harness

Automate iterative implementation review using an adversarial generator-evaluator loop.

3|Updated Mar 1, 2026
One-click install
npx skills add https://github.com/onblueroses/strata --skill harness-onblueroses
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: harness
Source: https://github.com/onblueroses/strata/tree/main/skills/harness
Command: npx skills add https://github.com/onblueroses/strata --skill harness-onblueroses

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires spec, evaluator, and includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

This Skill automates the iterative review and refinement of high-stakes implementation tasks using an adversarial generator-evaluator loop, reducing manual errors and improving consistency.

Core Features & Use Cases

  • Adversarial Generator-Evaluator Loop: Automates the generation of implementation candidates and their evaluation by a different-model-family evaluator.
  • Rotating Framings: Uses rotating adversarial framings to address different blindspots in each iteration.
  • Convergence-Based Termination: Stops when Stage A passes and Stage B has no CRITICAL findings, or a convergence signal fires.
  • Competitive Mode: Spawns multiple generators with different strategies per iteration, selecting the best candidate before evaluation.
  • Use Case: Suitable for high-stakes tasks like security audits, critical bug fixes, or complex feature implementations where multiple perspectives and iterative refinement are required.

Quick Start

Run the harness with the current spec as the criteria and a specific framing.

Frequently Asked Questions about harness

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I automate iterative code review for high-stakes implementation tasks?

Automate iterative code review by running an adversarial generator-evaluator loop that generates implementation candidates and evaluates them using a different-model-family evaluator to refine critical development tasks.

What is an adversarial generator-evaluator loop for implementation refinement?

An adversarial generator-evaluator loop is an automated process that generates implementation candidates and subjects them to evaluation by a different model family, using rotating adversarial framings to address different blindspots in each iteration.

When should I use an automated iterative review loop for critical bug fixes?

Use an automated iterative review loop for critical bug fixes, security audits, or complex feature implementations where multiple perspectives and iterative refinement are required to reduce manual errors and improve consistency.

Do I need a specific evaluator to run the harness for implementation review?

Yes, running the harness requires a compatible evaluator and a spec to define the criteria, enabling the system to automate the iterative review and refinement of your implementation tasks.

How does the adversarial loop know when to stop iterating during code refinement?

The adversarial loop uses convergence-based termination, stopping automatically when Stage A passes and Stage B has no CRITICAL findings, or when a specific convergence signal fires during the iterative review.

Can I use multiple generators with different strategies for implementation review?

Yes, competitive mode spawns multiple generators with different strategies per iteration, selecting the best candidate before evaluation to ensure high-quality implementation outputs for critical development tasks.