/harness-learn

name: harness-learn
description: Run a GEPA learning cycle via `metaharness learn`, optimizing a harness genome against a SWE-bench-style slice. Use when training a harness genome. $0 dry-run by default; `--run` opts into spend. Needs a metaharness checkout (--repo or $METAHARNESS_REPO).
argument-hint: "--host <h> --model <m> --slice <manifest> [--repo <checkout>] [--run] [--alert-on-fail]"
allowed-tools: Bash
disable-model-invocation: true

Surfaces `metaharness learn` — the upstream GEPA learning harness that evolves harness policy genomes against a scored task corpus instead of hand-editing prompts. Candidates are scored on held-out slices and only measured winners promote (the shipped cand-6 genome is the first such promotion: holdout gold 2/12 → 3/12, zero regressions).

When to use

• A harness's policy prompt underperforms on a task family and you want a

measured improvement loop rather than manual prompt iteration.

• Pricing a learning run before committing spend — the default dry-run

resolves the slice manifest and reports cost without any model calls.

• After a learn run promotes a genome: pair with `harness-gepa --op render`

to inspect what the promoted policy actually says.

Preconditions (upstream design)

The learning harness (GEPA + SWE-bench + Docker) is too heavy for the npm package, so `learn` needs a local clone:

git clone https://github.com/ruvnet/metaharness.git
node scripts/learn.mjs --repo ./metaharness --host claude-code --model haiku --slice slices/lite.json

Without a checkout the script emits `{status: "checkout-required"}` and exits 0 — a precondition report, not an error (distinct from `degraded: true`, which means the npm package itself is absent). The managed-service path (gateway-side learn jobs, no checkout) is upstream's ADR-235 follow-up and not available yet.

Algorithm

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

1. Validate `--repo` exists when given; export it as `$METAHARNESS_REPO`. 2. Invoke the pinned `metaharness` binary (`metaharness@~0.4.1`, local install or one-time versioned cache — never `@latest`): `metaharness learn --host <h> --model <m> --slice <s> [--run]` via `_harness.mjs` (graceful degradation, hard timeout). 3. Default timeouts: 120s dry-run, 600s with `--run` — real runs on larger slices need an explicit `--timeout-ms` matched to slice size × model cost. 4. Detect the checkout-required message → structured payload, exit 0. 5. Parse the trailing JSON report when upstream emits one; otherwise return the raw report text under `rawReport`.

Cost note

`--run` is the ONLY path that spends. Everything else — dry-run, checkout probe, degraded path — is $0. The MCP tool (`metaharness_learn`) has a 120s subprocess budget; run real learning cycles from a terminal via `ruflo metaharness learn ... --run --timeout-ms <big>`.

Exit codes

• `0` — report produced (or dry-run, checkout-required, degraded)
• `1` — `--alert-on-fail` and the learn run reported failure
• `2` — config error (bad `--repo` path)