Qwen tool calls trip Continue

This is a useful reminder that “supports tool calling” is not a portable guarantee across local inference stacks; the model, chat template, server parser, and client all have to agree on the same protocol.

• –Qwen3.5-9B advertises native function calling, but serving guides typically require a Qwen-aware parser or chat template rather than generic JSON-only tool handling
• –Continue.dev detecting the model and producing text does not prove tool support is wired correctly through Lemonade’s OpenAI-compatible API surface
• –Reports of hallucinated responses in Roo Code and Kilo Code point to a deeper prompt/template mismatch, not just one Continue configuration flag
• –For local coding agents, model variants fine-tuned for tool use still need harness-level compatibility; QwenCode, Aider, OpenCode, or a server with explicit Qwen parser support may be more reliable
• –The broader local-agent lesson is that small open models are getting capable enough, but tool-call interoperability remains brittle and under-documented