fix(s08): flush background task results before agent loop exits

[sirhappy]

Problem

Closes #46.

In s08_background_tasks.py (and s_full.py), the agent_loop exits immediately when stop_reason != "tool_use". Any background task results that arrive during the final LLM call, or from tasks that are still running at that point, are silently dropped — the notification queue is never drained again.

Minimal reproduction:

1. Ask the agent to run a slow command in the background (background_run("sleep 5 && echo done")).
2. The LLM responds with "OK, running in the background" and returns end_turn.
3. The loop exits. When the task finishes 5 seconds later its result sits in the queue forever.

Fix

At the exit branch (stop_reason != "tool_use") in agent_loop, before returning:

1. Immediate drain — flush any results that completed while the LLM was generating its response.
2. Wait loop — if tasks are still running, poll up to MAX_BG_WAIT (30 s) for them to finish, then drain again.
3. Re-enter loop — if there are pending results, inject them as a <background-results> message and continue so the LLM can acknowledge them before the session ends.

if response.stop_reason != "tool_use":
    pending = BG.drain_notifications()
    if not pending:
        still_running = [t for t in BG.tasks.values() if t["status"] == "running"]
        if still_running:
            deadline = time.monotonic() + MAX_BG_WAIT
            while time.monotonic() < deadline:
                if not any(t["status"] == "running" for t in BG.tasks.values()):
                    break
                time.sleep(0.5)
            pending = BG.drain_notifications()
    if pending:
        notif_text = "\n".join(...)
        messages.append(...)
        continue  # let the LLM see the late-arriving results
    return

The same fix is applied to s_full.py where BackgroundManager is also used.

Files changed

File	Change
agents/s08_background_tasks.py	Add import time, MAX_BG_WAIT constant, final-drain logic in agent_loop
agents/s_full.py	Add MAX_BG_WAIT constant, same final-drain logic in agent_loop

[vercel]

@sirhappy is attempting to deploy a commit to the crazyboym's projects Team on Vercel.

A member of the Team first needs to authorize it.