I thought the local LLM was weak — it was a bug in my own harness

I've been benchmarking local LLMs as coding agents inside my own IDE (Teaspoon IDE) on a CPU-only...