Оценка AI-сессий по методологии EDD

★ 7.9 · testing

eval-harness is a Claude Code skill that implements a formal evaluation framework for AI-assisted sessions following eval-driven development (EDD) principles. It organises evals into two categories: capability evals that verify new agent behaviours, and regression evals that guard against breaking existing functionality. Reliability is measured with pass@k («at least one success in k attempts») and pass^k («all k attempts succeed») metrics, with a typical target of pass@3 > 90%. Three grader types are supported: code-based (deterministic bash checks), model-based (Claude evaluates open-ended outputs on a 1–5 scale), and human (flags changes for manual review by risk level). Eval definitions, run logs, and regression baselines are stored under `.claude/evals/` and versioned alongside the codebase — making this skill suited for teams that need systematic quality control over AI agents across prompt or model changes.