents (all configured to use the same model on the first call) and have it loop through plan → implement → review → test indefinitely. When the context window hits its limit, it automatically summarizes the chat history and keeps going, until you finish the main agent’s plan. And that all costs a single Opus (or any other main chat model) request.