Hidden response support

<think> text </think> exists but that’s not stored in context. You can prompt models to use those tags and determine what’s inside them even if they aren’t reasoning models. I’d like a feature that supports something like <hidden> text </hidden> where everything there is stored inside a “Hidden” draw at the top of a response which can be accessed by users. The difference being that everything in the Hidden draw is part of the context llm’s deal with.
This is to support some prompt instructions such as tracking rpg like systems or inventories. Generally this system needs to not cause issues with reasoning models, perhaps allowing for two draws, one for “Thinking“ and the other for “Hidden“.