Stats
Actions
Available In
Tags
A Claude Code plugin that analyzes your project and auto-builds a tailored agent team
What is a "harness"?
Originally, a harness is the set of straps, reins, and gear you put on a horse — the tack that lets a rider hold a powerful animal in place and steer it toward useful work. The word crossed into engineering as a metaphor: a test harness is the wiring of inputs, stubs, and assertions that lets a unit run in isolation. The same idea spread to evaluation harnesses for ML models (e.g. lm-evaluation-harness), and in the LLM era it reappeared as the agent harness — the prompts, tools, and guardrails that hold an LLM in place while it does real work.
oh-my-harness takes that last meaning one step further. Instead of dropping in a generic agent scaffold, it scans your project first, detects the stack and architecture, and then builds a scaffold tailored to the evidence it found — a team of specialized agents where each one only carries the context it actually needs. The horse is still the LLM; the harness is now cut to the shape of your project.
In one line
Say "harness", get a team.
Scans your project to detect framework, architecture, and test stack, then injects that evidence directly into agent prompts to build a project-tailored agent team. No DDD/Hex/NestJS baked in — those lenses only activate when the actual code says so.
Why
Claude Code is powerful, but one agent doing design, implementation, testing, review, security, and debugging all at once leads to bloated context and inconsistent quality.
oh-my-harness fixes this:
- 🎯 Separation of concerns — each agent has a single responsibility and enforced boundaries
- 🤝 Multi-model orchestration — Claude Opus for reasoning, Codex (GPT-5.4) for coding, or your own choice
- 🔍 Project-aware generation — analyzes your stack and auto-generates conditional specialists (e.g.,
domain-expertonly when Hexagonal is detected) - 🚪 User-gated skill synthesis — skills are not pre-baked from a fixed matrix; the harness proposes evidence-based candidates with domain-derived names, and only the ones you approve get created (zero is a valid answer) (new in v1.0.1)
- 💾 One-time setup, persistent team — harness config lives in
.claude/, survives across sessions - 🎨 Detection → Injection architecture — framework and architecture style are not pre-baked; they are detected at runtime and injected via a
<Project_Context>block (new in v1.0.0)
Quick Start
1. Install the plugin
/plugin marketplace add MoonDongmin/oh-my-harness
/plugin install oh-my-harness
2. Codex CLI (optional, recommended)
npm install -g @openai/codex
codex login
Without Codex CLI, coding agents fall back to Claude Sonnet.
3. Build a harness in your project
/oh-my-harness:harness
Auto-detects project type and delegates to harness-be (backend) or harness-fe (frontend). You can also invoke them directly:
/oh-my-harness:harness-be # backend-only
/oh-my-harness:harness-fe # frontend-only
Commands
| Command | Description |
|---|---|
/oh-my-harness:harness | Router — auto-detects BE/FE and delegates |
/oh-my-harness:harness-be | Backend harness builder (framework/architecture/ORM detection) |
/oh-my-harness:harness-fe | Frontend harness builder (framework/rendering/a11y/perf detection) |
/oh-my-harness:tdd | TDD team pipeline (Red → Green → Refactor, category-aware) |
/oh-my-harness:implement | General team implementation (refactoring, migration, structural change) |
/oh-my-harness:verify | Verify changes (build, tests, typecheck) |
/oh-my-harness:help | Help |
Core 6 Agents
Each agent has distinct permissions and specialization. Their base prompts are framework-neutral; the harness injects actual project evidence via a <Project_Context> block at generation time.
🏛 architect — "Every claim must be traceable to specific code"
Strategic advisor. Analyzes architecture and diagnoses root causes. READ-ONLY — never touches code, only reads and recommends with file:line evidence.
