This is entirely based on the "agent skills" system. LLM agent only sees the one-line skill description in its context and "lazy loads" the rest of the skill file on demand.

The lazy loading approach is smart. We've been publishing agent skills too and the context budget is a real constraint; six skills with reference docs would blow past 30k tokens if loaded eagerly.

Filtering at load time based on what the agent actually needs makes a huge difference. Curious if the orchestrator/executor split causes issues with state handoff between the two context forks.