Parent session deadlocks when subagent tool execution hangs — no timeout or recovery #44783

Summary

Parent session becomes completely unresponsive when a subagent's tool execution hangs. No timeout fires, no circuit breaker kicks in, and user messages queue indefinitely without being processed. Session was stuck for 7+ hours until manually killed. Pressing escape multiple times does not interrupt the agent/subagent.

Steps to Reproduce

1. Launch two subagents in parallel (foreground/blocking) via the Agent tool
2. One subagent's Bash tool call (in this case, curl to GitHub API) hangs indefinitely
3. Parent session blocks waiting for both agents to return

Observed Behavior

• First subagent (a3d5e...): Only 5 messages in its log. Last message is a tool_use (Bash) with stop_reason=tool_use — waiting for tool results that never came. No timeout fired.
• Second subagent (aaa4c...): Completed its work (33 messages, multiple WebFetch calls, produced results) but its result was never delivered back to the parent session.
• Parent session: After the two Agent tool calls, there are zero tool_result entries. The parent was permanently blocked.
• User messages: Follow-up messages at 09:49 and 09:51 were enqueued but never dequeued. Multiple popAll/re-enqueue attempts ~7 hours later also failed to unblock the session.
• Only recovery: Manually killing the session process.

Expected Behavior

• Subagent tool calls should respect the standard 2-minute timeout
• If a subagent hangs, the parent should detect it (watchdog/circuit breaker) and either kill the subagent or return an error
• The parent session should remain responsive to user input even while subagents are running, or at minimum allow interruption
• Completed subagent results should be delivered to the parent regardless of sibling subagent state

Environment

• Claude Code CLI (remote session via claude-screen)
• Model: claude-opus-4-6
• Platform: Linux (Fedora 43)
• Session ID: 3893c609-e6ab-43d0-9073-ebdc6d4ed0d2

Session Log Evidence

Parent session log shows the deadlock clearly:

09:34:32 [queue-operation] enqueue (user message)
09:34:32 [queue-operation] dequeue
09:34:36 [assistant] stop=None
09:34:40 [assistant] stop=None (launches agents)
09:34:46 [assistant] stop=tool_use (Agent tool calls)
09:36:44 [user] (empty — likely agent notification that was never processed)
09:49:52 [queue-operation] enqueue (user follow-up — never dequeued)
09:51:50 [queue-operation] enqueue (another follow-up — never dequeued)
16:54:04 [queue-operation] popAll (7 hours later, user trying to unstick)
... multiple popAll/enqueue cycles, none successful

No tool_result entries exist after the Agent tool calls — the results were never delivered.