harness-gepa

Inspect, validate, and render GEPA genomes from the @metaharness/darwin library.

70.1k|8.4k|Updated Jun 2, 2025
One-click install
npx skills add https://github.com/ruvnet/claude-flow --skill harness-gepa
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: harness-gepa
Source: https://github.com/ruvnet/claude-flow/tree/main/plugins/ruflo-metaharness/skills/harness-gepa
Command: npx skills add https://github.com/ruvnet/claude-flow --skill harness-gepa

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires @metaharness/darwin, and includes scripts (resource) components.

What problem does it solve?

Evolved GEPA genomes are opaque JSON artifacts, making it hard to audit what system prompt they compile to, verify their structural validity before promotion, or diagnose why a run failed.

Core Features & Use Cases

  • Genome Loading & Validation: Load the shipped cand-6 promotion genome or any genome file and run structural validation, with an optional --alert-on-invalid exit code for CI gates.
  • System Prompt Rendering: Render the actual system prompt a genome compiles to, so you can review behavior before wiring a genome into a harness.
  • Transcript Failure Analysis: Classify failure modes in a run transcript using GEPA's failure-class taxonomy.
  • Use Case: Before adopting an evolved policy, run the render operation to read the exact system prompt cand-6 produces, then add the validate operation to CI so broken genome edits fail the build.

Quick Start

Run the harness-gepa skill with the render operation to show the system prompt that the shipped cand-6 genome compiles to.

Frequently Asked Questions about harness-gepa

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I validate a GEPA genome file in CI?

Run the validate operation with --path pointing at your genome file and add --alert-on-invalid. The script exits with code 1 when structural errors are found, so the CI job fails on broken genome edits.

How to see what system prompt a GEPA genome compiles to?

Use the render operation, which calls buildSystemFromGenome and outputs the compiled system prompt as JSON. Pipe it to jq -r .system to read the raw prompt text before wiring the genome into a harness.

Can this skill run the GEPA optimization loop itself?

No. gepaOptimize requires an in-process evaluate callback that cannot cross a subprocess boundary, so it is deliberately excluded. Use the library import directly or the harness-evolve skill for sandbox-scored evolution.

What happens if @metaharness/darwin is not installed?

The script first tries a local dynamic import, and on MODULE_NOT_FOUND falls back to a one-time npm install into a versioned cache directory. If the library cannot be installed at all, the skill degrades gracefully and still exits 0.

How do I classify failure modes from a GEPA run transcript?

Run the analyze operation with --transcript pointing at a JSON file containing the transcript entries. The script calls analyzeTranscript and returns failure classes from GEPA's failure-class taxonomy as a JSON object.