diff --git a/src/agents/models/openai_responses.py b/src/agents/models/openai_responses.py
index e060fb8ed..9f02d36bf 100644
--- a/src/agents/models/openai_responses.py
+++ b/src/agents/models/openai_responses.py
@@ -366,7 +366,7 @@ def _convert_tool(cls, tool: Tool) -> tuple[ToolParam, IncludeLiteral | None]:
                 "display_width": tool.computer.dimensions[0],
                 "display_height": tool.computer.dimensions[1],
             }
-            includes = None
+            includes = "computer_call_output.output.image_url"
 
         else:
             raise UserError(f"Unknown tool type: {type(tool)}, tool")
diff --git a/tests/test_openai_responses_converter.py b/tests/test_openai_responses_converter.py
index 34cbac5c5..8e6d9b31a 100644
--- a/tests/test_openai_responses_converter.py
+++ b/tests/test_openai_responses_converter.py
@@ -155,7 +155,9 @@ def drag(self, path: list[tuple[int, int]]) -> None:
     assert isinstance(converted.includes, list)
     # The includes list should have exactly the include for file search when include_search_results
     # is True.
-    assert converted.includes == ["file_search_call.results"]
+    assert "file_search_call.results" in converted.includes
+    # The includes list should also have the include for computer tool screenshots
+    assert "computer_call_output.output.image_url" in converted.includes
     # There should be exactly four converted tool dicts.
     assert len(converted.tools) == 4
     # Extract types and verify.
diff --git a/tests/test_reference_id_screenshots.py b/tests/test_reference_id_screenshots.py
new file mode 100644
index 000000000..277001d01
--- /dev/null
+++ b/tests/test_reference_id_screenshots.py
@@ -0,0 +1,53 @@
+"""Tests for the reference ID screenshot mechanism in ComputerAction."""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+from agents import Agent, RunConfig, RunContextWrapper, RunHooks
+from agents._run_impl import ComputerAction, ToolRunComputerAction
+from agents.items import ToolCallOutputItem
+from agents.tool import ComputerTool
+from openai.types.responses.response_computer_tool_call import (
+    ActionClick,
+    ResponseComputerToolCall,
+)
+from tests.test_computer_action import LoggingComputer
+
+
+@pytest.mark.asyncio
+@patch("agents.models.openai_provider._openai_shared.get_default_openai_client")
+@patch("openai.AsyncOpenAI")
+async def test_reference_id_screenshots(mock_openai, mock_get_client):
+    """Test that screenshots are sent with reference IDs."""
+    # Mock the OpenAI client to avoid API key requirement
+    mock_get_client.return_value = MagicMock()
+    computer = LoggingComputer(screenshot_return="test_screenshot")
+    comptool = ComputerTool(computer=computer)
+    # Create a dummy click action
+    action = ActionClick(type="click", x=1, y=2, button="left")
+    tool_call = ResponseComputerToolCall(
+        id="tool1",
+        type="computer_call",
+        action=action,
+        call_id="tool1",
+        pending_safety_checks=[],
+        status="completed",
+    )
+    # Setup agent and hooks
+    agent = Agent(name="test_agent", tools=[comptool])
+    run_hooks = RunHooks()
+    context_wrapper = RunContextWrapper(context=None)
+    # Execute the computer action
+    output_item = await ComputerAction.execute(
+        agent=agent,
+        action=ToolRunComputerAction(tool_call=tool_call, computer_tool=comptool),
+        hooks=run_hooks,
+        context_wrapper=context_wrapper,
+        config=RunConfig(),
+    )
+    # Verify that the output item has the correct structure
+    assert isinstance(output_item, ToolCallOutputItem)
+    assert "data:image/png;base64," in output_item.output
+    # Verify that the screenshot was generated
+    screenshot_calls = [call for call in computer.calls if call[0] == "screenshot"]
+    assert len(screenshot_calls) == 1