Investigate Codex Desktop GitNexus MCP process pile

[100yenadmin]

Summary

After consolidating Codex config to one gitnexus MCP server, Codex Desktop still accumulates many app-parented stdio GitNexus MCP child processes:

node /Volumes/LEXAR/repos/GitNexus/gitnexus/dist/cli/index.js mcp

This appears separate from PR #20's Voyage embedding/runtime fix. Direct local JSON-RPC smoke processes exit cleanly when the parent script closes them, but app-managed children remain parented to the Codex app server with live stdio pipes.

Current evidence

Observed after PR #20 merged and local main rebuilt:

• parent process: /Applications/Codex.app/Contents/Resources/codex app-server --analytics-default-enabled
• matching child count: around 20, with ages ranging from minutes to 12+ hours
• all children use the canonical merged dist path
• Codex config has one MCP server named gitnexus; no openclaw-code-index MCP block

Evidence folder:

/Volumes/LEXAR/Codex/gitnexus-reconciliation-2026-06-10/legacy-openclaw-embedding-regression-20260615T135034Z/

Why this matters

The original reconciliation concern included per-agent MCP child buildup and long-lived orphan-like processes. Even with one configured MCP server, stdio MCP still spawns per agent/session; stale cleanup appears to need a separate lifecycle strategy.

Acceptance criteria

• Distinguish live session MCP children from stale abandoned ones.
• Add/verify an exit path for stdio close, parent disconnect, SIGTERM, and idle abandonment where GitNexus owns the behavior.
• Document what Codex Desktop owns if process retention is entirely app-managed.
• Provide a safe cleanup command/runbook that does not kill active session transports.