# Anthropic Scion adopts Mononoke's derived data model, says a Mononoke engineer

> An engineer who helped build Meta's Mononoke says Scion copies its derived data design. Here is what that does and does not tell you about Scion on large repositories and agent-scale writes.

- Published: 2026-10-15
- By: Samir Sengupta (https://www.samcodeman.com/about)
- Canonical: https://www.samcodeman.com/writing/anthropic-scion-mononoke-derived-data-model
- How this was made: Drafted with Claude, Anthropic's model from 1 cited source, then checked against it in a second editing pass; read and approved by Samir Sengupta on 2026-10-17. Editorial policy: https://www.samcodeman.com/editorial-policy
- Disclosure: I build Wayne, an AI coding harness for VS Code, and Omnitron, an agent harness, so I work in this category.

## The short version

- The only Mononoke design choice the sunshowers.io write-up shows Scion adopting is derived data. Both systems use it to make operations like blames more efficient.
- Hacker News commenters go further. One says Scion's data model is a direct copy of Mononoke's down to the terminology. Another reads Scion as copying both Mononoke and the EdenFS filesystem. Those are commenter readings, not claims made in the write-up.
- Neither source gives any figure for repository size, write throughput or agent-generated commit volume, so Scion's behavior under heavy agent writes cannot be inferred from them.

According to a sunshowers.io post published 2026-10-14, Anthropic Scion, Anthropic's new version control system, uses a concept called derived data to make operations like blames more efficient. Mononoke, the source control server the author helped build at Meta, uses the same concept under the same name. That is the one design choice the write-up documents directly. It gives no numbers for repository size, commit rate or write volume, and it says nothing about how Scion handles agent-generated commits.

The post's author places Anthropic's description, published the day before, next to their own 2022 description of Mononoke's derived data, written for their Oxide candidate materials. Anthropic's announcement, as quoted by a Hacker News commenter, names Google's Piper and Meta's Sapling as prior art. According to the author, Sapling was credited only for "hard-won lessons" and Mononoke was not mentioned at all. For engineers, the useful question is how far the shared lineage tells you anything about Scion's behavior. On the sources available, it tells you about one mechanism and nothing about performance.

## Derived data is the documented Mononoke inheritance

The write-up rests on a side-by-side comparison of two descriptions, not on a feature list. Both descriptions present derived data as the way to make operations like blames more efficient. The author concludes that the resemblance is more than convergent thinking.

> I think the level of imitation goes well beyond having similar insights
>
> - sunshowers.io, Anthropic's Scion is heavily inspired by Mononoke

Neither source explains how derived data works internally. They do not say when it is computed, where it is stored, or what happens when it is missing for a commit. Teams evaluating Scion will need those answers from Anthropic's own documentation or from the Mononoke code. According to one Hacker News commenter, the post's Mononoke link points to a subdirectory of the Sapling repository.

## Where Mononoke sits in the Sapling stack

Some Hacker News commenters were unsure how Mononoke relates to Sapling. One quoted the README: "Mononoke is the server built for the Sapling Source Control System". Another laid out three parts that together form the core of Meta's source control setup:

- Sapling: either the system as a whole or just the client, sl, which was forked from hg.
- EdenFS: the virtualized FUSE filesystem, closely modeled after SrcFS at Google.
- Mononoke: the distributed system that implements the backend server storage and the API for checkout-less writes.

The same commenter's reading is that Anthropic built its own version of the client, copied EdenFS and Mononoke, and built a separate code review tool. That is one commenter's interpretation, not a claim the sunshowers.io post makes. The post's documented comparison covers derived data only.

## What Anthropic Scion's Mononoke lineage implies for huge monorepos

The author dates the Mononoke team's vision to 2017-18 and writes that Mononoke's design "has stood the test of time." Anthropic positions Scion in the same category. Its announcement, as quoted on Hacker News, says: "Source control for huge monorepos has prior art, notably Google's Piper and Meta's Sapling, so we were able to build Scion with their hard-won lessons in mind." The sources do not say how large Scion's repositories are or how Scion's design performs at that scale.

