Check your CLAUDE.md, etc. for any of the banned topics or anything adjacent. Also check your recent Git commit history, if you're running the agent in a repository, as well. The classifier trips on anything anywhere in context, so those could be sources of rejections.
adding rules for the agent to ignore version control and never touch git files or commands (except for git log) improve my slop coding 100%. no risk of sneak commits and much less token wasted.