Claude.app leaks claude-code subprocesses; ~150+ accumulate over 12 hours

Environment

• Claude Desktop 2.1.121
• macOS Darwin 25.2.0 ARM64

Steps to reproduce

1. Open 2 or 3 Claude Code sessions in Claude.app and leave them running for 8+ hours.
2. From a terminal: \ps -ef | grep claude | wc -l\

Expected

~6-12 claude processes (parent + child for each open session, plus a few MCP server helpers).

Actual

175+ claude processes. They accumulate at a rate of one new pair every ~5 minutes, parented by Claude.app itself (not by any session). Old processes are not killed when new ones spawn. Each consumes 2-4% CPU. Cumulative load can drive the system to 80%+ CPU.

Diagnostics

• \ps -eo pid,ppid,etime,command\ shows pairs of claude processes spaced at exactly 5-minute intervals across the full uptime of Claude.app.
• All leaked children are parented by the Claude.app PID (\application.com.anthropic.claudefordesktop\), not by any user-opened session.
• The 3 active session PIDs (matched by \--resume <session-id>\ in argv) are clean. Everything else is orphan.
• No user code or scheduled task in our project invokes the claude binary. Verified by grep.

Workaround

• Quit Claude.app fully (Cmd-Q) and relaunch. Process tree dies. Counter resets.
• This costs the user all in-flight session context.

Impact

• System CPU pegs at 80%+ over time. Laptop fans spin up. Battery drains.
• User must Cmd-Q + restart on a multi-hour cadence to keep the machine usable.
• Other agents in the same Claude.app instance experience slow tool execution due to CPU starvation.

Request

1. Identify the spawn source inside Claude.app (likely a session keepalive or health check on a 5-minute cadence).
2. Either kill the prior process before spawning the next, or stop spawning entirely if existing children are healthy.