Feature request: Process lifecycle management for spawned background tasks

Problem

When Claude Code spawns background processes (via run_in_background or long-running bash commands), these processes can become orphaned and consume significant resources without the user or Claude being aware.

Real-world example

During a session, I spawned a Python process to analyze Excel files using openpyxl. The process hung while parsing a large workbook. When the session context was compacted (due to token limits), awareness of this process was lost entirely.

Result:

• Process ran for 136 minutes
• Consumed 16GB RAM (24% of system memory)
• Used 68% CPU continuously
• Only discovered when user manually checked ps aux

Proposed Solutions

1. Process Registry

Track all spawned child processes with:

• PID
• Start time
• Command/description
• Resource usage (periodic sampling)

2. Auto-cleanup on Context Compaction

When a session's context is compacted due to token limits:

• Include active background process PIDs in the compaction summary
• Optionally terminate long-running processes
• Or at minimum, flag them for the resumed session to check

3. Resource Watchdog

Implement configurable thresholds that trigger alerts or auto-termination:

• Maximum runtime (e.g., 30 minutes default)
• Maximum memory usage (e.g., 2GB default)
• User-configurable via settings

4. Default Timeouts for Background Tasks

The run_in_background parameter should support (or enforce) a maximum runtime:

• Default timeout (e.g., 10 minutes)
• Explicit timeout parameter for longer tasks
• Warning when spawning potentially long-running processes

5. Session Resume Process Check

When resuming from a compacted session:

• Automatically check for orphaned processes from the previous session
• Report them to the user
• Offer to terminate them

Environment

• Claude Code version: 2.1.11
• OS: macOS (Darwin 25.2.0)
• The orphaned process was spawned via Bash tool with a Python script

Additional Context

The current /tasks command helps track background tasks, but:

• It doesn't survive context compaction
• It doesn't show resource usage
• It doesn't warn about runaway processes

This is particularly important for users running Claude Code for extended sessions or on resource-constrained machines.