“OpenAI's Codex is designed for long-running jobs by periodically running a context compaction process after each turn, which keeps the context window from filling up.”