zjp1997720/codex-model-routing-team

This project is a Skill for OpenAI's Codex coding assistant that adds a way to split a big task into smaller background jobs and send each one to a specific AI model, instead of every parallel task running on whatever model the main session happens to be using. One lead agent stays in charge of planning, verifying results, and putting the final answer together, while separate background tasks do the actual research or coding work. It is meant for genuinely complex, parallelizable work, things like researching several unrelated topics at once, writing a long document in sections, or building several independent pieces of a project that do not touch the same files. The README is explicit that it should not be used for simple questions, small edits, strongly sequential work, or anything involving publishing, payments, deleting things, or account and production changes, those stay with the lead agent. The system enforces limits so it cannot spiral out of control: at most three new background tasks per wave, six running at once, and eight total for one request. The very first task created is treated as a health check to confirm the background task system is actually working before more are created, and background tasks are not allowed to spawn further background tasks of their own. Completed tasks are reviewed and archived one at a time only if their results were actually used. Installing it uses a command line tool called skills, either globally for all Codex projects or scoped to one project, and the README shows how to verify the install copied all its supporting files correctly. It can be triggered explicitly by name in a request, or a user can add a standing authorization note to their own Codex configuration file so it activates automatically for suitable complex tasks, though the README notes this is a user configured instruction, not a hidden default behavior. The project is released under the MIT License.