I've never run into that problem, but these were coding solutions in codex with a strong plan, steps to work towards. It could be that if you're using massive

I've never run into that problem, but these were coding solutions in codex with a strong plan, steps to work towards. It could be that if you're using massive tokens on a 'plan' then then want to limit u in a way, or even if the objective is not perfectly clear they don't want semi-random token use. See if the token/sub solution behaves differently. Make sure that when it 'compacts' that it re-reads your instructions clearly.