OpenCode hits tool-calling walls with local Ollama models

The "last mile" of local LLM orchestration remains fragile, as even top-tier models like Qwen2.5-Coder struggle with the rigid JSON schemas required by terminal agents.

• –The issue often stems from Ollama's default context window, which truncates the long system prompts and tool definitions required by OpenCode's agentic logic.
• –Mismatched tool formats, such as capitalization errors in JSON keys, cause models to output raw text instead of triggering the terminal agent's execution engine.
• –Developers are increasingly forced into manual Modelfile configurations to bypass API limitations that fail to pass context parameters dynamically to local inference servers.
• –While VRAM is abundant on modern GPUs like the 7900XT, the software bridge between inference servers and agentic frameworks remains the primary bottleneck for local autonomy.