MetaHarness × RVM deployment-target partition

A minimal, opinionated agent-harness scaffold pre-configured for the RVM deployment-target partition. This scaffold gives you a working claude -p-style harness on disk in seconds: agents folder, MCP server wiring, settings.json with the RVM host conventions baked in, and a harness doctor health-check pass. It is intentionally minimal — it does NOT ship a vertical multi-agent template (no pre-built researcher/coder/tester crew), and it does NOT install or configure your underlying RVM runtime. Bring your own agents, bring your own runtime; this just gets the harness shape right.

Quickstart

npx @metaharness/rvm@latest my-bot
cd my-bot && npm install && harness doctor

If harness doctor returns green, you are ready to drop agents into agents/ and run them.

What you get

• settings.json — harness configuration pre-set with the rvm host partition (deployment-target conventions, path layout, permission defaults).
• agents/ — empty agent directory with the canonical layout the harness expects (one folder per agent, agent.json + prompt files).
• mcp.json — MCP server registration stub wired against the RVM host transport defaults.
• .harness/ — local state directory (gitignored) for session cache, doctor reports, and routing logs.
• package.json — harness CLI on the path via npx, plus a doctor and validate npm script.
• README.md — project-local readme stub you can overwrite.
• .gitignore — pre-populated with .harness/, node_modules/, and the RVM host's local artifact paths.

Advanced

Run the harness health check — verifies node version, settings.json schema, MCP server reachability, and host partition match:

$ harness doctor
[harness] node 20.x          ok
[harness] settings.json      ok (host=rvm, template=minimal)
[harness] mcp.json           ok (1 server registered)
[harness] agents/            ok (0 agents — add one to start)
[harness] rvm partition      ok

Validate just the config without touching the network:

$ harness validate
settings.json: valid
mcp.json: valid
agents: 0 (warning: no agents defined)

Run a one-shot prompt against the scaffold using Claude Code's plugin-dir mode (the harness is loadable as a plugin directory):

$ claude -p --plugin-dir my-bot "list the agents wired into this harness"

To re-scaffold over an existing directory (destructive — pass --force only if you mean it):

npx --yes metaharness@latest my-bot --template minimal --host rvm --force

FAQ

Q: Do I need RVM installed before running this? A: No. The scaffold only writes config files and directory structure that follow the RVM host partition conventions. You install/configure the actual RVM runtime when you are ready to deploy.

Q: Why is the agents/ directory empty? A: This is the minimal template — the point is to give you a clean harness shape, not opinions about which agents to run. If you want a pre-built multi-agent crew, look at the vertical templates instead of the minimal one.

Q: Can I switch the host later without re-scaffolding? A: Edit settings.json and change the host field, then run harness doctor. If the partition layout differs, doctor will tell you which paths to move. For non-trivial host switches, re-scaffolding into a fresh directory is usually cleaner.

License

MIT. Built on metaharness (https://www.npmjs.com/package/metaharness).

Deep-dive

Full explainer gist: https://gist.github.com/ruvnet/28353c033434e68f2c853546c0f99963