Windows Codex Desktop: opening/resuming historical subagent threads spawns duplicate MCP and node_repl process stacks #37453

app_server.request ... otel.name="thread/resume"
MCP server ... Processing request of type ListToolsRequest
codex_mcp::connection_manager::tool_catalog

### What steps can reproduce the bug?

## Summary

Codex Desktop on Windows creates duplicate local MCP and `node_repl.exe` process stacks when I open/resume historical chats that contain many completed subagents.

This occurs even when I do not send a new message or run a tool. Simply opening an old parent chat, and then opening a completed child thread in the Subagents panel, triggers `thread/resume` and starts another full set of local MCP processes.

The processes remain alive as children of the current `codex.exe app-server` instead of returning to a bounded baseline.

## Environment

- OS: Windows
- Codex Desktop package: `OpenAI.Codex_26.803.5235.0_x64`
- Main process:
  `codex.exe -c features.code_mode_host=true app-server --analytics-default-enabled`
- Configured local stdio MCP servers:
  - MarkItDown
  - brain-memory
  - brain-knowledge
- Also enabled: `node_repl` / Browser / Computer Use

## Reproduction

1. Fully quit Codex Desktop.
2. Launch Codex Desktop and do not open historical chats initially.
3. Baseline is one process stack:
   - 1 `node_repl.exe`
   - 1 MarkItDown MCP wrapper
   - 1 brain-memory MCP wrapper
   - 1 brain-knowledge MCP wrapper
4. Open one historical parent chat with many old subagents. Do not send a message.
5. The Subagents panel initially shows many old children as `Working`.
6. Open one or more old child threads. Their displayed state immediately changes from `Working` to `Done`.
7. Observe duplicate process stacks under the current `codex.exe`.

## Observed result

After opening a historical parent chat and inspecting old child threads, the current `codex.exe` had:

- 8 `node_repl.exe` direct children
- 8 MarkItDown MCP wrapper processes
- 8 brain-memory MCP wrapper processes
- 8 brain-knowledge MCP wrapper processes
- 32 direct helper processes total
- 48 Python processes in the Codex process tree

No new user task, subagent launch, or MCP tool call was requested.

The local app-server log shows `thread/resume` at the same timestamps as MCP startup / tool catalog initialization, including:

```text
app_server.request ... otel.name="thread/resume"
MCP server ... Processing request of type ListToolsRequest
codex_mcp::connection_manager::tool_catalog

### What is the expected behavior?

_No response_

### Additional information

Additional context
This appears related to MCP lifecycle/refresh handling. PR #19753 added explicit shutdown handling for MCP refresh and session shutdown:
https://github.com/openai/codex/pull/19753
Potentially related report:
https://github.com/openai/codex/issues/26869