Why Git pushes stall at AI speed, and the write-ahead log that ends the wait

The clone took hours. The rebase wedged. The index could not keep up.

For a team running a large monorepo against an AI coding agent that rewrites a hundred files in a minute, those failure modes are no longer edge cases. They are the daily experience of version control asked to operate at machine speed, and they trace to a single design choice that has shipped inside every Git push since 2013.

That design is Spokes, the packfile-replication system GitHub open-sourced in 2013, and the tax it charges shows up in two places. First, every push has to be confirmed by every replica through a three-phase commit (3PC), so the per-step latency is bounded by the slowest replica in the cluster. Add replicas to scale reads, and pushes get slower, not faster. Second, 3PC needs a quorum to work, which means every repo, no matter how small, has to live on at least three replicas. A throwaway branch from a coding agent pays the same floor as a production monorepo.

Cursor announced on X this week and followed with a detailed engineering breakdown of Origin, the in-house Git hosting platform that runs the everysphere monorepo, and a new storage system underneath it called Continuity. The architecture inverts the relationship between disk and truth. Local NVMe (the fast non-volatile storage drives that sit inside a server) is no longer where the repo lives; it is a cache. The source of truth is a write-ahead log (WAL) inside S3-compatible object storage, the same kind of bulk cloud storage that backs most modern data lakes.

A push only counts as acknowledged after the full WAL record is durable. The local repo gets the new commit, and a pointer to that commit is appended to a WAL index object. Because persistence is linear and the index is the single source of truth, no quorum round-trip is needed to publish a commit. Replication is then a separate concern. Origin uses optimistic gossip over UDP (the connectionless transport that prioritizes speed over delivery guarantees): when a commit lands, the writer pings replicas that might care, but the gossip is best-effort. Every replica independently verifies a read against S3 with a conditional GET (a read that fails if the object has changed since the writer last saw it) before it serves anything, so a dropped gossip packet cannot turn into a stale answer. The gossip is a hint, not a contract.

Repo-to-server routing is the third piece. Instead of a central database that says "repo X lives on server Y," Origin uses rendezvous hashing, a scheme where every server and every repo is hashed and the highest-scoring pair wins. If a server disappears, the repo is just materialized elsewhere on demand, because any server can rebuild it from the WAL.

An independent technical analysis from explainx.ai reports that Cursor's internal stress tests showed linear read throughput scaling up to 100 replicas, sustained push throughput around 120 commits per second on standard S3, and around 300 per second on S3 Express One Zone, AWS's higher-priced, single-region storage tier. Those numbers come from Cursor's own synthetic tests, not from independent benchmarks, and they describe one vendor's load, not the industry's ceiling. Spokes has run GitHub at production scale for thirteen years; one stress test on a hundred replicas does not invalidate that.

Cursor's stress-test result is one data point. The portable lesson is the tax itself: a way to test whether your Git host is paying it, and the design choices that flip it. On your own Git host, does push throughput fall as you add read replicas, and does every small agent-spawned repo carry a three-replica footprint you do not need? If both answers are yes, the design swap is worth reading on its own terms, as a vocabulary for asking whether your own stack is paying the same tax.

The same week, Block previewed Buzz Projects, a structurally different answer. Buzz is Nostr-based, meaning it leans on the same public relay network that powers the decentralized social protocol Nostr, and self-hosted. The two projects share a trigger: AI-generated code has outgrown the version-control systems that human-driven workflows were designed for. They do not share a solution. Origin changes the storage and consensus layer under Git. Buzz changes the identity and transport layer around it. Two projects, two structurally different approaches, shipped the same week. That is the signal: the Git platform layer is now being rebuilt across the industry.