Pick fix from #359 and revert max_steps = 1.

Signed-off-by: Peter Jin <[email protected]>

Author: pjin-nvidia

responses_api_agents/simple_agent/app.py

@@ -43,8 +43,7 @@
 class SimpleAgentConfig(BaseResponsesAPIAgentConfig):
     resources_server: ResourcesServerRef
     model_server: ModelServerRef
-    # max_steps: int = None
-    max_steps: int = 1
+    max_steps: int = None
 
 
 class SimpleAgentRunRequest(BaseRunRequest):
@@ -102,6 +101,9 @@ async def responses(
             output = model_response.output
             new_outputs.extend(output)
 
+            if model_response.incomplete_details and model_response.incomplete_details.reason == "max_output_tokens":
+                break
+
             all_fn_calls: List[NeMoGymResponseFunctionToolCall] = [o for o in output if o.type == "function_call"]
             all_output_messages: List[NeMoGymResponseOutputMessage] = [
                 o for o in output if o.type == "message" and o.role == "assistant"

responses_api_models/vllm_model/app.py

@@ -132,6 +132,7 @@ async def responses(
             metadata=body.metadata,
             instructions=body.instructions,
             user=body.user,
+            incomplete_details={"reason": "max_output_tokens"} if choice.finish_reason == "length" else None,
         )
 
     async def chat_completions(