eval-harness

Eval-driven development (EDD) for Claude Code is a formal evaluation framework that structures testing across model iterations. You need it when researchers and engineers require capability and regression evaluations with structured eval lifecycles stored under .claude/evals.

To run regression evals, define an eval feature with /eval define feature-name, implement tests, then execute /eval check feature-name. This workflow supports multiple grader options including Code, Model, and Human graders with clear success criteria.

Yes, Claude Code evaluations support pass@k metrics alongside multiple grader options. You can implement custom code graders, use model-based grading, or define human grading criteria to evaluate capability and regression eval definitions across model versions.

The Claude Code eval framework provides a structured eval lifecycle from define to report with storage in .claude/evals. It stores eval results and integrates with /eval commands to manage the end-to-end evaluation workflow for model versions.

The best way to structure capability evals across model versions is using the EDD framework's formal lifecycle: define eval features with /eval define, implement tests with appropriate graders, execute with /eval check, and store results under .claude/evals for regression tracking.