Why one lens per dispatch Shipped
A single reviewer told to check everything checks the things it happened to notice. Splitting the roster is a way of buying attention rather than tokens.
Ask one model to review a diff for security, performance, accessibility, data modelling and API design at once, and you will get a report about whichever of those the diff most obviously implicates. That is not a weakness of the model; it is what a single pass over a prompt produces. Giving each concern its own dispatch, with its own criteria, means the accessibility question gets asked whether or not anything about the diff looked like an accessibility problem.
The cost control is selection, not brevity
The way to make forty dispatches affordable is not to make each one shorter — it is to not dispatch the ones that have nothing to say. That is what scopes and signals are for, and it is why they run before anything agentic does.