metaharness/examples-packages/coding at main · ruvnet/metaharness · GitHub
Skip to content

coding

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

coding

Folders and files

README.md

MetaHarness: coding vertical

⚠️ Illustrative output. Transcripts and validation/run output shown in this README are representative examples, not captured from a specific run — actual output depends on your environment, models, and inputs. Run the commands to see real results.

A ready-made multi-agent engineering pod for Claude Code. One command scaffolds a working harness with four specialized agents — architect, implementer, reviewer, test-writer — wired into Claude Code's plugin directory layout with sensible permissions, hooks, and tier-appropriate model routing. Use it when you want a starting point for shipping real code with AI pair-programmers instead of staring at a blank .claude/ folder. It does NOT install a runtime, does NOT call a model on its own, and does NOT lock you into any specific MCP server — it is a scaffold, not a service.

Quickstart

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

harness doctor confirms the scaffold layout, agent definitions, and Claude Code plugin manifest are valid. After that you can point Claude Code at the directory (claude --plugin-dir .) and the four agents become spawnable via the Task tool.

What you get

  • agents/architect.md — design + decomposition agent, tier: opus
  • agents/implementer.md — code-writing agent, tier: sonnet
  • agents/reviewer.md — diff review + correctness/security pass, tier: sonnet
  • agents/test-writer.md — TDD-style test generation, tier: haiku
  • .claude/settings.json — pre-configured permissions, hook stubs, model routing defaults
  • CLAUDE.md — house rules describing the pod, the SendMessage pipeline, and when to escalate
  • harness.config.json — declares the vertical (coding) and host (claude-code) so harness doctor knows how to validate

Advanced

Validate the scaffold structure:

harness doctor
# ✓ plugin manifest .claude/settings.json
# ✓ 4 agents found (architect, implementer, reviewer, test-writer)
# ✓ host: claude-code
# ✓ vertical: coding

Strict-validate agent frontmatter and tier assignments:

harness validate --strict
# agents/architect.md      tier=opus     ok
# agents/implementer.md    tier=sonnet   ok
# agents/reviewer.md       tier=sonnet   ok
# agents/test-writer.md    tier=haiku    ok

Headlessly drive the pod from Claude Code:

claude -p --plugin-dir my-bot "Ship a typed REST client for the GitHub issues API with tests"
# spawns architect → implementer → reviewer → test-writer via SendMessage pipeline

Regenerate over an existing directory (force overwrite):

npx @metaharness/coding@latest my-bot --force

FAQ

Q: Do I need an Anthropic API key to run the scaffold? A: No. The scaffold itself is just files — agents, settings, and config. You need a Claude Code install (and whatever model access it's configured with) to actually run the agents.

Q: Can I swap the model tiers? A: Yes. Each agent's tier lives in the frontmatter of its agents/*.md file. Change tier: opus to tier: sonnet and rerun harness doctor to re-validate.

Q: How is this different from just copying a .claude/ folder from GitHub? A: The scaffold is versioned, includes harness doctor / harness validate so drift is detectable, declares its vertical so future tooling can target it, and is regenerable in place with --force without losing your custom CLAUDE.md edits (those live alongside, not inside, the generated files).

License

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

Deep-dive

Full explainer gist: https://gist.github.com/ruvnet/640c8b74b2e57690f81d130acf3cf439