boundary

Hey folks, is there anything special we have to do for the `allowed_role_metadata` flag for the gpt 5.6 models? From my testing, it seems those only work for claude & cache_control, but not prompt_cache_breakpoint

Thanks! If it's worth anything we're on baml-cli 0.221.0

Ahh ok, that the breakpoint has to be put on an input block, which BAML doesn't support.

From Claude:


Even with:

 ```baml
   allowed_role_metadata ["cache_control", "prompt_cache_breakpoint"]
   {{ _.role("system", cache_control={"type": "default"},
      prompt_cache_breakpoint={"mode": "explicit"}) }}
 ```

 the OpenAI Responses wire body was:

 ```json
   {
     "model": "gpt-5.6",
     "prompt_cache_options": { "mode": "explicit" },
     "input": [{
       "role": "system",
       "content": [{ "type": "input_text", "text": "Static instructions." }]
     }]
   }
 ```

 Both metadata fields are dropped. prompt_cache_options and prompt_cache_key do survive, but BAML 0.221.0 cannot render OpenAI’s content-block
 prompt_cache_breakpoint.