What is claude-codex-orchestrator?

claude-codex-orchestrator is a skill for Claude Code that splits work between two AI coding tools so you spend your more expensive model's time on judgment rather than typing. The idea is simple: Claude plans, breaks a task into pieces, and checks the results, while OpenAI's Codex actually writes the code. The README frames this as sending the thinking to the model best suited for design and review, and sending the mechanical writing to a fast, flat rate tool. The workflow runs in a loop. First Claude decomposes a problem into small, self-contained work packets, each one listing its goal, how to know it succeeded, which files it touches, what it should not touch, and the exact command that proves it worked. Codex then writes the code for one packet at a time, without making its own design decisions, and hands back both a code change and a written report describing what it did. Claude does not trust that report on its own, it treats the actual code change as the real evidence, re-runs every check the report claims passed, and watches for signs like weakened tests or claimed work that was not actually done. Based on that check, Claude marks the result verified, verified with caveats, or refuted. If a packet fails its checks twice, the system reroutes it to a different, more capable setup rather than blindly retrying the same approach a third time. Codex is configured through several named profiles, each pinned to a specific underlying model and a chosen reasoning effort level, for tasks like everyday implementation work, harder or more uncertain implementation, code review, deep diagnosis of a stuck task, purely mechanical bulk edits, and getting a second opinion. Setting it up requires having Claude Code and the Codex command line tool installed and authenticated, plus an official Codex plugin for Claude Code, after which you copy the provided profile files and the skill file into your configuration folders. The README walks through several worked examples, such as building a feature from a written spec, fixing a bug that has a known way to reproduce it, doing a large mechanical rename across many files, and digging into a task that keeps failing for unclear reasons. The project explicitly separates itself from a related project that routes between different Claude model tiers, saying the two should not be mixed. It is released under the MIT license.