What is codex-skill?

codex-skill is a configuration package for OpenAI's Codex command line tool that sets up a strict way of coordinating multiple AI coding sessions, all using a model the README calls GPT 5.6 Sol. One session acts as the root planner, deciding what work needs doing, while three separate limited roles do the actual work: one that only explores and reads code, one that is allowed to write code but only within a scope it was explicitly given, and one that only reviews changes without touching files. A rare fourth mode, called an Ultra takeover, temporarily replaces the root planner with a stronger reasoning setting for unusually hard architecture or security decisions, and requires a human to type an explicit confirmation flag before it runs. What makes this project distinct is that it does not simply trust whichever AI model claims to be running. After starting a Codex session, the launcher scripts read the actual session log to confirm which model and reasoning setting were really used, and it refuses to accept the result if they do not match what was requested. Installation clones the repository and runs a setup script that links the skill into the user's global Codex configuration folder, adjusts some default settings, and installs safety hooks, without disturbing any of the user's existing unrelated Codex settings. After installing, the user must manually approve the two new hooks inside Codex before they take effect. Each coding task is started by piping written instructions into a launcher script along with a chosen role and a sandbox mode such as read-only or workspace-write. Every run prints back a single structured result showing whether it completed, what was changed, and whether the model routing was verified. A locking system prevents more than three of these tasks from running against the same files at once. The README does not mention a license.