(feat): Prompt engineering to remind o1 to generate a patch

evaluation/swe_bench/run_infer.py

@@ -91,6 +91,14 @@ def get_instruction(instance: pd.Series, metadata: EvalMetadata):
             "Your thinking should be thorough and so it's fine if it's very long.\n"
         )
 
+    instruction += """
+        <IMPORTANT>
+        - You MUST generate only one action per turn!
+        - A patch is a set of changes to the source code of the codebase that you are given
+        - You MUST generate a patch that attempts to fix the issue described in the <pr_description>
+        </IMPORTANT>
+        """
+
     if RUN_WITH_BROWSING:
         instruction += (
             '<IMPORTANT!>\n'

openhands/agenthub/codeact_agent/codeact_agent.py

@@ -508,8 +508,12 @@ def _get_messages(self, state: State) -> list[Message]:
                 None,
             )
             # do not add this for function calling
-            if latest_user_message:
-                reminder_text = f'\n\nENVIRONMENT REMINDER: You have {state.max_iterations - state.iteration} turns left to complete the task. When finished reply with <finish></finish>.'
-                latest_user_message.content.append(TextContent(text=reminder_text))
+            if not latest_user_message:
+                return messages
+
+            # Build environment reminder text
+            reminder_text = f'\n\nENVIRONMENT REMINDER: You have {state.max_iterations - state.iteration} turns left to complete the task. When finished reply with <finish></finish>.'
+
+            latest_user_message.content.append(TextContent(text=reminder_text))
 
         return messages