Checked against the subagent reference rather than written from memory, and
three things were wrong.

A description is what the harness MATCHES to decide whether an agent is chosen
at all, and every custom agent's description is charged against one shared
budget. Ours were prose taken from the role's first LINE, so `reviewer` was
published describing itself as "says where the THINKING is wrong —", ending on
the dash where the sentence continued. Fixed at the source: a role states its
own `description:`, and the fallback is the first SENTENCE — never a line, which
truncates, and never a paragraph, which taxes every other agent.

The body IS the system prompt and REPLACES the whole of it rather than adding to
it. Ours were written in the third person, about the agent, so each one spent
its opening telling the agent what it was instead of what to do. Every role is
now addressed to it, and asks for the skills it needs by name — nothing is
inherited.

A role now states its own `model:`, because the cost of a role is a property of
the work it does and the role is the only place that knows what that work is.
Reviewing and judging idiom get opus; filing and merging get sonnet.

Declarations no longer repeat in the body, and the generated stamp names a
command that exists.