Has anyone else experienced Gemini in Firebase Studio losing track of the current instruction after a long coding conversation?
I have noticed that after several successful small changes in the same chat, Gemini sometimes starts carrying old requirements into a completely new request. Even when the latest prompt clearly says not to touch certain areas, it may continue modifying them, introduce duplicate structures, rewrite working logic, or make assumptions that were never requested. What makes it more confusing is that the model can acknowledge the scope correctly in one response and then still behave differently in the next turn.
I have also tried keeping completion/context notes to clarify what has already been finished and what should not be touched, but the effect seems inconsistent. Sometimes they help, sometimes the model appears to prioritize older conversation context over the latest instruction.
This becomes especially frustrating in larger projects because a simple change can turn into unnecessary refactoring or changes across unrelated parts of the codebase. At that point, even repeatedly telling it to stop expanding the scope does not always prevent the behavior.
Has anyone found a reliable way to make Gemini treat the latest prompt as the current scope while keeping useful conversation context? I am mainly interested in understanding whether this is a context-management/instruction-priority issue in Firebase Studio, or whether there is a recommended workflow for long-running coding chats that avoids this behavior.