Claudex Loop

Why it's ranked here

The cross-checking is enforced in code, not left to prompt wording. Approval is bound to the plan's SHA256 and invalidated on change, inspections fingerprint staged, unstaged, deleted and untracked files, and the adapter refuses results from a different resumed session. Reviewers run read-only: Claude with only file read and search tools and no MCP, Codex in its read-only sandbox. A unit suite with fake CLIs and throwaway Git repositories covers these contracts, with CI configured on Windows, macOS and Linux. MIT licensed, no pip packages or extra API keys at runtime.

What's good

Failure handling is explicit. An empty output file, a session-start event alone, a timeout or a malformed reply never counts as approval, and an approval cannot carry unresolved high or medium findings. A timeout kills the whole process tree. Requested and observed model names are logged separately, and there is no silent fallback to another model or provider. The validation record is candid: its live checks seeded one obvious data-loss defect and it states plainly that this says nothing about comparative model quality. Rounds are capped, five for plan review and two each for fixes and inspections.

Tradeoffs

The full loop needs both the Claude Code and Codex CLIs installed and logged in, plus Python 3.10 or newer, so you need working accounts with both vendors. Live model tests were run on Windows only; other platforms rely on the fake-CLI suite. The Codex reviewer's sandbox does not restrain external MCP tools already configured, so you must audit those yourself. Delegated builds require a clean checkout. The docs also note that a structured verdict cannot prove a model's findings or claimed coverage are true, and the companion compatibility commands break if copied without the main skill.

How to use it well

Reach for it on changes big enough to deserve a written plan: migrations, data-handling code, anything where a second opinion before building is cheaper than a rollback. Start in whichever agent you prefer as planner, set a proof command up front so the host can verify the build itself, and keep diagnostics out of the repository. For a quick second opinion or a single delegated task, the lighter routing skill needs no Python. It is not for trivial edits, and it does not replace your own review of the final diff.

Observed

License

MIT (LICENSE)

Language

Python 3.10+, standard library only (skills/claudex-loop/scripts/runner.py)

Packaging

Claude Code plugin manifest plus manual skill copy for Codex or Claude Code (.claude-plugin/plugin.json, README.md)

Interfaces

Agent skills (claudex-loop, codex-review, codex-build, standalone routing skill) driving the claude and codex CLIs

Reviewer isolation

Claude reviewer limited to Read, Glob, Grep with no MCP; Codex reviewer in read-only sandbox (skills/claudex-loop/references/runtime.md)

Approval binding

Approval recorded against plan path and SHA256; plan change invalidates it (README.md)

Tests

unittest suite using fake CLI processes and disposable Git repositories, no model calls (tests/test_runner.py)

Platform support

CI configured for Windows, macOS and Linux; live model checks run on Windows (VALIDATION.md)