Upskill routes agents to vetted skills

The idea is strong because it attacks a real failure mode in agent workflows: not model quality, but bad task routing and weak reuse of existing procedures.

• –Best part: it turns “search and select the right skill” into an explicit system behavior instead of hoping the model remembers or finds the right repo on its own.
• –The hybrid ranking story is credible; exact-match retrieval matters for flags, APIs, and install steps, while embeddings help with intent.
• –The security review layer is necessary, because skill registries are a prompt-injection and supply-chain magnet.
• –The main risk is operational: the product only works if the indexed skills are genuinely high quality and the routing is good enough to justify the extra step.
• –The optional env-aware ranking is useful, but it will need very clear privacy guarantees to avoid skepticism.