An interesting combination, indeed. Running local models through Claude Code opens up many possibilities for teams that can't push code to external APIs.
However, I've noticed something with local LLMs: they're much more prone to context window issues than hosted models. A 7KB model with an 8KB context fills up quickly when reading files, and once it overflows, it starts displaying hallucinations in the function signatures. Have you encountered this problem with Unsloth models? If so, how do you handle context management?
