open_scion puts agent branch permissions and server-side landing into a first public release

Give each agent a place to land changes

open_scion has published version 0.1.0, its first public release of a Git-compatible server aimed at agent workflows. The independent MIT-licensed project forks Gitaly and adopts ideas from Anthropic’s public source-control writing. It is neither an Anthropic product nor a reproduction of the company’s private Scion system.

Its Graft extensions move several operations toward the server: clients can fetch selected objects, use sparse workspaces and request history or search results remotely. The project describes UserLand as rebasing and landing a change with a compare-and-swap update, with concurrent landing requests on a branch batched together. These are the project’s documented capabilities, not an AITrending throughput test.

The quickstart makes the permission model tangible. A planner and a builder receive separate tokens. The planner can land on main; the builder can land on a feature branch but receives a policy refusal on main. Agent requests use their token identity for these decisions in standalone mode. Shared administrative credentials have broader authority and should not be confused with an agent’s restricted token.

A server and CLI, with AI kept optional

The build requires Linux or macOS, Go 1.26 and Gitaly’s build dependencies. make build compiles the binaries and bundled Git, and examples/graft/quickstart.sh runs the local server, cache and two-agent demonstration in a temporary directory. This is infrastructure for an application or harness to use, rather than a plugin that installs into an existing coding assistant.

The server itself does not call a language model. Conflict resolution is off unless an operator configures it and a client requests it. An external resolver can use deterministic strategies or an optional model provider; provider credentials and any inference charges are separate from the MIT software. The resolver documentation warns that sending a conflict to a model shares the file versions, including any secrets they contain, with that provider.

First-release limits still matter

The project explicitly lacks a hosting UI, issue tracker and CI service. Operators remain responsible for protecting configuration and tokens, creating repositories and isolating verification commands. Interfaces may change before 1.0, storage still lacks pack garbage collection and checkpoint compaction, and the release says real S3, GCS and Azure services have not yet been tested.

AITrending read the release and tagged English documentation but did not deploy or benchmark the server. The useful new option is specific: a community project now packages agent identities, branch policies and server-side landing into a public release that developers can examine and try.