Anthropic experienced an explosion in AI-generated code: daily merges rose from about 2,500 in May 2026 to over 30,000 by September, exposing fundamental scaling limits in git-based developer infrastructure. The monorepo’s main-branch history approached 100 GB, compressed working copies topped 25 GB, and CI alone performed on the order of 25 million checkouts per day - loads that made hundreds of internal git mirrors and packfile maintenance untenable. Rather than shoehorning git into an always-online, agent-driven workflow, Anthropic built Scion and migrated the monorepo in five months. Scion embraces clients-as-online, server-centric workflows, cheap lock-free server-side commits, sparse and virtualized working copies, horizontally scalable services, and a rich source-control API that lets CI and automation query repository metadata without full clones. These choices also enabled complementary tooling like a new code-review surface, code search, and IAM systems.
Scion’s core is a write-ahead log of “changesets” that record file deltas and content-addressed blobs; derived data structures (blame, path history, skip lists) are computed recursively from parents and can be materialized eagerly or lazily. Servers act as large, cacheable workers backed by Spanner and never require a fully materialized repo; clients fetch specific trees and blobs on demand via the scn protocol rather than telling servers what they already have. That design eliminates packfiles and manual repo maintenance, reduces CPU and bandwidth pressure, and makes operations like diffs, merges, bisects, and automated reverts practical at agent scale.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.