Signal: Codex was yesterday's lead. Commit 6b5f574, merged on June 29, is a narrower follow-up worth tracking because it moves skill behavior into model metadata.
Public source: PR #29740 adds a false-by-default include_skills_usage_instructions model metadata field, enables it for bundled gpt-5.5 metadata, consumes it in core and extension skill rendering, and removes the old hardcoded legacy-model path.
Implication: skill guidance is becoming a model capability flag. That is a cleaner integration contract for agent builders than asking code paths to infer behavior from model names.
Caveat: the evidence does not show a public release or a general OpenAI API rule. It is a Codex repository change with tests across app-server, core, extensions, and model metadata.
What to watch: more agent UI behavior may move into explicit model metadata. This stays compact because it continues yesterday's Codex trust/runtime beat and is narrower than Tau's new public onboarding surface.