/harness-gepa

name: harness-gepa
description: Inspect and audit GEPA genomes via the `@metaharness/darwin/gepa` library entry (darwin 0.8.0) — load/validate a genome (default: the shipped cand-6 promotion), render the system prompt a genome compiles to, or classify failure modes in a run transcript. The `gepaOptimize` loop itself is library-only (bring your own evaluator) and not surfaced here — use `harness-evolve` for sandbox-scored evolution. Degrades gracefully when @metaharness/darwin is absent.
argument-hint: "--op genome|validate|render|analyze [--path <genome.json>] [--transcript <t.json>] [--alert-on-invalid]"
allowed-tools: Bash

Surfaces the GEPA (genetic-evolution prompt-adaptation) *library* exports from `@metaharness/darwin/gepa`. Unlike the other skills in this plugin there is no CLI binary behind this — the script dynamic-imports the library (local resolution first, versioned cache install as fallback) and calls the subprocess-safe subset.

When to use

• **Adopting an evolved policy**: `--op render` shows the actual system

prompt a genome compiles to — read THAT, not the raw JSON, before wiring a genome into a harness.

• **Auditing a promotion**: `--op genome` loads + validates the shipped

cand-6 genome (first holdout-confirmed cheap-tier promotion; provenance ships in the package) or any genome file you point at.

• **CI gate on genome edits**: `--op validate --alert-on-invalid` exits 1

on structural errors.

• **Debugging a bad run**: `--op analyze --transcript run.json` classifies

failure modes (GEPA's failure-class taxonomy) from a transcript array.

What is deliberately NOT here

`gepaOptimize` — the optimization loop takes an in-process `evaluate(candidate)` callback ("bring your own evaluator") that cannot cross a subprocess boundary. Two supported paths instead:

1. **Library consumers**: `import { gepaOptimize, loadCand6Genome } from '@metaharness/darwin/gepa'` 2. **Sandbox-scored evolution**: `harness-evolve` (darwin CLI `evolve`), which pairs GEPA with its own sandbox evaluators.

Algorithm

Implementation: [`scripts/gepa.mjs`](../../scripts/gepa.mjs).

1. `import('@metaharness/darwin/gepa')`; on MODULE_NOT_FOUND fall back to a one-time `npm install --prefix ~/.ruflo/darwin-cache-0.8.0` and import the cached `dist/gepa/index.js` (versioned dir → pin bumps invalidate). 2. Dispatch `--op`:

• `genome` → `loadGenome(fs, path)` or `loadCand6Genome()` + `validateGenome`
• `validate` → `validateGenome(rawJson)` (raw parse so broken files reach

the validator instead of throwing in the loader)

• `render` → `buildSystemFromGenome(genome, ext?, glob?)`
• `analyze` → `analyzeTranscript(entries)`

3. Emit one JSON object; exit 0 (or 1 under `--alert-on-invalid`, 2 on bad input).

Examples

node scripts/gepa.mjs --op genome                          # cand-6 + validation
node scripts/gepa.mjs --op render | jq -r .system         # what does cand-6 SAY?
node scripts/gepa.mjs --op validate --path my-genome.json --alert-on-invalid
node scripts/gepa.mjs --op analyze --transcript run.json

Exit codes

• `0` — op completed (or degraded — darwin not installable)
• `1` — `--alert-on-invalid` and validation found errors
• `2` — config error (unknown op, missing/broken input file)