wtf is codex-advisor?

Codex Advisor is a routing tool for the Codex coding assistant that tries to give each task to the cheapest model that can actually handle it, instead of always calling the biggest one. It defines a small set of roles: Luna searches the repository and gathers evidence, Sol makes architecture decisions and turns approved evidence into a plan, and Terra implements one narrow, well defined slice of work and runs its own checks. A root agent still keeps control of scope, verification, and final decisions, and can skip the other roles entirely if it already has what it needs. The project is built around Codex CLI and runs on macOS or Ubuntu, with Bash, Git, Python 3, Perl, jq, and rg installed alongside an authenticated Codex CLI session. Windows is not supported yet. Installation is a git clone followed by a setup script, which can either link the checkout as the active install so a later git pull updates it, or copy it as a separate install. Uninstalling means removing the skill folder and four agent configuration files. To check whether the routing actually works, the project includes two types of tests. A deterministic test suite spins up a temporary Git repository and runs each role through a fake Codex binary, checking things like model selection, sandboxing, timeouts, cleanup, and write permissions. A separate live test suite asks the real Codex to route eighteen different tasks without being told the expected answer, covering simple fixes, messy changes that touch multiple parts of a codebase, security related edits, and cases designed to tempt it into using too many or too few models. The project states that a release needs at least ninety percent exact routing accuracy with no critical failures, and the checked in results show one miss out of sixteen cases in the full run. The README is direct about the limits of what this project protects against. It changes what an agent is allowed to do, not whether the underlying repository is safe to run code from. Users are advised to only point it at trusted repositories, since source files can contain hidden instructions, and to know that a sandbox limits authority rather than hiding local files from the process entirely. The project is released under the MIT license and is not affiliated with or endorsed by OpenAI, even though it works alongside their Codex tool.