run-evals

Run a full eval round on stellar-raven-codemode from Codex, Claude Code, or another CLI agent — pick the right instruments (routing gate, QA headline, agentic, plan, live-data), distinguish the orchestrating agent from the spawned answering and judge agents, record results, review every answer/verdict, triage failures to root cause, and file evidence-backed upstream service-improvement findings in improvements/. Use when asked to run evals, check gates, measure a scoring/catalog/executor change, review QA verdicts, understand eval model roles, or close an eval round. The primary artifact of every round is upstream findings, not the scores.