Summon extension causes agent to get confused when using async

[alexyao2015]

Describe the bug
Since #6964

Using gpt-5.2-codex and telling the agent to use a subagent to call a tool in an mcp server, it gets confused when it delegates the task then immediately tries to load it. The agent then calls a sleep command of a random number of seconds. The extension should instead notify the agent when the task is done instead of relying on the agent to guess when the task is done.

See the logs of the agent attempting to load the result of a task and it just sleeping

  ▸ load
    source: 20260219_5
    cancel: false

Annotated {
    raw: Text(
        RawTextContent {
            text: "Error: Task '20260219_5' is still running. Use load(source: \"20260219_5\", cancel: true) to stop.",
            meta: None,
        },
    ),
    annotations: None,
}

  ▸ shell developer
    command: sleep 296

-32603: request timeout after PT30S
  ▸ shell developer
    command: sleep 1

Annotated {
    raw: Text(
        RawTextContent {
            text: "",
            meta: None,
        },
    ),
    annotations: Some(
        Annotations {
            audience: Some(
                [
                    User,
                ],
            ),
            priority: Some(
                0.0,
            ),
            last_modified: None,
        },
    ),
}

  ▸ load
    source: 20260219_5
    cancel: false