Escalate to Opus by need
A run picks its model from the work, not only the agent (src/crew/model_choice.ts):
- an issue serving the P0 pillar, or labelled
opus→opus - a planner's run →
opus - the next round on a pull that has two or more changes-requested reviews →
opus(not for the reviewer's own run) - anything else → the agent's own model
The run row records model and model_reason (agent, p0, label:opus, planner, bounced:N).
Settings → Crew (/settings/crew, "Per model per day"), the terminal's "Crew scorecard" list (admin only) and GET /api/v1/crew/scorecard (modelDays) show runs, spend and landed pulls per model per day.