LM Studio, OpenCode hit local context wall

This is less a product announcement than a reality check for anyone trying to run serious coding agents fully offline on consumer hardware.

• –Commenters say OpenCode can consume roughly 12K tokens before the model even starts working, so an 18K context window leaves almost no room for repo state or tool traces.
• –The looping behavior likely comes from repeated context overflow rather than a simple LM Studio bug, which makes prompt budgeting and task chunking critical for local agents.
• –A bigger machine helps, but the thread's practical advice is to first test higher context settings, smaller tasks, or lighter models before jumping to a 64GB Mac Studio.
• –The discussion highlights a broader local-AI tradeoff: privacy-first offline setups are feasible, but agentic coding workloads still demand far more context and memory than basic chat.