I think you are spot on. Also if you are actively listening you may learn that the fault is in how you are communicating your ideas instead of the ideas themselves being flawed
Is there a local harness designed around the local model use case that is claude code like? Opencode has been problematic at times, pi works for one off for me but not back and forth conversations with the LLM. Considering I only use Qwen or Gemma models I'm close to just writing my own at this point