How to Configure Luna Max as a Cost‑Efficient Sub‑Agent in Codex

Tip for Codex: ask Sol to configure Luna Max as a sub‑agent.

In the DeepSWE benchmark, Luna Max scored 67% versus 73% for Sol Max. At the same time, the average cost of executing a single task turned out to be almost 14 times lower: $0.61 versus $8.39. Therefore, it makes perfect sense to delegate individual well‑defined tasks to Luna.

Prompt:

Создай кастомного агента luna_worker в файле ~/.codex/agents/luna-worker.toml.
Используй следующие настройки:
model = "gpt-5.6-luna"
model_reasoning_effort = "max"
Добавь описание и инструкции для выполнения чётко ограниченных делегированных задач. Остальную конфигурацию не меняй. Проверь совместимость с установленной у меня версией Codex, покажи diff, а затем используй luna_worker для задач субагента.

Another option is to do without a custom agent file. Keep Sol in the main thread and ask it to launch a separate Luna Max thread for each task.

In Codex, a thread is a separate dialog with its own context. Multiple threads can run in parallel, and Sol collects their results in the main thread.