@ruah-dev/eval

Eval

Same task — which executor wins, and what did it cost?

A fast micro-benchmark harness for AI coding agents. Run one task specification across multiple executors in fresh, isolated sandboxes with deterministic acceptance criteria and cost scorecards.

Source

See it

Quick start

Up and running in seconds

@ruah-dev/eval
$ ruah eval run spec.json --runs 5
# mock-pass 1.0   mock-fail 0.0

Features

01Ephemeral sandbox isolation per run (/tmp/ruah-sandbox-*)
02Multi-executor orchestration (Claude Code, Codex, OpenCode, Scripts)
03Deterministic criteria referee (powered by Ruah Verify)
04Multi-run variance analysis (--runs N with median & spread)
05Exportable scorecards in JSON, Markdown, and self-contained HTML
06Real cost-per-successful-delivery calculations

N

Executors

tmp

Sandbox

md|html

Scorecard

Command Reference

Commands

ruah eval run <spec.json>Execute benchmark spec across all configured executors
ruah eval run <spec.json> --runs 5Run multiple iterations for statistical variance
ruah eval compare a.json b.jsonDiff two benchmark scorecards side-by-side
ruah eval report <scorecard.json>Render scorecard as markdown or shareable HTML

Dive deeper

Full API reference, guides, and examples.