eval-harness

Eval-driven development for Claude Code sessions is set up by defining markdown-based capability and regression evals, configuring graders, and storing results in the .claude/evals directory to track reliability across revisions.

Software testing evals support three types of graders: code-based graders for automated validation, model-based graders for AI evaluation, and human evaluators for manual review, ensuring flexible validation across different testing contexts.

To prevent regressions in Claude Code projects, define and run regression evals that verify existing functionality remains intact, storing baselines and results in .claude/evals to audit changes across revisions.

Yes, eval results and baselines can be tracked across revisions by storing eval definitions and outputs in the .claude/evals directory, enabling audit and iteration throughout the development lifecycle.

Automating capability checks for Claude Code features is best done by defining targeted capability evals in markdown, applying code-based or model-based graders, and using predefined workflows to scaffold, execute, and report on evals.

No external dependencies are required to run evals for Claude Code sessions. The framework operates independently, using markdown-based eval definitions and built-in grader configurations stored locally in .claude/evals.