LSP is for when you already know the symbol and want an exact answer, like go-to-definition or find-references. graft is for the step before that: you ask "where does auth happen" and it hands the agent the right files and lines to start from, so it reads a lot less to get oriented.

so graft solves for making claude code understand your codebase every session.

And I feel you can save a lot more tokens and also increase accuracy by just switching to new claude session without worrying about your codebase context. (doesn't solve for the session context yet, but we are still figuring out how can we solve for that. Maybe hooks that can manage your CLAUDE.md)