Active subagent work can be interrupted when requesting additional capacity

What issue are you seeing?

In a multi-agent session, requesting another worker while one is already running can interrupt the active worker and reassign its task. The original work then stops part-way through, and the replacement does not reliably resume from the exact checkpoint. This leaves the session unsure which work completed.

What steps can reproduce the bug?

1. Start a long-running task with worker A.
2. While worker A is working, request an additional worker for an independent task.
3. Observe capacity handling interrupt or reassign worker A instead of adding worker B.
4. Observe that worker A’s unfinished work is not clearly resumable and the user must reconstruct the missing work.

What is the expected behavior?

Active work should remain running when additional capacity is requested. If no slot is available, report that clearly and leave current workers untouched. If reassignment is explicitly necessary, preserve the full task state and checkpoint, explain the lifecycle transition, and ensure the replacement continues to completion.

Additional information

This report intentionally omits project, customer, financial, and task-specific details. The impact is lost continuity, duplicated coordination, and uncertainty about whether the original deliverable is still progressing. A visible lifecycle event log and an atomic handoff would make this diagnosable.

Start by reproducing the multi-agent capacity request described in the issue, then trace the worker lifecycle, capacity handling, reassignment, and checkpoint or handoff entry points. Done means active workers remain untouched when capacity is unavailable, or any required reassignment preserves task state, records the lifecycle transition, and resumes the work to completion.