Feature request: add a global explicit-only mode for skills

What feature would you like to see?

Please add a global user-level setting that disables implicit skill invocation for all skills, while keeping skills available for explicit $skill-name invocation.

Today, a skill author can set:

policy:
  allow_implicit_invocation: false

in agents/openai.yaml, but that is a property of each individual skill. As a user, I want to install skills without allowing Codex to automatically select or invoke any of them unless I explicitly ask for one.

A single config option would solve this cleanly, for example:

[skills]
allow_implicit_invocation = false

or an equivalent explicit-only mode.

Expected behavior

With the global setting enabled:

• Codex must not implicitly select, load, or invoke any skill.
• Explicit $skill-name invocation should continue to work normally.
• Skills should remain installed and discoverable; this should not be equivalent to enabled = false.
• The user should not need to modify agents/openai.yaml inside every installed skill.
• Ideally the setting can be overridden per profile/session later, but a global user-level switch is the core request.

Why this is useful

Implicit skill routing is not desirable for every workflow. With a large skill library, automatic invocation can add unwanted workflow behavior and context overhead. More importantly, the policy of whether Codex may invoke a skill automatically should be controllable by the user, rather than requiring every skill author to opt out individually.

This is especially useful for users who want skills to behave like explicit commands: installed and ready, but never activated unless explicitly requested.

Relationship to existing issues

• #16172 requests explicit-only skill injection in general.
• #36248 requests local per-skill overrides for implicit invocation.
• #34885 discusses overly aggressive implicit skill selection and mentions an explicit-only mode.

This request is intentionally narrower: provide one global user-level switch that makes all skills explicit-only without editing or enumerating individual skill configurations.

Start by tracing how the existing per-skill policy.allow_implicit_invocation setting in agents/openai.yaml is read and how implicit and explicit skill invocation are selected. Review issues #16172, #36248, and #34885, then define the user-level configuration entry point and tests that verify global blocking, explicit invocation, installation, and discoverability.