chaseai-yt/claudex-loop

What it solves

AI-assisted coding often suffers from "echo chambers" where the model that writes a plan is the same one that reviews it, leading to unbuildable subsystems or data corruption. Claudex-loop eliminates this by using two rival AI models (Claude and OpenAI Codex) to cross-examine and harden plans and code, ensuring that whoever creates an artifact never grades it.

How it works

The project implements a four-phase workflow:

1. Recon: Claude explores the codebase or researches prior art to create an Assumptions Ledger, preventing redundant questions.
2. Interrogate: Claude asks the user "load-bearing" questions one by one to lock in a PLAN.md file.
3. Review: OpenAI Codex attacks the PLAN.md in a read-only sandbox, identifying flaws. Claude arbitrates these critiques and revises the plan until Codex approves or a round limit is reached.
4. Build (Optional): One model writes the code while the other cross-inspects the final diff against the original plan.

Who it’s for

Developers using AI coding assistants who want to ensure their architectural plans are technically sound and free of hallucinations before committing to implementation.

Highlights

• Cross-Provider Validation: Uses both Claude and OpenAI Codex to prevent single-model bias.
• Read-Only Sandboxing: The reviewer model cannot modify files during the review phase.
• Decision Mapping: Distinguishes between critical "load-bearing" decisions and cosmetic ones during the interrogation phase.
• Audit Trail: Generates a PLAN-REVIEW-LOG.md documenting the full argument and revision history.