OpenCode local agents drift on small models

This is less an OpenCode bug report than a broader local-agent reality check. OpenCode's local support is real, but the ceiling is still model quality and loop design, not whether you can squeeze a model onto an M4.

• –OpenCode officially supports local backends like LM Studio, Ollama, and llama.cpp, so the plumbing exists.
• –Its docs say only a few models are good at both code generation and tool calling, and recommend far stronger models than a 7B local coder.
• –The community advice matches the failure mode: short steps, resets, and external checks are the difference between useful and spiraling.
• –Consumer hardware changes speed and context headroom, but it does not fix a model that is already bad at tool use.
• –For real local agent work, think bounded assistant first, autonomous coder second.