One commenter, addressed in the thread as the post's author, said several more organizations are going to reach the scale Meta reached a while ago. They added that they are unsure whether those organizations will choose a monorepo or a polyrepo direction. Another commenter, wishing Sapling could be used outside Meta, listed what they value in it: a centralized model, a virtual file system, all commits automatically on the cloud, large binary files handled without trouble, and a Mercurial CLI. Neither source establishes whether Scion offers each of these.

## Agent-generated commits and checkout-less writes are not addressed

For teams running coding agents, the key question is how Scion handles a high volume of machine-generated commits. Neither source answers it. Neither gives commit rates, latency, repository size limits, or any account of agent workloads on Scion. The nearest relevant detail is one commenter's description of Mononoke exposing an API for checkout-less writes. The sources do not describe that API further, and they do not confirm that Scion has one.

The same gap applies to derived data on the write side. The sources do not say whether derived data is computed at commit time or on demand. That means the cost it adds per commit under agent-scale write volume is unknown. Until Anthropic publishes figures, the sources support no conclusion about Scion's write throughput.

## Is crediting Sapling enough to cover Mononoke

One objection on Hacker News was that Mononoke is a component of Sapling, so crediting Sapling for "hard-won lessons" arguably covers it. The author's position is that the borrowing goes beyond lessons, and that crediting only Sapling while omitting Mononoke "borders on plagiarism." They write that the post "comes off as Anthropic claiming that Claude designed all this." One commenter argued that the data model is a direct copy "down to the terminology." Another used a car analogy: crediting the Corvette is not enough if the new car's engine is a near copy of the Corvette's.

The author is explicit that copying itself is not the complaint. Mononoke was released as open source partly "to benefit humanity through the spread of good ideas." They also note that Mononoke drew on prior art, from Google's SrcFS to write-ahead logs in databases. The complaint is about credit. In the author's view, LLMs used to design new systems are likely to index heavily on existing designs. When asked about prior art, Opus 5.5 and GPT-6 Astra both immediately point out the similarities to Mononoke.

The author concludes that giving credit is now "easier - and more important - than ever." One commenter added a point aimed at operators. In their view, agents treat a prior-art search as a detour from their goal, so the responsibility for attribution has to stop with the people running them.

## Cursor's system has similar insights but different design

In a footnote, the author points to Vicent Marti's blog post for Cursor as an example of a system with similar insights but many key design differences. The distinction matters when evaluating Scion: sharing the derived data idea does not make two systems interchangeable. The sources do not list those design differences. A comparison of Scion, Mononoke and Cursor's system on write paths and storage therefore has to come from the systems' own documentation.

## Questions this raises

### What does Anthropic Scion share with Meta's Mononoke?

According to a Mononoke engineer's post, Scion uses derived data to make operations like blames more efficient, the same concept under the same name as Mononoke. That is the only design choice the post documents directly.

### How does Scion perform with large repositories and agent-generated commits?

The sources give no numbers for repository size, commit rate, latency or write volume. They also do not say whether derived data is computed at commit time or on demand, so its per-commit cost under agent-scale writes is unknown.

### Did Anthropic credit Mononoke as prior art for Scion?

No. Anthropic's announcement names Google's Piper and Meta's Sapling, crediting Sapling for hard-won lessons, but does not mention Mononoke. The author says omitting Mononoke borders on plagiarism, while stressing that copying itself is not the complaint.

## Key facts

- A sunshowers.io post dated 2026-10-14 compares Scion's derived data to Mononoke's, one day after Anthropic's 2026-10-13 announcement.
- The author's Mononoke derived data description dates to 2022, written for their Oxide candidate materials.
- The author dates the Mononoke team's vision to 2017-18.
- Anthropic's announcement credits Google's Piper and Meta's Sapling for prior art; Mononoke is not mentioned.
- Asked about prior art, Opus 5.5 and GPT-6 Astra both point out the similarities to Mononoke.

## Sources

1. [Anthropic's Scion is heavily inspired by Mononoke](https://sunshowers.io/posts/anthropic-mononoke/) - sunshowers.io

Discussion read (practitioners' reactions, not a source): https://news.ycombinator.com/item?id=50094694