功夫 · Dated record · 2026-10-14

Convergence: KungFu and Scion

On 2026-10-13 Anthropic published Redesigning source control at Anthropic, describing Scion, the system that replaced git for their monorepo. Much of its diagnosis, and several of its design choices, match what this project has said publicly since 2026-05-10. This page records where the two converged, the evidence for when, and what is not being claimed.

The claim, stated precisely

KungFu reached the same diagnosis independently and published it five months earlier.

The two dates are not the same kind of evidence.

First publicVerifiable by
KungFu2026-05-10GitHub's server timestamps, available to anyone
Scion2026-10-13The article's publication date

Scion's article describes internal work before that: a chart beginning in January 2026 and an ingestion prototype written "earlier in the year." That is Anthropic's own account of private work, and it cannot be checked from outside. This page makes no claim about when Scion began internally.

It claims only what the public record shows: KungFu was public first, by five months, and could not have drawn on Scion, which was not public. That is the more useful fact anyway. A system built and run at Anthropic's scale arriving at the same conclusions is evidence the conclusions are right.

The evidence

Every date links to the record it comes from.

Server-timestamped set by GitHub

These timestamps are set by GitHub, not by the author, and cannot be backdated. GitHub records an edit timestamp on issue bodies; none of these has ever been edited.

WhatWhen (UTC)Relevance
Repository k8nstantin/kungfu created, public2026-05-10 17:39:15The anchor
Issues #1–#52026-05-10 20:54–20:55Central sync server, binary files, intent dashboard
PR #8 opened and merged2026-05-11 20:15 / 20:32Agent coordination through the shared code
Issue #92026-05-12 12:31:47kf ingest: importing existing git projects
Issue #112026-05-12 16:28:41Disposable server; cloud object storage as the source of truth
Issue #132026-05-17 11:31:26Block-level convergence over parsed code

Commit-dated set by the author

Commit dates are set by the author, and history was rewritten on 2026-06-26 (eef8ae5) to purge a committed key, so the current hashes postdate that rewrite. These are supporting evidence; the server timestamps above are the anchor.

IdeaFirst appearsDate
Agent-first, operation-based, a write-ahead log instead of snapshotsb3102e22026-05-10
Trunk-only, no branches6238e142026-05-10
One monorepo; "agents do not clone"27d24822026-05-10
Access control on logical boundaries inside the monorepo27d24822026-05-10
A continuous verification gate replacing branch-based safety27d24822026-05-10
"Agents write, humans review"829e6642026-05-10
Hot buffer plus cloud object storage as the durable ledger004168a2026-05-12
Lock contention on branch refs as the failure pointf0639692026-05-14
Tiered history: recent locally, deep history fetched centrally033fac92026-05-14

Where they converged

The same diagnosis, and many of the same design choices.

KungFuScion
The diagnosisGit was built for humans exchanging patches and offline trust; agents break its assumptions.Git was built for distributed, offline clients; nearly all of theirs are always-online agents and jobs.
The bottleneckLock contention on branch refs, repository congestion.No global merge lock; commits are lock-free; a ref update is a single compare-and-swap.
TrunkTrunk-only. Named releases kept.Refs discouraged except main and release branches.
One repositoryNative monorepo; agents work without cloning.One monorepo; source control exposed as an API so jobs need no local copy.
Changes, not snapshotsAn attributed operation stream with a write-ahead log.A changeset is a write-ahead-log entry holding changes relative to its parent, with no tree pointer.
Derived stateBlocks, paths, review state, deploy state and last green are derived, never declared.Blame, path history and graph indices are derived data, invisible to clients.
TopologyPer-machine daemon, central hot path, cloud object storage underneath.Per-node daemon, per-region cache, authoritative server on a database plus cloud storage.
Access inside one repoNarrow grants for sensitive code instead of exiling it to a separate repository.Per-directory access control to protect sensitive code within the monorepo.
ReviewThe reviewer judges whether the change is the right thing to do.Review judges the intent of a change rather than every line.
SafetyBranches were never the safety mechanism; the pipeline is.No global merge-queue lock; post-merge and pre-deploy automated testing.
Migrationkf ingest (issue #9); replaying real history is the first piece of work.A git-to-Scion ingestion pipeline.
RestraintThe clean cut (db74a9c) and the third law: a feature that only makes sense in a story about the future of software is not core.Much of the engineers' work was telling the model no when it over-engineered for hypothetical concerns.

Where they did not

These are as informative as the agreements.

Online versus local-first

Scion assumes clients are always online: a thin client, a bounded cache, authority on the server. KungFu's local daemon owns the state and converges offline.

The unit of change

Scion keeps commits and merges, made cheap and moved server-side. KungFu removes them in favour of operations and derived blocks.

Conflicts

Scion lets the model repair them. KungFu proposed exactly that early on (c349feb, 2026-05-14) and then rejected it: model-based resolution is a suggestion, not a gate. Convergence may produce garbage inside a draft; garbage cannot promote.

Virtual filesystems

Scion materialises files on demand. KungFu rejected virtual filesystems (3c271c4, 2026-05-14), on the grounds that the only one that works runs on a managed fleet, which is what Scion's clients, mostly CI and agent jobs, are.

When the gate runs

KungFu's machine gate runs before work is accepted. Scion's post says they could not make a global merge-queue lock work, and tests after the merge instead.

When work becomes visible

In Scion, from the moment of commit. In KungFu, from the first keystroke, as a draft.

Scope

Scion says plainly that it is not a general git replacement and hopes its ideas flow back into git. KungFu is meant for everyone, including teams still on SourceSafe.

What neither invented

Most of the shared mechanisms predate both projects.

Google's Piper and CitC

  • One monorepo and trunk-based development
  • Per-directory permissions
  • Virtual working copies
  • Already cited in KungFu's README

Meta's Mononoke and Sapling

  • Server-side changesets stored relative to their parents
  • Derived data such as blame, computed on the server

What is newer is the reason: agents, not humans, are now the primary client of source control, and that changes which of git's assumptions still hold. That is the part both projects arrived at in 2026.

What it means for KungFu

Two claims corroborated, one challenged, one question still open.

  1. The diagnosis is corroborated. A team with Anthropic's resources and a ten-fold rise in merge volume reached the same conclusions about git, trunks and monorepos.
  2. "The pipeline is the real gate" is corroborated. They did not use a global merge-queue lock, and rely on post-merge and pre-deploy testing instead.
  3. "You cannot break the build" is challenged. Two drafts can each pass the machine gate against trunk and fail together. Re-validating every draft against a moving trunk is a merge queue, the thing Scion's post says they could not make work. The answer to dev, qa and prod on one trunk has to account for this.
  4. Release pointers survived. Even the most trunk-committed design kept release branches.
  5. The open question is still open. The post describes no CRDT, so it says nothing about convergence at repository scale. The replay measurement remains KungFu's to do.