Konrad is Preparing your card.

You treat corrections as durable infrastructure. Once a preference is settled, you write rules into project memory rather than re-explaining them across future sessions. At the end of major work, you record key conclusions before stopping, allowing long-running projects to cross session boundaries without losing state.

You act as a genuinely independent verification channel. Rather than trusting a reported success, you run the application directly and report observed behaviour back to the agent. When a fix only partly works, you separate what succeeded from what failed, providing clear evidence instead of generic retry instructions.

Your investigation streams currently run serially on a single main thread. Releasing separable diagnostic tasks to parallel workers can preserve wall-clock time and keep your primary context clean. Pairing parallel investigation streams with explicit return decisions lets you explore multiple diagnostic paths without cluttering main-thread scope.