Fix GPT-5 codex empty patches

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

@@ -216,6 +216,17 @@ def step(
         # LLMResponse already contains the converted message and metrics snapshot
         message: Message = llm_response.message
 
+        # Check if this is a reasoning-only response (e.g., from reasoning models)
+        # or a message-only response without tool calls
+        has_reasoning = (
+            message.responses_reasoning_item is not None
+            or message.reasoning_content is not None
+            or (message.thinking_blocks and len(message.thinking_blocks) > 0)
+        )
+        has_content = any(
+            isinstance(c, TextContent) and c.text.strip() for c in message.content
+        )
+
         if message.tool_calls and len(message.tool_calls) > 0:
             if not all(isinstance(c, TextContent) for c in message.content):
                 logger.warning(
@@ -254,16 +265,18 @@ def step(
 
             if action_events:
                 self._execute_actions(conversation, action_events, on_event)
+            return
 
-        else:
-            logger.debug("LLM produced a message response - awaits user input")
-            state.execution_status = ConversationExecutionStatus.FINISHED
-            msg_event = MessageEvent(
-                source="agent",
-                llm_message=message,
-                llm_response_id=llm_response.id,
-            )
-            on_event(msg_event)
+        # No tool calls - emit message event for reasoning or content responses
+        if not has_reasoning and not has_content:
+            logger.warning("LLM produced empty response - continuing agent loop")
+
+        msg_event = MessageEvent(
+            source="agent",
+            llm_message=message,
+            llm_response_id=llm_response.id,
+        )
+        on_event(msg_event)
 
         # If using VLLM, we can get the raw prompt and response tokens
         # that can be useful for RL training.

openhands-sdk/openhands/sdk/llm/message.py

@@ -458,15 +458,10 @@ def to_responses_dict(self, *, vision_enabled: bool) -> list[dict[str, Any]]:
             for c in self.content:
                 if isinstance(c, TextContent) and c.text:
                     content_items.append({"type": "output_text", "text": c.text})
-            if content_items:
-                items.append(
-                    {
-                        "type": "message",
-                        "role": "assistant",
-                        "content": content_items,
-                    }
-                )
+
             # Include prior turn's reasoning item exactly as received (if any)
+            # Note: Reasoning item must come BEFORE message/tool_calls for proper
+            # ordering
             if self.responses_reasoning_item is not None:
                 ri = self.responses_reasoning_item
                 # Only send back if we have an id; required by the param schema
@@ -491,6 +486,17 @@ def to_responses_dict(self, *, vision_enabled: bool) -> list[dict[str, Any]]:
                     if ri.status:
                         reasoning_item["status"] = ri.status
                     items.append(reasoning_item)
+
+            # Add message item after reasoning (if content exists)
+            if content_items:
+                items.append(
+                    {
+                        "type": "message",
+                        "role": "assistant",
+                        "content": content_items,
+                    }
+                )
+
             # Emit assistant tool calls so subsequent function_call_output
             # can match call_id
             if self.tool_calls: