Thanks for the reply. In response to your points:

1. The models do support tool calling and I'm not claiming the prompt is preventing them from calling tools categorically. They do, however, fail significant percentages of tool calls, especially gpt-oss-120b (see original post).
2. The behavior persists with the tool_choice_mode set to required.
3. The tool call failures are specific to the gpt-oss models. I have attached a chart comparing the tool call error counts for the 5 models we tested across ~48,000 API calls.
4. This behavior was observed in isolated API calls outside of an agentic loop.

I will speak to our Databricks rep to file a ticket.