Qwen 3.5 tool calling needs client-side safety nets

Qwen 3.5 is brilliant but its "plumbing" is currently broken in almost every major local inference engine.

• –XML tool calls frequently leak as plain text or get buried inside thinking blocks, causing agent loops to hang or crash.
• –Stock Jinja templates fail on basic argument filtering; switching to "barubary-attuned" or Unsloth templates is a non-negotiable requirement for stability.
• –Servers like llama.cpp suffer from "thinking leaks" where internal reasoning tags poison multi-turn context, requiring aggressive client-side stripping.
• –LM Studio v0.4.9 currently leads the pack by natively handling Qwen's specific parsing quirks that vLLM and Ollama still struggle with.
• –The model's reasoning capabilities are elite, but developers must treat it as a "raw" output source rather than relying on server-provided JSON fields.