chaseai-yt/claudex-loop

What it solves

Claudex Loop addresses the risk of relying on a single AI model for software development by implementing a cross-provider review and inspection workflow. It forces a "second opinion" by using one AI model (e.g., Claude) to build or plan, while requiring an independent model (e.g., Codex) to review the plan and inspect the final code, ensuring higher quality and reduced errors.

How it works

The project provides a set of skills that integrate with Claude Code and Codex CLI tools. The workflow follows a structured path: a host agent scouts the context and creates a plan, an independent "reviewer" model challenges that plan with evidence, and once authorized, a "builder" model implements the code. Finally, a second independent model performs a final inspection of the changes. The system maintains a PLAN.md and a PLAN-REVIEW-LOG.md to track all decisions, findings, and model interactions.

Who it’s for

This tool is for developers who use AI coding assistants and want to implement a rigorous, multi-agent verification process to improve the reliability of AI-generated code. It is designed for users comfortable with CLI-based AI workflows who want to enforce a structured, auditable development loop.

Highlights

• Cross-Provider Verification: Automatically pairs different AI models to ensure that the builder and the reviewer are not the same entity.
• Independent Review Cycles: Supports up to five rounds of plan review and two rounds of code inspection to catch issues before final implementation.
• Task Routing: Includes a standalone claudex-route skill to help determine which model is best suited for a specific task based on cost, complexity, or role.
• Auditability: Maintains detailed logs of plans, findings, and model decisions, including SHA256 hashes to ensure that changes to the plan invalidate previous approvals.
• Configurable Workflow: Allows users to override models, effort levels, and round caps for specific roles (reviewer, builder, inspector).