If you properly keep documents, architecture, and decision records, token consumption can be pretty less. Iam managing everything with two codex plus sub. Repo size is 300 k loc ( backend).