`claudex-loop`
The four-phase entry point that handles recon, interrogation, review, and optional build.
Plan hardening workflow
Two AI models harden the plan before the build.
Claudex Loop closes both failure gaps in AI-assisted coding: alignment with the human, then adversarial review by a rival model. Claude and Codex swap jobs so whoever built the thing never grades it.
Phase 0 · Recon
Claude scouts code, docs, prior art, and known pitfalls before the interview starts.
Phase 1 · Interrogate
Only load-bearing decisions are surfaced, each with a recommendation and risk.
Phase 2 · Review
Codex attacks the locked plan in a read-only sandbox until it is approved or deadlocked.
Phase 3 · Build
The builder and reviewer swap jobs so the final code is always judged by the rival model.
Skill set
What ships in this repo
Static Pages root for discovery, with the plugin and skill content unchanged behind it.
`codex-review`
A standalone adversarial plan review loop when the plan already exists and only needs stress-testing.
`codex-build`
Codex implements a frozen spec while Claude reads the diff, runs proof, and arbitrates fixes.
Receipts
What a real run looks like
The first end-to-end run converged from 26 findings down to zero across five review rounds, surfacing one fatal architecture issue and multiple data-corrupting design flaws before any code shipped.
- 55 findings across 5 rounds.
- 1 fatal unbuildable subsystem caught before implementation.
- Several missing subsystems and wrong data models surfaced early.
