[BUG] Main agent stalls after a sub-agent (Agent/Task tool) completes — result not propagated, no auto-resume

Preflight

• [x] Searched existing issues. Closest I found is possibly #47936 ("(Async) Subagents stopping early"), but that is framed as the subagent terminating early; my symptom is the parent failing to resume after a sub-agent that

completed successfully, and it appears to be a recent regression (I could not confirm #47936's details, so filing separately).

• [x] Single bug.
• [x] On the latest version (2.1.187).

### What's wrong
When the main agent spawns sub-agents via the Agent/Task tool and awaits their completion, the main agent is not reliably resumed when a sub-agent finishes. The sub-agents complete their work correctly, but:

• the main agent receives only an idle_notification heartbeat (idleReason: "available") instead of the sub-agent's final result, and
• frequently the main agent stalls entirely — the CLI appears to stop mid-process and emits nothing further until I manually type a message.

To retrieve each finished sub-agent's output I had to manually message it and ask it to relay its result — results that should have been delivered automatically.

### What should happen
When a sub-agent completes, the main agent should be resumed automatically with that sub-agent's final result (as happens correctly for background Bash tasks via <task-notification>), and the workflow should continue without
manual intervention.

### Error messages
None. No error is shown; the session silently halts. The only signals received are idle_notification heartbeats from the finished sub-agents.

### Steps to reproduce

1. From a main session, spawn several named sub-agents with the Agent/Task tool, each doing a few minutes of work (e.g. a review fan-out of 5–15 agents).
2. Let the main agent await their completion.
3. As sub-agents finish, they emit idle_notification heartbeats, but their final results are not delivered to the parent; the main agent does not auto-continue and the session appears to halt.
4. A manual user message un-sticks the parent; results then have to be pulled per-agent via follow-up messages.

### Frequency / regression
Happened multiple times within a single multi-agent session today. The same multi-agent workflows ran fine before a recent CLI update.

### Environment

• Claude Code: 2.1.187
• Node: v25.2.1, npm: 11.12.1
• OS: Ubuntu 25.10 (Linux 6.17.0-35-generic, x86_64)
• Shell: bash
• Contrast: background Bash tasks (run_in_background) DID deliver a proper completion <task-notification> that correctly resumed the main agent in the same session — so the defect appears specific to the sub-agent

(Agent/Task) result-propagation / parent-resume path.