feat(agent): add push_notification to inject messages during multi-turn prompts

[lazytanuki]

Hi!

This PR adds a push_notification method to Agent that allows injecting user messages into an already-running multi-turn prompt, similar to how OpenCode makes it possible to send input to an active agent mid-conversation.

How it works

• Agent::push_notification(conversation_id, message) enqueues a message into a per-conversation pending queue.
• At the start of each loop iteration (both streaming and non-streaming), the queue is drained and messages are injected as User messages into the conversation history.
• The queue is stored behind Arc<Mutex<...>>, so cloning the agent shares the same queue, enabling cross-task patterns like spawning a background watcher that calls push_notification when an external event occurs.

Example

let agent = openai::Client::from_env()?
    .agent(openai::GPT_5_2)
    .conversation_id("my-conv")
    .toot(some_tool)
    .build();
let agent_clone = agent.clone();

let prompt_task = tokio::spawn(async move {
    // Start a long running multi-turn prompt.
    let resp = agent.prompt("Do this and that with your available tools.").await.unwrap();
    println!("{resp}");
})

let some_watcher_task = tokio::spawn(async move {
    // ... wait for some condition ...
    agent_clone.push_notification("my-conv", "Continue, but you should know that ...".to_string());
});

// Wait on both tasks...

Disclosure

• Feature code was handwritten.
• Tests were AI-generated then checked.

Thanks!