Request to submit deferred MCP discovery fix

Open 💬 0 comments Opened Aug 5, 2026 by a13onat9

Hello,

I have a reviewed and tested fix for a deferred MCP discovery issue in Codex.

Root cause:
Configured MCP server names mentioned in normal language were not added to the turn’s required server set before the initial tool catalog was captured. A fresh conversation could therefore incorrectly report a configured MCP server as unavailable before deferred discovery completed.

Fix:

  • Generic whole-name, case-insensitive matching for configured MCP server names
  • Wait for explicitly mentioned configured servers before initial tool-catalog capture
  • Handles the fresh-session current_config() race by combining the turn’s configured server map with the resolved runtime catalog
  • Adds secret-free discovery lifecycle logging
  • No provider-specific, ContextSync-specific, authentication, authorization, or MCP protocol changes

Validation:

  • codex-mcp: 165/165 passed
  • focused mention regressions: 2/2 passed
  • delayed-discovery runtime regression: passed
  • no-MCP-mention behavior: passed
  • scoped Clippy: passed
  • formatting and git diff --check: passed

Prepared branch:

  • Fork: a13onat9/codex
  • Branch: fix/deferred-mcp-server-mentions
  • Commit: 1ecc5224978492114a314b6f3dcd8a29b70ab4ab

GitHub currently blocks PR creation because only collaborators may open pull requests.

Could you please either:

  • invite a13onat9 as a collaborator with permission to open the PR, or
  • open the PR from the prepared fork branch?

Suggested release note:
“Codex now waits for deferred discovery of enabled, configured MCP servers explicitly mentioned by name, case-insensitively and with whole-name matching, before capturing the turn’s initial tool catalog.”

Thank you.

View original on GitHub ↗