Support an opt-in authoritative external skill provider

Resolved 💬 2 comments Opened Jul 21, 2026 by rleungx Closed Jul 21, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

Summary

Support an opt-in authoritative external skill provider that can be used directly by Codex CLI.

Motivation

Organizations may store and manage skills in databases or internal services instead of the local filesystem. They may also need custom routing based on permissions, metadata, availability, or organization-specific policies.

Currently, there is no supported way for Codex CLI to delegate automatic skill discovery and selection to such an external system.

Expected behavior

When an external skill provider is configured:

  • The provider becomes authoritative for automatic skill discovery and selection.
  • Codex does not expose its native skill catalog to the model for implicit selection.
  • The provider decides which skills, if any, apply to each user turn.
  • Codex loads and injects the context for the selected skills.
  • Returning no skill is a valid result.
  • Native skills remain available through explicit user invocation.

When no external provider is configured, Codex keeps its existing skill behavior unchanged.

Use cases

  • Store and manage skills in a database.
  • Apply organization-specific access control and routing policies.
  • Select skills using external metadata, search, or ranking systems.
  • Keep skill management consistent across machines and CLI sessions.

Compatibility

This feature should be opt-in and should not change the default Codex experience. Existing sandbox, approval, and permission boundaries should continue to apply.

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 1 month ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #33945

Powered by Codex Action

etraut-openai contributor · 1 month ago

Implicit skill selection is performed by the model from the skill catalog Codex exposes; the harness cannot delegate that decision to an external per-turn provider. You can influence selection through instructions in AGENTS.md or explicitly invoke a skill.

Organization-managed plugin marketplaces can centrally control which skills are distributed and available, but they do not provide authoritative per-turn routing.