The OpenAI Codex provider accepts reasoning_effort but silently discards it. Wire it through as {"reasoning": {"effort": ...}} in the request body so the config option actually takes effect.
