GSD is super powerful but still... right now when Claude starts a session, it knows nothing about your codebase. Every task begins with exploration - grepping, globbing, reading files to figure out what exists and how it fits together. Today I'm shipping a solution to this: A continually updating living graph database that tracks: - What files exist and what they do - Which files depend on which - Where the high-impact "change carefully" files are The planner agent reads this before writing plans. Instead of "figure out the codebase structure" as step 1 of every phase, it already knows.