openai[patch]: support multi-turn computer use

[ccurme]

Here we accept ToolMessages of the form

ToolMessage(
    content=<representation of screenshot> (see below),
    tool_call_id="abc123",
    additional_kwargs={"type": "computer_call_output"},
)

and translate them to computer_call_output items for the Responses API.

We also propagate reasoning_content items from AIMessages.

Example

Load screenshots

import base64

def load_png_as_base64(file_path):
    with open(file_path, "rb") as image_file:
        encoded_string = base64.b64encode(image_file.read())
        return encoded_string.decode('utf-8')

screenshot_1_base64 = load_png_as_base64("/path/to/screenshot/of/application.png")
screenshot_2_base64 = load_png_as_base64("/path/to/screenshot/of/desktop.png")

Initial message and response

from langchain_core.messages import HumanMessage, ToolMessage
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="computer-use-preview",
    model_kwargs={"truncation": "auto"},
)

tool = {
    "type": "computer_use_preview",
    "display_width": 1024,
    "display_height": 768,
    "environment": "browser"
}
llm_with_tools = llm.bind_tools([tool])

input_message = HumanMessage(
    content=[
        {
            "type": "text",
            "text": (
                "Click the red X to close and reveal my Desktop. "
                "Proceed, no confirmation needed."
            )
        },
        {
            "type": "input_image",
            "image_url": f"data:image/png;base64,{screenshot_1_base64}",
        }
    ]
)

response = llm_with_tools.invoke(
    [input_message],
    reasoning={
        "generate_summary": "concise",
    },
)
response.additional_kwargs["tool_outputs"]

Construct ToolMessage

tool_call_id = response.additional_kwargs["tool_outputs"][0]["call_id"]

tool_message = ToolMessage(
    content=[
        {
            "type": "input_image",
            "image_url": f"data:image/png;base64,{screenshot_2_base64}"
        }
    ],
    #  content=f"data:image/png;base64,{screenshot_2_base64}",  # <-- also acceptable
    tool_call_id=tool_call_id,
    additional_kwargs={"type": "computer_call_output"},
)

Invoke again

messages = [
    input_message,
    response,
    tool_message,
]

response_2 = llm_with_tools.invoke(
    messages,
    reasoning={
        "generate_summary": "concise",
    },
)

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Mar 24, 2025 3:25pm

[bracesproul]

cha-ching!