Tokenization is fragile, especially in agent loops. Do the correct thing and focus on append-only, token-in token-out. This will save you countless times from silently wrong training runs.