Redesigning source control at Anthropic

This post details the design and implementation of Scion, Anthropic's in-house source control system, built to handle the extreme scale of AI-driven code generation. Key contributions include: a client-server architecture assuming online clients, client-server asymmetry where the server handles complex operations, horizontally scalable components with no fully materialized repo, cheap/incremental/lock-free server-side commits, and optional/sparse/virtualized working copies. It introduces 'changesets' as the unit of work and 'derived data' computed recursively from parent changesets for efficient querying (e.g., blame, history). The client protocol focuses on requesting specific objects (trees and blobs) recursively, enabling tiered caching and efficient, resumable operations without large packfiles.