What problem does it solve?
Manually iterating on harness policy prompts is slow and unmeasured. This Skill runs a GEPA learning cycle via metaharness learn that evolves harness genomes against a scored SWE-bench-style task corpus, promoting only candidates that measurably improve on held-out slices.
Core Features & Use Cases
- Measured genome evolution: Candidates are scored on held-out slices and only winners promote, replacing hand-edited prompt iteration.
- $0 dry-run by default: Resolves the slice manifest and reports cost without any model calls;
--run is the explicit spend opt-in.
- Graceful degradation: Without a metaharness checkout it reports
checkout-required with clone instructions; without the npm package it reports a degraded state rather than failing.
- Use Case: A harness policy underperforms on a task family, so you run a dry-run to price the learning cycle, then execute
metaharness learn --run from a terminal with an appropriate timeout to promote an improved genome.
Quick Start
Ask the assistant to run a harness-learn dry-run against your slice manifest with your chosen host and model to see the estimated cost before committing any spend.
