Issue draft: Codex Desktop appears to leak MCP/helper process trees during UI debugging workflows

Title

Codex Desktop accumulates duplicate MCP/helper process trees and multi-GB RSS during UI/debugging workflows

Summary

I am seeing Codex Desktop accumulate many duplicate MCP/helper child-process trees during a recent UI debugging workflow involving browser/app automation, screenshots, and opening/closing app screens. The leak is visible outside the app with ps: a single Codex Desktop session currently has more than one hundred Codex-related processes in its process envelope, including repeated @playwright/mcp, xcodebuildmcp, context7-mcp, node_repl, and Computer Use helper processes.

At the latest sample, the Codex-related process envelope was using about 7.27 GiB RSS, with 105 helper/MCP processes using about 4.98 GiB RSS. This is high enough that I had to write an external local kill-switch monitor to prevent the machine from becoming unstable if the growth continues.

This looks related to, but not necessarily identical to, previously reported cleanup/leak issues around app-server, MCP tools, and Playwright MCP stdio processes:

• #12491
• #17574
• #17832
• #18881

Environment

• App: Codex Desktop for macOS
• Codex version: 26.422.71525
• Bundle version: 2210
• macOS: 26.2 build 25C56
• Architecture: arm64
• Main app process observed: /Applications/Codex.app/Contents/MacOS/Codex
• App-server process observed: /Applications/Codex.app/Contents/Resources/codex app-server --analytics-default-enabled

Workload that appears to trigger it

The recent task was primarily UI/debugging oriented:

1. Use Codex Desktop for local development assistance.
2. Invoke tools/plugins that spawn MCP/helper processes, especially browser/UI/simulator/debugging helpers.
3. Open/close UI screens, run screenshot or browser-style automation, and inspect app state repeatedly.
4. Continue in the same Codex Desktop session/thread for some time.

I do not yet have a minimal clean-room reproducer, but the process growth is visible in the live system state after this workflow.

Expected behavior

When a tool call, browser automation session, UI/screenshot inspection, simulator/debugging session, or Codex task finishes, Codex Desktop should clean up tool-specific MCP/helper processes that are no longer needed.

At steady state, I would expect:

• No unbounded growth in npm exec ... mcp wrappers.
• No unbounded growth in their child node .../.bin/*-mcp processes.
• No accumulation of stale helper process groups after UI/debugging tasks complete.
• Codex Desktop RSS and helper RSS to return close to baseline after work completes.

Actual behavior

A live sample taken on 2026-04-30 07:44:50 UTC showed:

Codex-related process count: 125
Focused Codex/MCP/helper ps rows: 111
Total Codex-envelope RSS: 7266.42 MiB
Helper/MCP process count: 105
Helper/MCP RSS: 4979.00 MiB
External guard decision in dry-run: kill_helpers
Guard reason: helper_count_limit

Representative largest processes from that sample:

PID    PPID   RSS MiB  Age(s)  Command
1106   653    453.52   1100    /Applications/Codex.app/.../Codex Helper (Renderer) --type=renderer ...
10766  653    271.89   30      /Applications/Codex.app/.../Codex Helper (Renderer) --type=renderer ...
653    1      227.03   1111    /Applications/Codex.app/Contents/MacOS/Codex
8987   4400   224.05   306     /Applications/Google Chrome.app/Contents/MacOS/Google Chrome ...
8998   8987   190.92   306     /Applications/Google Chrome.app/.../Google Chrome Helper (Renderer) ...
990    653    157.64   1101    /Applications/Codex.app/Contents/Resources/codex app-server --analytics-default-enabled
9964   9796   128.62   110     node .../.bin/xcodebuildmcp mcp
9908   9799   103.53   110     node .../.bin/playwright-mcp
6990   6867   97.02    427     node .../.bin/xcodebuildmcp mcp
5930   5828   94.58    513     node .../.bin/xcodebuildmcp mcp
9796   990    87.61    111     npm exec xcodebuildmcp@latest mcp
9800   990    86.45    111     npm exec @upstash/context7-mcp
9799   990    85.62    111     npm exec @playwright/mcp@latest
9285   9258   80.72    209     node .../.bin/playwright-mcp
4400   4289   79.70    730     node .../.bin/playwright-mcp
7450   7361   79.39    420     node .../.bin/xcodebuildmcp mcp
9904   9800   76.88    110     node .../.bin/context7-mcp
8594   8404   75.84    369     node .../.bin/xcodebuildmcp mcp
7467   7356   75.36    420     node .../.bin/playwright-mcp
7024   6862   75.34    427     node .../.bin/playwright-mcp

The important pattern is not one single large process. It is repeated process groups with the same shape, for example:

codex app-server
  npm exec @playwright/mcp@latest
    node .../.bin/playwright-mcp

codex app-server
  npm exec xcodebuildmcp@latest mcp
    node .../.bin/xcodebuildmcp mcp

codex app-server
  npm exec @upstash/context7-mcp
    node .../.bin/context7-mcp

Those groups appear many times in the same Desktop session.

