eval-harness

To set up regression tests for Claude Code workflows, you apply eval-driven development by defining capability and regression evals, configuring deterministic code-based graders, and storing structured evaluation data in the .claude/evals directory to measure system reliability.

Eval-driven development formalizes evaluation for Claude Code sessions by standardizing testing. You should apply it when you need to validate new changes, measure reliability with pass@k metrics, and ensure audit-friendly documentation across model versions and prompts.

Yes, you can use model-based graders alongside code-based and human graders for Claude Code evaluation. This combination provides robust assessment capabilities, allowing deterministic code checks and human oversight options for audit-friendly documentation.

You measure regression stability across model versions by applying an evaluation framework through iterative cycles. This involves tracking pass@k and pass^k metrics to quantify reliability and maintain system performance when prompts or models change.

Yes, eval-driven development works without external dependencies because the framework uses configurable graders and structured eval storage. It relies on deterministic code-based graders and human oversight options to generate audit-friendly documentation natively.