OpenCode struggles with local Qwen 3.5 via Ollama

Running local LLMs with complex agentic coding tools remains brittle, exposing the gap between raw model capability and reliable tool integration.

• –"Plan mode" failures suggest the model is dropping system prompts or failing to output the structured JSON required by the tool.
• –Context window management in "build mode" appears flawed, causing the model to forget previous session history.
• –Local file handling issues indicate potential mismatch between the tool's expected context format and what Ollama provides to the model.