Qwen3.6 Tool Calls Fray in n8n

This looks like a harness problem masquerading as a model problem. Qwen’s own docs emphasize tool-use support and recommend specific server-side tool-call parsers, which is exactly where n8n integrations tend to break.

• –n8n wants structured OpenAI-style `tool_calls`; raw `<tool_call>` XML or chat text will often be ignored.
• –Qwen3.6’s official model card says tool use is supported and shows explicit `--tool-call-parser qwen3_coder` setup for serving stacks like vLLM and SGLang.
• –Roo Code and OpenWebUI can hide a lot of backend mistakes because they include client-side parsing logic that n8n does not.
• –If llama.cpp is emitting the wrong chat template or a non-OpenAI tool format, the fix is usually in the server config, not in the workflow.
• –For local agent setups, the practical test is the raw `/v1/chat/completions` response: if it does not return structured `tool_calls`, n8n will stay flaky.