What kind of work is coming through — SDLC category by cost & prompt share (the headline).
Is the right model used for the task? (e.g. Opus for “1+1” = overkill).
One-shot vs rework vs clarification, per prompt (PES). Descriptive, not a target.
What’s being used — incl. subagent spawns (first-class).
Cost attribution by rchain node (manager rollup). k-anon ≥5; your own slice shown to you.