Tool calls fail on reasoning deployments that refuse function tools unless reasoning_effort is "none" #1353

What happens

On an Azure OpenAI reasoning deployment, every request that carries tools comes back 400 ("Function tools with reasoning_effort are not supported for ") unless the body sets reasoning_effort: "none". Chat then falls back to a single retrieval step, and the governance second look, which needs tools, cannot run, so the pairs it would settle go to people instead.

What exists

#1232 learns max_completion_tokens and drops temperature from a single 400. Main also has a reasoning_effort setting, but with_reasoning_effort puts it into every chat body (crates/utopia-llm/src/lib.rs), so making tool calls work means turning reasoning off for extraction and alignment too.

Proposal

Learn it the way #1232 does: when a request with tools gets that 400, retry with reasoning_effort: "none" and remember it for tool calls only (chat_tools and chat_tools_stream), leaving the other calls as configured. Splitting the setting by call kind (extraction, alignment, tool calls) would also work.

Workaround we run

A small proxy in front of the endpoint that sets reasoning_effort: "none" when tools is non-empty.

Seen on v0.1.0-rc9; checked against main on 2026-10-21.