oh-my-harness

Oh My Harness

Spec-driven, weight-aware autonomous Claude Code harness. Define the goal — it loops until done.
An autonomous loop that self-verifies and cross-verifies, weight-aware routing with multi-model verification, a living on-disk state anchor, plus smart defaults, test enforcement, model routing, and multi-agent orchestration — all through native hooks.

한국어 · Get Started · Features · Multi-Agent · Config · Architecture

---

Why Oh My Harness?

Claude Code is powerful out of the box — but it stops at the end of a turn even when the job isn't done, it doesn't verify its own work against the goal, and it treats every request the same regardless of complexity.

Oh My Harness (OMH) turns Claude Code into a spec-driven autonomous harness: you define the goal once in a SPEC.md, and OMH loops — implementing, self-verifying, and cross-verifying — until the spec is objectively met. The loop runs on Claude Code's native hooks, so the harness owns when to keep going and when to stop — never the model's self-assessment. Around the loop sit the same lightweight smart defaults (test enforcement, dangerous-op guard, model routing, multi-agent) that make every session safer.

graph LR
    A[You type a prompt] --> B{OMH Hooks}
    B --> C[Ambiguity? Ask first]
    B --> D[3+ tasks? Plan mode]
    B --> E[rm -rf? Warn]
    B --> F[Code changed? Test reminder]
    B --> G[git commit? Convention check]
    style B fill:#7C3AED,color:#fff

---

Philosophy

Autonomy with real walls.

Earlier OMH was "warnings instead of walls." The autonomous loop changes that where it counts: a loop that can't be stopped is dangerous, and a loop that stops too early is useless — so the loop has real walls. The harness forces continuation while the goal is unmet and under budget, and forces termination on objective signals (verify ladder green + cross-verify, or a guardrail: iteration/wall-clock budget, no-progress, oscillation). The model never decides it's "done"; the harness does, against machine-checkable acceptance criteria.

Everywhere else, OMH stays the harness you barely notice — smart defaults that guide with warnings, and project-specific skills auto-scaffolded from your detected stack (test conventions, review checklists, lint workflows) that you own and customize.

• Built-in skills (agent management, setup) stay in the plugin
• Project skills (code-review, test-write, lint-fix) live in .claude/skills/ — your project, your rules
• Run /init-project to scaffold, then customize freely

---

Quick Start

Install via the Claude Code plugin marketplace — one copy-paste:

claude plugin marketplace add Hoya324/oh-my-harness
claude plugin install oh-my-harness@oh-my-harness

That's it. Zero setup required — OMH works on sensible defaults the moment it's installed, and harness features (including the autonomous loop) activate automatically. /harness-setup is optional and only needed if you want to tune harness.config.json.

---

Updating

When a new version is released, update to get the latest hooks, detection patterns, and features.

# Pull the latest plugin version
claude plugin update oh-my-harness@oh-my-harness

# Re-apply updated hooks and dictionary
/harness-setup

Note: updating preserves your existing harness.config.json. Only hooks, commands, and CLAUDE.md instructions are refreshed.

---

Features Overview