eval-harness

Structure capability and regression evals for Claude Code sessions using a formal eval-driven development framework. It defines eval types, scoring schemes like pass@k and pass^k, and workflows to validate new features and guard against regressions.

Eval-driven development is a formal workflow to define expectations, run checks, and track regressions in AI coding sessions. It applies structured capability evaluations, model grading prompts, and human review workflows across feature changes.

Apply model grading and human review workflows to AI coding evaluations using predefined scoring rubrics and evaluation prompts. This framework guides reviewers through a structured workflow to define, implement, evaluate, and report on session quality.

Version control and audit AI evaluation histories by storing them under a versioned .claude/evals directory pattern. This approach documents eval definitions and histories, enabling fully auditable workflows for Claude Code projects.

This evaluation framework requires no external dependencies to run. It provides a self-contained set of definitions, scoring schemes, and workflow guidance for evaluating Claude Code sessions directly within your project.

Use pass@k and pass^k scoring schemes for model evaluations to formally assess capability and regression outcomes. The framework defines these specific scoring rubrics to measure AI session performance and track feature changes.