[Windows Desktop 26.810.41047] Local STDIO MCP process groups accumulate and consume ~7 GB RAM

Title: [Windows Desktop 26.810.41047] Local STDIO MCP process groups accumulate and consume ~7 GB RAM

Environment:

• Windows
• ChatGPT/Codex desktop app 26.810.41047
• Local STDIO MCP servers: Playwright, todo-query, node_repl

Actual behavior:
After starting/resuming multiple Codex threads, local MCP process groups accumulate and are not reclaimed when most threads become idle or unloaded.

Observed:

• 26 identical MCP process groups
• 183 processes total
• Approximately 7 GB RAM and 1,200 threads
• Each group contains:
  • 1 node_repl.exe
  • 2 Python processes for todo-query
  • 2 cmd.exe and 2 node.exe processes for Playwright MCP
• The app listed only 1 active and approximately 4 idle threads at the time
• System became noticeably unresponsive

Logs show each thread/start initializing a new MCP manager and launching the configured STDIO servers. Some temporary groups exit gracefully, but many remain attached to the main Codex process.

Expected behavior:
MCP processes belonging to closed, unloaded, or expired thread runtimes should be terminated or reused instead of accumulating.

Workaround:
Restarting the app or disabling local MCP servers releases/prevents the accumulated processes.

Start at the thread/start flow that initializes the MCP manager and launches the configured STDIO servers; the payload does not name source files or tests. Reproduce on Windows with the listed local servers, then verify that closed, unloaded, or expired runtimes terminate or reuse their MCP processes instead of accumulating groups.