OpenAI Agents SDK client_session_timeout_seconds with Sume jobs_wait

The OpenAI Agents SDK for Python connects to MCP servers through classes such as MCPServerStreamableHttp. The SDK's MCP page says client_session_timeout_seconds sets the client session timeout: a positive value is a finite timeout, and None or 0 disables it, read 2026-10-03. Sume's jobs_wait can hold a call for up to 55 seconds, so the number you pick decides whether a long wait succeeds.

Pick a number against the hold

The Sume tools and gates page says jobs_wait holds 50 seconds by default and 55 at most; larger requests are clamped and the response carries wait_slice_clamped. A client timeout shorter than the hold turns a normal wait into an error even though the job is fine. Set it above 55 seconds, or leave it off and rely on your own outer deadline.

import asyncio, os
from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp

async def main() -> None:
    key = os.environ["SUME_API_KEY"]
    async with MCPServerStreamableHttp(
        name="sume",
        params={
            "url": "https://mcp.sume.com/mcp",
            "headers": {"Authorization": f"Bearer {key}"},
        },
        client_session_timeout_seconds=70,
        cache_tools_list=True,
    ) as server:
        agent = Agent(name="Reader", mcp_servers=[server],
                      instructions="Read job status with Sume tools.")
        r = await Runner.run(agent, "Wait on job abc and report.")
        print(r.final_output)

asyncio.run(main())

What a timeout means

A client-side timeout is a transport failure, not a job outcome. The Sume jobs guide says a client timeout does not cancel the job and it keeps billing. When Sume answers wait_slice_expired, call the wait again with the same job ids. Never resubmit the paid create, which would start a second job unless the idempotency_key repeats.

The code uses a read-only agent on purpose. Give an agent the create tools only when its task needs them.

Checklist

• Set the timeout above 55 seconds, or disable it and keep a deadline of your own.
• Persist job ids before waiting.
• Use wait_for of all or any with up to 20 ids.
• Treat 524, 522, 523 and 525 as transport failures.