Qwen3.6-35B-A3B stalls inside agent wrappers

My read: this is more likely a wrapper mismatch than a broken model. Qwen3.6 is explicitly aimed at agentic coding, but its official docs assume specific reasoning and tool-call parsers that local agent shells may not be wired for by default.

• –Qwen’s docs show Qwen3.6 defaults to thinking mode and recommend `reasoning-parser qwen3` plus `tool-call-parser qwen3_coder` for tool use, which means a generic agent client can stall even when plain chat works.
• –Context is worth checking, but it is probably not the root cause: the model is natively 262K tokens, and Qwen says 128K+ helps preserve thinking behavior. A 4K default would be bad for agents, but not the only plausible failure mode.
• –Qwen also documents a non-thinking path via `chat_template_kwargs.enable_thinking: false`, so some agent workflows may need thinking disabled or preserved explicitly rather than left to defaults.
• –The practical takeaway is that “works in `ollama run`” does not prove agent compatibility; tool protocols, chat templates, and wrapper support matter more than raw generation speed here.