Async / Long-Running Tool Calls

## Goal
Enable an agent to launch slow-running tool calls, remain responsive (continue conversation or perform other tasks), and automatically deliver results as soon as they are ready—without waiting for another user message.

## User Story
As an end user,
I want the agent to start time-consuming operations (large queries, heavy ML jobs, etc.) without freezing the conversation,
so that I can keep chatting or receive interim updates, and the agent can push the final result the moment it completes - even if I have not sent a new query.

## Acceptance Criteria

ID | Description
-- | --
AC-1 | A step can tag any tool call with async: true and an optional timeout.
AC-2 | When an async call is triggered, it is placed in a queue with metadata {step_id, tool_name, params, timeout, started_at}.
AC-3 | While queued calls run, the agent can: (a) not transition to other steps, (b) send user-facing progress messages, (c) handle new user input normally, and (d) run non dependant tool calls.
AC-4 | If a queued call finishes within timeout, the result is injected into the originating step’s context and the agent immediately produces the next decision (no user prompt required).
AC-5 | If timeout expires, the agent must ask the user whether to extend the wait or skip the tool.
AC-6 | Declining to wait removes that tool from the step’s tool list and unlocks normal step transitions.
AC-7 | All queued calls are isolated per step; results cannot leak across steps unless explicitly shared.

```mermaid
sequenceDiagram
participant U as User
participant A as Agent
participant Q as Async-Queue
participant T as Tool (slow)

U->>A: Ask question requiring heavy job
A->>Q: enqueue(tool_call, timeout)
Q->>T: execute tool async
A-->>U: Processing… you can keep chatting

alt user keeps chatting
U->>A: New question
A-->>U: Answers immediately
end

T-->>Q: Result ready
Q->>A: deliver(tool_result)
A-->>U: Push final answer (no new prompt)

alt timeout expires
Q->>A: Timeout notification
A-->>U: Still running—wait longer?
U->>A: Yes, extend / No, skip
A->>Q: extend or cancel
end

```

## Summary
This asynchronous pattern lets the agent behave more like a real assistant: it can start long-running tasks, remain interactive, and proactively deliver results as soon as they are available, improving both user experience and agent productivity.