Codex desktop leaks orphaned MCP helper processes under launchd

Summary

Codex Desktop appears to repeatedly leak MCP helper processes. Over time, they become re-parented to launchd (PPID 1) and continue consuming CPU/RAM after the Codex session/tool invocation that created them is gone.

This has happened repeatedly on my machine. In the latest cleanup, I found and removed 501 stale MCP-related processes.

Environment

• App: Codex Desktop
• Codex app version: 26.429.30905 (CFBundleVersion 2345)
• Bundle id: com.openai.codex
• macOS: 26.3.1 (a) build 25D771280a
• Architecture: arm64

What I observed

A host-level process snapshot showed hundreds of stale MCP helpers with PPID 1, mostly:

• php /Applications/Herd.app/Contents/Resources/herd-mcp.phar
• node /private/tmp/bunx-501-@delorenj/mcp-server-trello@latest/node_modules/.bin/mcp-server-trello

Before cleanup:

target_count=501
total count=501 cpu=190.3 rss_mb=4013.5
php count=252 cpu=190.3 rss_mb=1660.3
node count=249 cpu=0.0 rss_mb=2353.2

After killing only MCP-related processes with PPID 1:

remaining_target_count=0
total count=0 cpu=0.0 rss_mb=0.0

There were also two STAT=Z defunct processes, but they were not the material resource problem because both had 0% CPU and 0 RSS.

Why this looks like a Codex leak

The live Codex app had active helpers under the current Codex process tree, which I left alone. The stale processes were specifically the MCP helpers re-parented to launchd with PPID 1, which suggests their original parent exited without terminating/reaping them.

This matches a recurring pattern I have seen where many Codex/MCP helper processes accumulate across sessions/tool invocations.

Expected behavior

When a Codex session, tool server, or MCP client process exits, Codex should terminate/reap any MCP helper process it spawned, or otherwise ensure those child processes do not become long-lived orphaned PPID 1 processes.

Actual behavior

MCP helper processes can accumulate under launchd and persist for hours/days, consuming significant CPU and RAM.

Detection command used

ps -axo pid,ppid,stat,%cpu,%mem,rss,etime,comm,args \
  | awk 'NR>1 && $2==1 && /mcp-server-trello|herd-mcp\.phar|chrome-devtools-mcp|playwright-mcp|watchdog\/main\.js/'

Cleanup command used

ps -axo pid,ppid,stat,%cpu,%mem,rss,etime,comm,args > /tmp/codex_ps_before_cleanup.txt
awk 'NR>1 && $2==1 && /mcp-server-trello|herd-mcp\.phar|chrome-devtools-mcp|playwright-mcp|watchdog\/main\.js/ { print $1 }' /tmp/codex_ps_before_cleanup.txt > /tmp/codex_stale_mcp_pids.txt
xargs kill < /tmp/codex_stale_mcp_pids.txt

I intentionally scoped cleanup to MCP-related commands with PPID 1 so live Codex-attached helper processes were not killed.