harness-kit

🔧 HarnessKit

Harness Engineering: A reliable AI agent is not just a raw model. It is defined as: $$\text{Reliable Agent} = \text{Model (AI)} + \text{Harness (Controls)} + \text{Human Auditor}$$

HarnessKit is a complete AI-assisted software engineering methodology built on Harness Engineering—the principle that true reliability comes from enclosing generative models inside structured execution scaffolds and human-driven governance loops.

---

👑 The Core Engine: Autonomous Orchestration & Live Auditing

At the heart of HarnessKit is the autonomous-orchestrator skill. Once provided with the initial task scope, it runs an atomic, continuous execution cycle without stopping, pausing, or asking redundant questions—fully automating domain planning, TDD execution, and multi-agent code reviews.

However, the human engineer is never replaced: your role evolves into Live Auditing.

While the orchestrator executes continuously, you act as the Human Auditor in the cockpit, tracking the live stream through your coding workspace (Claude Code, Cursor, OpenCode, Gemini, Copilot, etc.). The AI moves with sovereignty, but you maintain continuous telemetry and oversight.

⚡ Hot-Interception: Absolute Human Command

Because the engine runs seamlessly without waiting for permissions at every step, you use this live observability to dynamically intercept the loop when necessary:

• Pull the Emergency Brake: Forcefully kill the execution (Ctrl+C) the moment you notice the AI has adopted an incorrect architectural premise.
• Live In-Flight Injections: Hot-patch the active backlog, append newly uncovered constraints, or update domain specifications while the loop is running.
• Dynamic Parameter Tweaking: Modify configuration thresholds on the fly—lower the validation score target (default 0.70) to accept a minor style debt, or increase maxReworks directly inside the configuration files.

🔍 Socratic Code Review in Action

To prevent systemic risks (such as N+1 queries, memory leaks, security vulnerabilities, or database connection exhaustion), HarnessKit employs a Socratic Code Review model. The orchestrator invokes the the-grumpy-tech-lead to validate the code inferentially by asking deep architectural questions rather than providing copy-paste solutions.

Below is a visual example of how this interactive code review occurs under your watch:

🤖 Autonomous State Machine

The loop is driven by a robust Product State Machine. It tracks feature backlogs and dynamic transition gates, ensuring that a feature only progresses when quality criteria are fully satisfied.

Based on gate scores, the orchestrator updates the project state machine into four terminal statuses:

• COMPLETED: Approved and ready for your final PR review.
• RETRY: Scores fell short; the engine compiles a REWORK-LOG.md and loops back to code automatically.
• BLOCKED: Critical crash/break—the engine triggers a circuit breaker and halts for immediate human intervention.
• FAILED: Non-blocking tech debt—the pipeline logs the issue and moves to the next feature, leaving the debt for you to audit later.

---

Installation & Quick Commands

HarnessKit is distributed as a command-line plugin compatible with major AI developer ecosystems.

⚠️ IMPORTANT! This project requires the Superpowers skill. Install it before initializing HarnessKit:

/plugin install superpowers@claude-plugins-official

Claude Code

/plugin marketplace add romabeckman/harness-kit
/plugin install harness-kit@harness-kit
/harness-kit:project-memory --help

GitHub Copilot CLI

copilot plugin marketplace add romabeckman/harness-kit
copilot plugin install harness-kit@harness-kit

Gemini CLI

# Install the extension
agy plugin install https://github.com/romabeckman/harness-kit

---

What's Inside

To prevent role contamination, the orchestrator isolates operational contexts by dispatching highly specialized agent personas equipped with dedicated skills.

🛠️ Skills (/skills)