https://shane.greaves.casa/categories/llms/ Recent content in LLMs on Guru Dude Hugo -- 0.131.0 en shane@greaves.casa (Shane Greaves) shane@greaves.casa (Shane Greaves) Mon, 02 Jun 2025 12:00:00 -0500 https://shane.greaves.casa/posts/2025-06-02-qwen-thinking-tokens-tool-calling-corruption/ Mon, 02 Jun 2025 12:00:00 -0500shane@greaves.casa (Shane Greaves) https://shane.greaves.casa/posts/2025-06-02-qwen-thinking-tokens-tool-calling-corruption/ The Symptom Cron jobs started returning this odd refusal: [System Limitation Notice] I am an AI assistant and cannot actually execute bash scripts... Except the agent can execute bash. That’s literally its job. The refusal was fake — a content filter hallucination triggered by something else going wrong deeper in the stack. The Real Problem Qwen3.5 was dumping its internal reasoning monologue into the tool-calling JSON. When calling Qwen via OpenAI-compatible /v1/chat/completions with tool definitions, the model should generate clean JSON like: