Claudex Loop Guide: Claude Code and Codex Cross-Review

*This research-based guide covers the open-source chaseai-yt/claudex-loop project as documented on September 17, 2026. The project is community software, not an Anthropic or OpenAI product, and DiscoverAI did not independently benchmark defect detection or token cost. The official project name is styled “Claudex Loop”; some searches render it as “ClaudeX Loop.”*

Claudex Loop is a community-built orchestration skill that connects Claude Code and OpenAI Codex across four stages: reconnaissance, requirements, adversarial plan review, and an authorized build followed by independent inspection. The model that builds should not grade its own work, while the host arbitrates findings and keeps the cycle within configured round limits.

Choose it when a software change is important enough to justify structured planning and a second provider's review. Skip it for tiny, reversible edits where setup, duplicated context, and review cost exceed the likely benefit.

The current repository describes a host-neutral workflow rather than a single hard-coded Claude-builds/Codex-reviews sequence. The current host scouts the code and context, resolves material requirements, writes a plan with observable acceptance checks, and asks the other provider to challenge it. After authorization, a selected builder implements the plan, proof checks run, and the other provider inspects the final code in a fresh session.

The plan and an append-only review log preserve findings, dispositions, model choices, proof, and remaining uncertainty. The documented defaults bound plan review, fix, and inspection rounds rather than promising an endless autonomous cycle.

1. Reconnaissance

Inspect the repository, relevant documentation, dependencies, and greenfield assumptions. Surface an assumptions ledger rather than quietly filling material gaps.

2. Requirements and acceptance checks

Resolve questions that would change the outcome. The plan should name files or systems in scope, user-visible behavior, constraints, proof commands, and decisions that still need human authority.

3. Adversarial plan review

The opposite provider examines the plan and relevant evidence. The host accepts, rejects, or incorporates findings instead of treating every model comment as truth. Review ends at an explicit verdict or the configured round cap.

4. Build, proof, and inspection

An authorized builder implements the locked plan. Deterministic checks run independently, and the non-building provider inspects the final changes. Remaining findings are reported rather than hidden behind a consensus label.

The repository says the full workflow requires authenticated Claude Code and Codex CLIs plus Python 3.10 or newer. It supports Claude Code plugin installation and manual skill installation for both hosts. Model access, rate limits, platform subscriptions, and supported command behavior can change, so verify the repository's current runtime documentation before installation.

Treat third-party skills and installers as executable code. Inspect the repository and release you intend to install, understand which commands it can invoke, avoid secrets in the working tree, and start in a disposable branch or repository.

Its strongest use case is plan risk. A second model may notice missing migrations, unsafe data assumptions, rollback gaps, authorization errors, or acceptance checks that cannot prove the promised outcome. The persistent review log also makes the reasoning easier for a human to audit than a long, unstructured chat.

Cross-provider diversity is useful but not magical. Claude and Codex can share blind spots, misunderstand the same requirement, or agree on code that fails in production. Tests, types, linting, security tooling, real data constraints, observability, and human review remain the evidence layer.

Each handoff repeats repository context and consumes tokens. Large plans or broad codebases can make multiple review rounds expensive and slow. Reviews may also produce contradictory preferences, churn a correct plan, or prioritize theoretical concerns over the user's actual deadline.

Set a task-specific round budget, narrow the files and evidence each reviewer needs, and rank findings by user impact. Never let model consensus grant new authority: publishing, committing, deleting data, changing infrastructure, sending messages, or spending money should still follow the user's explicit instructions.

Claudex Loop is a disciplined option for consequential coding work where plan quality matters and both CLIs are already available. Its advantage over a casual “ask another model” workflow is the explicit lifecycle, evidence log, bounded rounds, and independent final inspection. The payoff must be measured in defects caught and rework avoided—not in how many agents participated.