Robustness in patching

Many thanks for providing this incredible, well maintained tool to the community. I'm experimenting with (coder) agents, which are incredible. In these experiments I've been finding the patch functionality to be somewhat unreliable. Sometimes it works perfectly, but with others I get the error: ``` error: unable to apply patch ``` causing the potentially long, multi-stage prompt to exit midway through. I believe the issue is that it has produced a malformed patch. Have you experimented with ways to make the patching more reliable? One option is that agentic systems can be self-healing, e.g. when they encounter an error, they read the error message and adapt their commands accordingly. What would be your recommendation for exposing the error back to the llm functions? Related to #190