agent.md files of your agents, and the artifacts of your models. The platform scans it to register what it declares, checks it out at a pinned commit to run a build, and commits back what a training writes.
A repository is identified by res.repository.…. The row on the platform keeps its identity; the URL, the provider and the token are kept by the connection service, encrypted, write-only.
Declaring a repository
credential_fields naming what is stored. PUT /repositories/{id}/credentials rotates the token wholesale; PATCH changes the rest; DELETE removes the repository, not the transforms that were scanned from it.
Which credential a build uses
Most repositories carry no token of their own. When the repository has none, the platform uses the organization’s git credential if the URL is on the organization’s git host, else the platform’s own for its internal host, else none, which is right for a public remote. The credential never leaves the connection service: a build receives a bundle of the repository, cloned locally with its history and tags but no remote, so nothing in the checkout can reach the origin.What a scan reads
A scan walks the files of a branch and registers, in the repository’s project:
A scan runs when the repository is saved, after each successful build of it, and on demand with
POST /orchestration/repositories/{id}/scan-lineage?ref=. A scan of the default branch is authoritative: it registers what it finds, takes over what was merged from a branch, and removes the transforms whose declaration is gone. A scan of another git branch adds and updates only what was born on that branch, tagged with it, so the default branch remains the single source of the lineage.
A dataset owned by a transform of another repository is never taken over: the scan reports it and moves on.
Builds and branches
A build checks out the repository at the commit resolved when the build starts and pinned on it, so the same build run again runs the same code. On a data branch, the build checks out the git branch of the same name when the repository has one, and the configured branch otherwise. See Builds for what the checkout receives. A training commits the model’s artifacts back into the repository, under the model’s directory, and that commit becomes the model’s version. See Running.Reading and writing files
The git endpoints read and change a repository through the platform, without a clone:
This is what the Atlas code editor and Klara’s
read_code and write_code tools use.
These endpoints are served by the platform and used by the Atlas interface; they are being opened on the public gateway progressively, and the Python library does not wrap them yet. The API reference lists what the gateway exposes today.