Qwen3.6 35B 搭配正確的程式碼框架，讓我的本地設置在真實 Go 任務上達到 9/10 的成績

I wanted to test a slightly different question than "can one open model beat GPT-5.4 Codex?" The question was: Can a combination of local models, scaffolding, repair loops, and routing policies running on home hardware get close enough to frontier coding models on my actual workload? Short version: yes, surprisingly. On my first curated 10-task Go eval set, a routed local process got to 9/10 passing tests. Links: - little-coder: https://github.com/itayinbarr/little-coder - The write-up tha