[Feature Request] Improve model reasoning consistency and memory adherence in Claude Code

Bug Description
Claude 4.7 is not usable for paid work in its current state. 17+ rounds of bot reviews on a single PR, constantly deviates from memory.md, no longer anticipates the obvious. In general, compared to 4.6, reasoning of your current model did not graduate. It's bad enough that I'm actually considering a fresh review of your competitors' capabilities to determine if better fit. While I understand that you need to manage capacity, this result is absolutely a regression/not satisfactory. Also, respectfully requesting consideration for credit of tokens that have been unnecessarily utilized by the new model.