Are Tool Calling broken for some of you?

[BradKML]

Would love to check on this since it is known that Claude Code has a hard time with the newer models having custom API calling formats (yes this includes Harmony, Anthropic, and Qwen3-Coder). Right now we know GLM-4.5 variants and Qwen3-Instruct at least can play fair (the same applies for DeepSeek v3.1 updates and R1 but not the original v3, so the opposite of Alibaba's situation)

Ideally we should have a defaulty setup that mixes small models (cheaper and ideally Ollama-able on consumer GPU) with large models (OpenRouter or some other provider but should be as economical as possible)

Chats elsewhere UfoMiao/zcf#31 (reply in thread) https://gist.github.com/WolframRavenwolf/0ee85a65b10e1a442e4bf65f848d6b01?permalink_comment_id=5733908#gistcomment-5733908