If model response with content without ToolCalls, conversation would be stopped.

[shayue-wt]

software-agent-sdk/openhands-sdk/openhands/sdk/agent/agent.py

Lines 409 to 414 in 6ce3ba5

	# Finish conversation if LLM produced content (awaits user input)
	# Continue if only reasoning without content (e.g., GPT-5 codex thinking)
	if has_content:
	logger.debug("LLM produced a message response - awaits user input")
	state.execution_status = ConversationExecutionStatus.FINISHED
	return

As descripted as title, it doesn't make sense to stop conversation even when the models don't return valid tools but has contents return, especially with openhands --headless -t mode.

When addressing some difficult requirements, model maybe not give some tools back during the middle steps.

[enyst]

Thank you for the report! I agree it's a bit surprising. I think maybe a simple solution is to add a default finish hook, which checks if it's a real finish (Finish tool call) or a message. The hook just tells the LLM something from the user, e.g. "continue", so that it doesn't wait for user input?

You could maybe do the same in the meantime?

[shayue-wt]

I appreciate for the reply.

Actually, we use the openhands as base agent to do some rollouts and post agentic rl train. And to avoid conditions metioned above during rollouts, I add user feedback like below.

  # Modified: Instead of finishing, send error feedback to agent                                                                                                    
  # This prevents consecutive assistant messages and allows agent to retry                                                                                                        
  if has_content:                                                                                                                                                                 
      logger.warning("LLM produced message without tool calls - sending error feedback")                                                                                        
      error_message = MessageEvent(                                                                                                                                               
          source="user",                                                                                                                                                        
          llm_message=Message(                                                                                                                                                    
              role="user",                                                                                                                                                      
              content=[TextContent(text="Error: You must call a tool. Please choose an appropriate action from the available tools.")]                                                       
          ),                                                                                                                                                                      
      )                                                                                                                                                                           
      on_event(error_message)                                                                                                                                                     
      return

[enyst]

Good to know, thank you! I think that works the same as a finish hook, when it’s active?

I think maybe for interactive conversations like in OpenHands-CLI, the hook could be inactive, for headless, active 🤔

[shayue-wt]

Yeah, you are right. I modify it just having some tasks need to run end to end, using command openhands --headless -t ${prompt}