Comment by crackalamoo

This seems like a solvable engineering problem. For example, you could have a lightweight subagent with its own context for reading the skills and determining which to use