LM Studio users hit tool call loops

Local agentic workflows remain brittle, as the gap between reasoning-heavy models and structured parsers creates "infinite thought" loops.

• –The "reasoning trap" in models like Qwen 2.5 causes LM Studio to misinterpret `<think>` blocks as live tool calls, triggering recursive errors.
• –Small context windows (under 32k) lead to "amnesia," where the model forgets it has already received the tool output and re-runs the request.
• –High repeat penalties (1.2+) can paradoxically worsen loops by forcing the model to hallucinate new calls to avoid repeating "final answer" tokens.
• –Using deterministic models like IBM Granite 4H-Small or Llama 3.1-Instruct with a 0.0 temperature is currently the only reliable fix.