Configure sources and connections

Connect governed source systems and understand crawl, checkpoint, permission, and revocation behavior.

Source connections tell the worker which provider to crawl, what scope to use, where to publish, and how to authenticate. The generic administration surface is adapter-driven: installing an adapter contributes its profile, configuration fields, credential probe, scopes, and crawl source.

Supported adapters

The current repository includes Slack, Google Drive, and GitHub adapters. Manual upload remains a first-class source. A deployment offers only the sources whose adapters are installed; naming an unavailable source is a request error.

Connection journey

Select a source reported by the deployment and create a disabled connection for the current organization.

Configure the adapter-specific scope and a target Knowledge Space. Enabling a crawl with nowhere to publish is refused.

Submit the provider credential through the write-only credential field. It is encrypted at rest and is never returned by list, detail, test, or audit responses.

Run the adapter-owned credential probe. Confirm the remote workspace or installation identity without exposing the submitted credential.

Enable the connection and inspect the next crawl attempt, independent content, permission, and membership checkpoints, and any completeness reason.

Credential boundary

Credentials use an authenticated AES-256-GCM envelope in a separate ledger table. A row that fails its authentication tag is refused. The worker loads current configuration and credentials on each poll, so rotation or scope changes apply without a process restart.

The API reports whether a credential is present and whether its probe succeeds; it does not reveal the credential in any form. Every mutation records only that a credential changed.

Crawl completeness

The connector contract versions identity, membership, content, and permissions independently. Each component carries its own cursor and completeness state. Unsupported versions fail before work begins.

COMPLETE permission evidence may seal a new resource ACL generation.
COMPLETE membership may atomically advance a source-group head.
INCOMPLETE observations remain visible for health but cannot become the last successful authorization state.
a permissions-only refresh does not need to read document bodies;
a complete crawl may retire objects it no longer observes;
an incomplete crawl never treats absence as deletion.

Provider-specific trust

Slack maps messages and whole threads, stable users, and channel membership. Bot/deactivated accounts and Slack markup are removed. A channel that cannot be read is omitted rather than interpreted as an empty ACL.

Google Drive maps supported files and stable permission IDs. User, group, and domain grants are preserved; anyone-with-link does not grant public access. Incomplete search or unreadable sharing withdraws completeness.

GitHub admits organization installations and private repositories with Issues enabled. Issues and pull requests inherit a repository-reader group derived from effective collaborators. GitHub email is never trusted implicitly.

Verify a connection

Recorded-response and integration tests cover rate limits, pagination, credential redaction, restart resume, changed content, group revocation, tombstones, and incomplete evidence. A live Slack workspace, live Drive, and live GitHub production proof are not claimed by those tests.