fix(metadata): give gpt-5-mini reasoning headroom for tagging/validation succeed

src/helpers/metadata_llm.py

@@ -32,5 +32,6 @@ def get_metadata_model_response(prompt: str, max_output_tokens: int) -> str:
     return _get_metadata_model_run().get_response(
         prompt=prompt,
         max_output_tokens=max_output_tokens,
+        reasoning={"effort": "minimal"},
         safety_identifier=get_openai_safety_identifier(),
     )

src/metadata/tag_questions/main.py

@@ -34,7 +34,7 @@ async def _get_category_single(index, row, semaphore):
             response = await asyncio.to_thread(
                 metadata_llm.get_metadata_model_response,
                 prompt=prompt,
-                max_output_tokens=50,
+                max_output_tokens=512,
             )
             category = response.strip('"').strip("'").strip(" ").strip(".")
             logger.info(

src/metadata/validate_questions/main.py

@@ -32,7 +32,7 @@ async def _validate_single_question(index, question, semaphore):
             response = await asyncio.to_thread(
                 metadata_llm.get_metadata_model_response,
                 prompt=prompt,
-                max_output_tokens=500,
+                max_output_tokens=1024,
             )
             if "Classification:" not in response:
                 logger.error(f"'Classification:' not in response for: {question}")

src/tests/test_model_request_params.py

@@ -138,6 +138,7 @@ def fake_get_model_run(model_run_key):
             "Classify this question.",
             {
                 "max_output_tokens": 123,
+                "reasoning": {"effort": "minimal"},
                 "safety_identifier": "forecastbench-safety-id",
             },
         ),