>>fugu2+(OP)
Maybe you can log all the traffic to and from the proprietary models and fine tune a local model each weekend? It's probably against their terms of service, but it's not like they care where their training data comes from anyway.
Local models are relatively small, it seems wasteful to try and keep them as generalists. Fine tuning on your specific coding should make for better use of their limited parameter count.