- BudgetTracker tracks LLM token cost separately from tool cost
- LlmClient auto-records cost into request-scoped budget
- Router selects tier from classify confidence, intent risk, escalation count
- Supervisor short-circuits via regex fast-path when rule confidence is high
- Onboarding skill loads user escalation history to inform routing
Add skill priority routing and duplicate manifest handling.
Split eval golden sets by skill and add moderation rule lookup coverage.
Move onboarding runtime helpers out of the main skill class.