Codex skills for bounded auto-research in harness engineering.

EvoHarness turns code, traces, metrics, and candidate history into falsifiable harness hypotheses, executable candidate code, and held-out evaluation discipline. It also treats harness evolution as a versioned research process: bad optimizations are kept for diagnosis but the active frontier can roll back to the best earlier version.

This repository currently contains:

• skills/evo-harness: a workflow skill for setting up and running bounded auto-research loops over task-specific harness code.

The skill is intended for narrow, benchmarkable harness surfaces such as memory systems, retrieval wrappers, prompt/context builders, and tool-use scaffolds. It is not a replacement for building a full production agent product or doing open-ended scientific research.

The bundled plot_scores.py script can turn evolution_summary.jsonl into a version-vs-eval-score chart for tracking the frontier over time.

Copy or symlink skills/evo-harness into your Codex skills directory:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R skills/evo-harness "${CODEX_HOME:-$HOME/.codex}/skills/"

Then invoke it with:

Use $evo-harness to run bounded auto-research over this benchmark's harness.