How do you manage context size and your coding harness?
Hey folks!
I'm on a 48 GB M5 Pro MacBook that I picked up a couple months ago and I've been trying to get into local agentic coding. At work I'm fortunate enough to use frontier models within VSCode's Copilot (no Claude Code, Codex, etc), and it's so easy there to never worry about context size, manual compaction, etc.
On the Mac I've tried a few couple harnesses with oMLX, namely Claude Code and OpenCode, but I can't quite figure out the right workflow yet. For example I'll run into situations where part way through a session the prefill OOM guard kicks in. I've been using Qwen 3.6 35B A3B oq4 and a 65K context window, which I thought should be manageable with my 48 GB RAM. With nothing loaded and all my apps closed activity monitor shows roughly 16 GB usage, seems excessive, but I can't figure out what other system stuff I can get rid of to leave more room for the model + context.
I know I can keep turning down the context window, use a smaller model, etc., but it just feels like I'm missing something... I'd like to know immediately when I load a model with a given context size if it'll eventually hit the OOM guard or not.
I suppose I don't have a clear question, but I've been reading through this sub for a bit and still nothing has quite landed well for me. Any additional tips?