Assertion framework for AI agent evaluations - supports skill invocation checks, build validation, and LLM-based judging