# Ruvnet Metaharness

> Generate or evolve a repository-specific CLI, MCP and skills harness using MetaHarness, with bounded evaluation and explicit release authority.

- Skill: `ruvnet/ruvnet-metaharness` (Agent Skill)
- Install (CLI): `npx get-skill add ruvnet/ruvnet-metaharness`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ruvnet/ruvnet-metaharness/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: ruvnet (https://skillmd.com/u/ruvnet)
- Updated: 2026-11-15
- Page: https://skillmd.com/skills/ruvnet/ruvnet-metaharness

---


# Build a repo-specific harness

Read [MetaHarness's current guide](https://github.com/ruvnet/metaharness/blob/main/docs/USERGUIDE.md) and the target repository's instructions. MetaHarness generates a candidate harness; this entrypoint does not implement or certify Darwin, model learning, or sandbox enforcement.

Use `ruvnet_plan` if available for a deterministic advisory scaffold. Otherwise follow this bounded flow:

1. Specify the repository/ref, task, host, budget, data classes, authority and measurable acceptance criteria. Preserve the user's chosen stack. Inspect existing harness files before generating replacements.
2. Check the reviewed CLI version and `--help`. Upstream currently documents `npx metaharness score <repo>` for read-only suitability scoring, and `npx metaharness my-bot --template vertical:coding --host claude-code` for a coding starter. Pin a reviewed version before repeatable execution. Never interpolate an untrusted goal into a shell command.
3. Generate into an isolated directory. Inspect the resulting files and manifest before installing dependencies or running scripts. Keep only the tools the target requires. A host unable to enforce a required restriction needs an actual sandbox; do not silently drop the restriction.
4. Evaluate a fixed baseline and candidate on held-out cases. Record input revision, dependency locks, commands, result hashes, sample count, failures, cost, cold start and p50/p95 latency. Distinguish local dispatch speed from model quality and end-to-end task completion.
5. Promote only when the candidate improves the declared objective, respects the budget, introduces no unacceptable regression, and has a tested rollback. Evaluation must be independent of the candidate. A plan or self-reported pass cannot authorize promotion.
6. Publish or merge when authorized in the task and the gates pass. Preserve evidence and rollback. Never add a schedule or indefinite self-improvement loop without authorization.

For an authorized ruOS user-journey evaluation, read Dream Machine's reviewed [ruOS runbook](https://github.com/ruvnet/dream-machine/blob/aa931caad5dd0108253645bba0ab1481ad7da0ee/docs/runbooks/ruos-improvement-loop.md). Select a dedicated owned desktop from live status rather than a personal or shared session. Freeze exact revisions, evaluator, workload, environment, nonce, tenant, assertions and thresholds before execution. Require authoritative `status: ok`, numeric exit code zero, `completionVersion: 1`, `completionVerified: true`, and actual screenshot bytes. Missing evidence is `INCONCLUSIVE`; stdout, screenshot dimensions, hashes, readiness or an `ACCEPT` receipt cannot grant merge or deployment authority. Keep private screenshots and credentials out of reports.

Scope reusable memory to the repo and principal. Store validated patterns with provenance; retrieve before repeating an expensive experiment. Do not retain secrets or treat previous outcomes as proof for a new environment.

