I read all 22 GB of my Codex conversations byte by byte. 71% was screenshots

Using the vision feature feels so token bloated to me. I generally delegate that to a fresh context if possible.