Task tool subagents become orphans when parent session crashes

Problem

When a Claude Code session crashes (OOM, network timeout, user interrupt), subagent processes spawned by the Task tool become orphaned and continue running indefinitely, consuming memory.

Reproduction

1. Start a Claude Code session
2. Use the Task tool to spawn a subagent
3. Kill the parent session (e.g., kill -9 or let it OOM)
4. Observe the --resume process still running:

``
ps aux | grep 'claude.*--resume'
``

Example orphaned process

/Users/user/.local/bin/claude --output-format stream-json --verbose \
  --input-format stream-json --model claude-sonnet-4-5 \
  --resume 9c33ca78-1244-4e57-8cc8-184e144261c7 \
  --disallowedTools Bash,Read,Write,Edit,... \
  --permission-mode default

Impact

• Each orphaned subagent holds ~300-400MB RAM
• Multiple crashed sessions compound the leak
• Users hit OOM conditions, causing more crashes, spawning more orphans
• Workaround requires manual cleanup: pkill -f 'claude.*--resume'

Suggested fixes

1. Process groups: Spawn subagents in the same process group as parent so SIGHUP propagates on parent death

1. Heartbeat/watchdog: Subagents periodically check if parent is alive (e.g., via parent PID or IPC), self-terminate if not

1. Cleanup on startup: Claude Code could kill stale --resume processes from previous sessions on startup (with appropriate staleness check)

Environment

• macOS 15 (Darwin 25.2.0)
• Claude Code CLI
• Apple Silicon