It is impossible to get an agent to not use the ask_user tool for Q/A style chats.

Example:
"/grilling we need do a minor addition to this feature branch. look at the specs so far to see where we are on the change management additions for auto generating users. What I want to do is introduce gating for users with a long username.
DO NOT USE the ask_user tool."

(/grilling is a q and a format for defining a spec)

This results in the ask_user tool being called.

I don't want the Q and A format because it is restrictive in many ways. For example, it will not ask bulk questions like the skill specifies. It will also not allow uploading of pictures or documents. No matter what I tell the agent (GPT-6 Astra in this case) it will not stop using the ask_user UI.

This is after many tries to restart and convince the agent that this is possible.

Start with the /grilling Q&A flow and the ask_user tool behavior described in the report. Reproduce the example while explicitly forbidding ask_user, then trace why the agent continues invoking that UI; done means the instruction is respected without preventing ordinary chat, bulk questions, or file and image uploads.