@AIMFllyYS

DeepSeek Harness settings page: cap every configured model to a working context window, clamped to what it can actually hold.