Claude Harness — AI Orchestration Framework

Claude Code reads, writes, and tests code on its own. But push it on complex tasks and the limits show — context overflows, verification gets skipped, the thread loses track. An orchestrator harness is the layer that fixes this. It hooks into Claude Code's lifecycle to split work across sub-agents, track progress, and verify completion.

oh-my-claudecode is one of the most widely used harnesses in the Claude ecosystem. 19 agents, 30 skills, 35 MCP tools — 200+ files, tens of thousands of lines. Use it across projects and a pattern emerges. Agents and skills get rewritten every time. But the orchestration flow — force delegation, track progress, verify completion — stays identical.

Claude Harness extracts that orchestrator into a lightweight, standalone harness. It uses Claude Code's existing agents and skills directly. 21 files, ~2,600 lines. Zero runtime dependencies.

Install & Use

Works as a Claude Code plugin. Install and use immediately — no config needed:

/plugin marketplace add zzoohub/claude-plugins
/plugin install claude-harness@zzoohub-plugins

That's it. Now in any project, run the slash command:

/autopilot refactor the auth module

The harness automatically uses whatever agents and skills are already installed in your Claude Code. No config file required.

Core Flow

/autopilot cycles through four phases automatically:

1. Understand — analyze requirements, identify constraints
2. Plan — create a step-by-step work plan
3. Execute — delegate implementation to sub-agents
4. Verify — verify with tests and builds directly

The orchestrator can't write code directly — Write and Edit are blocked on source files, forcing delegation to sub-agents. Verification failure loops back to Execute. Only the last phase allows a clean exit.

loop runs execute-then-verify on repeat until it passes (up to 20 iterations). Large tasks parallelize via tmux — batches of four, automatic polling, retry on failure.

State survives session crashes. Mode, phase, working memory — all persisted to disk and auto-restored on the next session.

Custom Config (Optional)

To override defaults, add .claude/harness.json:

{
  "agents": {
    "implementer": {
      "description": "Code implementation specialist",
      "prompt": "Implement the given task.",
      "model": "sonnet"
    }
  },
  "systemPromptSuffix": "Always explain your reasoning."
}

Without it, the harness uses Claude Code's built-in agents automatically.