Codex Desktop log observations

Recent logs were under:

~/Library/Logs/com.openai.codex/2026/04/30/codex-desktop-...-653-t0-i1-072628-0.log

Over the last 7 days, focused log scans found:

any_mcp=4835
workspace_dependencies_errors=4576
title_timeout=15
render_process_gone_mentions=3872
unresponsive_mentions=3872
xcodebuildmcp=0
playwright=1
Computer_Use=0
node_repl=0

A note on the render_process_gone_mentions / unresponsive_mentions counts: these are mentions inside serialized Electron event objects in repeated codex-home request log lines, not necessarily proof that those events fired. I am including them because those fields appear repeatedly while the process envelope is growing.

Representative recent log lines, with UUIDs redacted:

2026-04-30T07:36:16.281Z info [AppServerConnection] response_routed ... method=mcpServerStatus/list ... durationMs=3295 ...
2026-04-30T07:43:02.542Z info [AppServerConnection] response_routed ... method=mcpServerStatus/list ... durationMs=3927 ...
2026-04-30T07:44:20.744Z info [electron-fetch-handler] codex-home request hostId=undefined params={... "_events":{"render-process-gone":[...],"destroyed":[...],"unresponsive":[...]},"_eventsCount":32} ...
2026-04-30T07:44:20.744Z info [electron-fetch-handler] codex-home request hostId=local params={... "_events":{"render-process-gone":[...],"destroyed":[...],"unresponsive":[...]},"_eventsCount":32} ...
2026-04-30T07:44:50.570Z info [electron-fetch-handler] codex-home request hostId=undefined params={... "_events":{"render-process-gone":[...],"destroyed":[...],"unresponsive":[...]},"_eventsCount":32} ...
2026-04-30T07:44:50.571Z info [electron-fetch-handler] codex-home request hostId=local params={... "_events":{"render-process-gone":[...],"destroyed":[...],"unresponsive":[...]},"_eventsCount":32} ...

Earlier in the same log stream, there were repeated unsupported-feature errors:

unsupported feature enablement `workspace_dependencies`: currently supported features are apps, memories, plugins, tool_search, tool_suggest, tool_call_mcp_elicitation

I do not know if those feature-enable errors are causally related to the process leak. They are included only because they are frequent in the same recent desktop logs.

Commands used to collect evidence

Process sample:

ps -axo pid=,ppid=,rss=,etime=,command= \
  | grep -E '/Applications/Codex.app|@playwright/mcp|playwright-mcp|xcodebuildmcp|context7-mcp|node_repl|SkyComputerUseClient|chrome-devtools-mcp'

App/bundle version:

/usr/libexec/PlistBuddy -c 'Print :CFBundleShortVersionString' /Applications/Codex.app/Contents/Info.plist
/usr/libexec/PlistBuddy -c 'Print :CFBundleVersion' /Applications/Codex.app/Contents/Info.plist

Log scan, simplified:

find ~/Library/Logs/com.openai.codex -type f -mtime -7 -name '*.log' -print0 \
  | xargs -0 grep -hE 'mcp|workspace_dependencies|generate thread title|render-process-gone|unresponsive|xcodebuildmcp|playwright|Computer Use|node_repl'

Impact

The practical impact is that a long Codex Desktop session can accumulate enough helper processes and RSS to threaten overall system stability. In this session, helper processes alone were already near 5 GiB RSS. If the growth continues, the OS can become sluggish or run into memory pressure.

As a temporary mitigation, I wrote an external monitor that samples the Codex process envelope and, only after sustained threshold breaches, terminates Codex helper processes or the whole Codex tree. That is only a local safety net; ideally Codex Desktop should own this lifecycle cleanup internally.

Suggested investigation areas

• Whether MCP server lifecycle cleanup is tied to thread/session close but not to tool call completion.
• Whether repeated UI/browser/screenshot/debugger operations start new MCP server groups instead of reusing existing ones.
• Whether stdio MCP subprocesses survive when their owning renderer/tool invocation is done.
• Whether app-server retains references to old MCP server instances, preventing child cleanup.
• Whether the repeated mcpServerStatus/list calls reflect many still-registered servers rather than only active servers.
• Whether Desktop should enforce a maximum number of duplicate MCP helpers per server type and clean stale ones.

What would help confirm the fix

A fix would be easy to validate externally:

1. Start Codex Desktop fresh.
2. Run a UI/browser/screenshot/debugging workflow that invokes Playwright MCP, xcodebuild MCP, Context7 MCP, Computer Use, and node repl tools multiple times.
3. Wait for the tool calls to complete.
4. Check that stale npm exec ... mcp wrappers and child node .../.bin/*-mcp processes are cleaned up or reused.
5. Confirm process count and RSS return near baseline instead of growing monotonically.