diff --git a/spoon_ai/agents/toolcall.py b/spoon_ai/agents/toolcall.py
index 6676659..a055c92 100644
--- a/spoon_ai/agents/toolcall.py
+++ b/spoon_ai/agents/toolcall.py
@@ -25,7 +25,9 @@
 FINAL_RESPONSE_PROMPT = (
     "You have reached the tool budget. Do not call any more tools. "
     "Using only the tool results already in memory, provide the final user-facing answer now. "
-    "Summarize the concrete results and do not describe future actions."
+    "Follow the latest user's requested output format exactly. "
+    "Do not replace it with a recap or progress summary unless the user explicitly asked for one. "
+    "Do not describe future actions."
 )
 
 class ToolCallAgent(ReActAgent):
@@ -705,15 +707,12 @@ def _should_finish_execution(self, name: str, result: Any, **kwargs) -> bool:
 
     def _should_terminate_on_finish_reason(self, response) -> bool:
         """Check if agent should terminate based on finish_reason signals."""
-        # For Anthropic: native_finish_reason="end_turn" maps to finish_reason="stop"
-        # For OpenAI: both finish_reason and native_finish_reason are "stop"
         finish_reason = getattr(response, 'finish_reason', None)
-        native_finish_reason = getattr(response, 'native_finish_reason', None)
-
-        if finish_reason == "stop":
-            # Accept either "stop" (OpenAI) or "end_turn" (Anthropic) as valid termination signals
-            return native_finish_reason in ["stop", "end_turn"]
-        return False
+        # The normalized finish_reason is the contract spoon-core relies on.
+        # Native provider reasons differ across APIs (e.g. Responses API may
+        # surface "completed"), so requiring a legacy native value can turn a
+        # valid final answer into a repeated loop.
+        return finish_reason == "stop"
 
     async def _call_llm_with_middleware(
         self,
diff --git a/tests/test_agent_llm_integration.py b/tests/test_agent_llm_integration.py
index 00bf39f..253c0df 100644
--- a/tests/test_agent_llm_integration.py
+++ b/tests/test_agent_llm_integration.py
@@ -67,6 +67,32 @@ async def test_toolcall_agent_with_manager(self, mock_chatbot_manager, tool_mana
         mock_chatbot_manager.ask_tool.assert_called()
         assert "I'll help you with that task." in result
 
+    @pytest.mark.asyncio
+    async def test_toolcall_agent_terminates_on_stop_even_when_native_reason_is_completed(
+        self,
+        mock_chatbot_manager,
+        tool_manager,
+    ):
+        """Responses API may use native status values like 'completed' for terminal answers."""
+        mock_chatbot_manager.ask_tool.return_value = LLMResponse(
+            content="FINAL_OK",
+            tool_calls=[],
+            finish_reason="stop",
+            native_finish_reason="completed",
+        )
+
+        agent = ToolCallAgent(
+            name="test_agent",
+            llm=mock_chatbot_manager,
+            available_tools=tool_manager,
+            max_steps=1,
+        )
+
+        result = await agent.run("Reply with exactly: FINAL_OK")
+
+        assert result == "FINAL_OK"
+        assert mock_chatbot_manager.ask_tool.await_count == 1
+
     @pytest.mark.asyncio
     async def test_toolcall_agent_forwards_thinking_flag_to_llm(self, mock_chatbot_manager, tool_manager):
         mock_chatbot_manager.ask_tool.return_value = LLMResponse(
@@ -446,6 +472,17 @@ async def test_toolcall_agent_uses_final_tool_free_summary_after_budget_exhausti
         assert result == "Final summary after tool execution."
         assert mock_chatbot_manager.ask_tool.await_count == 1
         mock_chatbot_manager.ask.assert_awaited_once()
+        final_messages = mock_chatbot_manager.ask.await_args.kwargs["messages"]
+        assert any(
+            "Follow the latest user's requested output format exactly." in msg.content
+            for msg in final_messages
+            if isinstance(getattr(msg, "content", None), str)
+        )
+        assert any(
+            "Do not replace it with a recap or progress summary" in msg.content
+            for msg in final_messages
+            if isinstance(getattr(msg, "content", None), str)
+        )
     
     @pytest.mark.asyncio
     async def test_agent_memory_consistency(self, mock_chatbot_manager, tool_manager):