OpenAI Agents

From the root of an AMW checkout, install the OpenAI wrapper and its Python SDK dependency. The wrapper is not published to PyPI. Install the optional OpenAI dependency if you will call a model:

python3 -m pip install -e ./sdk/python
python3 -m pip install -e './integrations/openai-agent-middleware[openai]'
export OPENAI_API_KEY='YOUR_OPENAI_API_KEY'

With the local quickstart running, register its tool with the runner and pass the returned definition to Chat Completions:

import asyncio
from openai import AsyncOpenAI
from openai_b2a import B2AClient, GovernedToolRunner, JsonFileOperationKeyStore

async def main():
    async with B2AClient(
        api_key="YOUR_WALLET_KEY", base_url="http://127.0.0.1:8000"
    ) as trust:
        runner = GovernedToolRunner(
            trust,
            wallet_id="YOUR_WALLET_ID",
            run_id="YOUR_STABLE_RUN_ID",
            key_store=JsonFileOperationKeyStore("operations.json"),
        )
        tools = [runner.register_tool(
            "partner.notes.write",
            description="Write a note",
            parameters={"type": "object", "properties": {"text": {"type": "string"}}},
        )]
        response = await AsyncOpenAI().chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": "Write a note that says hello"}],
            tools=tools,
        )
        for result in await runner.run_all(response.choices[0].message.tool_calls or []):
            print(result.receipt.receipt_id)

asyncio.run(main())

The runner derives the Request key from the model tool call ID and persists it before sending the Call. Keep the same run_id, store, wallet, and tool definitions when resuming a run. For Responses API function calls, pass each function_call item to runner.run and append result.as_function_call_output(). The wrapper's CI uses recorded tool call shapes; a live OpenAI model call is not verified there.