Background agents/tasks not cleaned up, causing 100% memory usage

Description

When using background tasks (run_in_background: true with Bash tool or Task tool with background agents), child processes are not properly cleaned up when:

1. The task completes but TaskOutput is never called to retrieve results
2. The session ends or conversation switches before output is retrieved
3. There's an error during execution

These orphaned processes continue running and consuming memory, eventually leading to 100% memory usage.

Steps to Reproduce

1. Start a Claude Code session
2. Run a long-running command with run_in_background: true:

``
Bash tool with run_in_background: true
command: "npm test -- --testPathPattern='some-pattern' --testTimeout=30000"
``

1. Either:

• End the session before calling TaskOutput
• Switch to a different conversation
• Let the command complete but don't retrieve output

Expected Behavior

Child processes spawned by background tasks should be:

• Tracked by the parent Claude process
• Terminated when the session ends
• Cleaned up if TaskOutput is never called (with a reasonable timeout)

Actual Behavior

• Child processes become orphaned (PPID changes or parent exits)
• Processes continue running indefinitely
• Memory usage grows unbounded
• Task output files in /private/tmp/claude/*/tasks/ remain as 0-byte files

Evidence

Task output directory shows incomplete tasks (0-byte files):

$ ls -la /private/tmp/claude/-Users-*/tasks/
-rw-r--r--  1 user  wheel  2551 Jan 19 15:15 b12ce7a.output  # completed
-rw-r--r--  1 user  wheel     0 Jan 19 15:12 b15840a.output  # orphaned
-rw-r--r--  1 user  wheel     0 Jan 19 15:06 b5dbd8d.output  # orphaned
-rw-r--r--  1 user  wheel  2273 Jan 19 15:10 b6cfad7.output  # completed
-rw-r--r--  1 user  wheel     0 Jan 19 15:04 b82d28e.output  # orphaned
-rw-r--r--  1 user  wheel     0 Jan 19 15:03 bea612e.output  # orphaned

Environment

• OS: macOS (Darwin 25.2.0)
• Claude Code version: 2.1.12
• Shell: zsh

Suggested Fix

1. Register child processes in a process group
2. Set up SIGTERM/SIGINT handlers to clean up child processes on session end
3. Implement a timeout mechanism for background tasks that auto-terminates after N minutes if output isn't retrieved
4. Consider using process.on('exit', cleanup) to ensure cleanup on normal exit

Workaround

Users can manually clean up orphaned processes:

# Kill orphaned test runners
pkill -f "vitest run"
pkill -f "tsx scripts"

# Clean up old task files
find /private/tmp/claude -name "*.output" -mmin +60 -delete