I've been thinking, with the right harnesses, opencode deepseek sidex could be more powerful than Claude code, LLM gaps always boil down to bad memory or poor training data which can be fixed by examples, a simple pre-prompt that measures complexity and inserts example refs