ovstorage-user-authenticate-to-backend

Use when configuring credentials or driving interactive authentication for a direct ovstorage backend connection.

SKILL.md

GitHub

Authenticate to a Backend

Goal

Create or refresh credentials for a backend connection without leaking secrets into logs, route metadata, or tool output.

When to use this

Use this when adding a new direct-mode connection (file, http, omniverse-storage-service, s3, gcs, azure, huggingface, opendal, or any other backend kind), or when an existing connection reports AuthRequired, CredentialExpired, or an inactive auth_state_kind in ovstorage_doctor.

Recipe

  1. Inspect the configured backend layers and connections with ovstorage-user-getting-started.
  2. Read the backend descriptor and credential schema before prompting for anything secret. Stack callers get this from the root Layer's list_kinds; CLI users can use the connect wizard.
  3. For file, no external credential is normally required. Register the root path and keep credentials empty.
  4. For http, prefer public URLs where they exist. If an HTTP route needs credentials, pass them through the connection credential schema — bearer_token, username + password, signed_query, or secret_headers — never through the address string; a query in root_url or prefix is rejected as InvalidArgument at instantiate.
  5. If authenticate_connection returns an auth event stream, surface OpenBrowser, DeviceCode, Progress, Succeeded, Failed, and Cancelled events to the user exactly as events, then stop on a terminal event.
  6. If it returns Unsupported, no flow ran and the connection's state is unchanged, so keep the registration and report the state rather than treating it as a failure. Backends that always answer this way: s3, gcs, azure, huggingface, opendal; broker connections on a direct endpoint (any address that is not http(s)://); and file and http, which have no connection-auth driver at all. Unsupported can also come from a layer that does not implement the call — a credentialless visibility-override row, or a remote service that has not implemented the RPC — so read it as "no flow was offered here", not as a statement about the backend kind. Where a credential is what is missing, supply one the origin accepts; re-running the flow is not the fix.

Backend-specific auth

omniverse-storage-service

The omniverse-storage-service backend authenticates over OIDC when its address names a discovery root. The plugin reads /api/v1/auth-config from that root to learn the OIDC issuer and client; the standard OIDC discovery doc supplies the auth / token / device endpoints.

Two interactive flows:

  • Browser (InteractiveAuthCapability::Browser): PKCE with a loopback redirect. The plugin opens the user's browser to the authorization endpoint, captures the code at a transient localhost callback, and exchanges it for the token bundle.
  • Headless (InteractiveAuthCapability::Headless): RFC 8628 device flow. The plugin requests a device code, surfaces a user-code + verification URL via AuthEvent, and polls the token endpoint until the user completes the flow on a separate device.

In both cases the resulting OAuthToken lands as the connection's oauth credential bundle. The plugin proactively refreshes the access token in a background task at ~90% of the token's TTL using the refresh-token grant (or the client-credentials grant if that's the configured method). No intervention required for long-running processes.

To pre-seed credentials (e.g. for CI), populate oauth with a long-lived refresh token; the plugin will exchange it for an access token on first use.

When address names a grpc:// or grpcs:// endpoint instead, there is no discovery root and therefore no OIDC configuration: no sign-in flow, no client-credentials grant, and no background refresh. Such a connection takes an access token the host already holds — put it in the oauth field with no refresh token — and the host rotates it by calling update_connection_credentials again with a fresh one. An empty bundle removes it.

s3

The descriptor exposes three credential methods:

  • static_key — long-lived IAM user key. Populate aws_access_key_id + aws_secret_access_key.
  • session — short-lived STS / SSO credentials. Populate aws_access_key_id + aws_secret_access_key + aws_session_token.
  • aws_credentials_file — read from an INI section. Populate file_path (defaults to ~/.aws/credentials) and profile (defaults to default); the per-connection profile config key picks the section.

Each credential field's descriptor default is the matching ${AWS_*} env-var placeholder (e.g. ${AWS_ACCESS_KEY_ID}); the host's bundle-resolution layer expands these before the bundle reaches the plugin. From the plugin's point of view the resolution chain is bundle fields → env vars (via placeholder expansion) → shared credentials file. IMDS, AWS SSO, web-identity tokens, and STS role assumption are not implemented in-process; operators that need them run their session helper (aws sso login, etc.) to materialise a static-key profile.

gcs

The descriptor exposes two credential methods:

  • service_account_key — paste an ADC JSON blob into the service_account_key field. The plugin auto-detects the JSON type: service_account triggers JWT-bearer exchange (and is the only path that can mint V4-signed read URLs); authorized_user triggers an OAuth2 refresh-token exchange; external_account runs the workload-identity STS exchange (the subject token is re-fetched from its file or url source on every exchange, with an optional impersonation hop; token and impersonation endpoints must be https *.googleapis.com or loopback, the subject-token url https or loopback, and workforce-pool configs are rejected); impersonated_service_account mints its source credential's bearer and spends it on generateAccessToken. Every type except service_account reads through the plugin (alt=media) because only an RSA private key can V4-sign.
  • gcloud_adc_file — populate file_path with the path to an ADC JSON file of any of those four types — e.g. written by gcloud auth application-default login (optionally with --impersonate-service-account) or by gcloud iam workload-identity-pools create-cred-config (defaults to ~/.config/gcloud/application_default_credentials.json, with ~ expanded against HOME / USERPROFILE + HOMEDRIVE).

The plugin does not consult GOOGLE_APPLICATION_CREDENTIALS or any other env var; if you need env-var expansion, do it in the host's bundle-resolution layer before handing the bundle to the plugin. GCE/GKE metadata-server identity and the AWS / executable external_account source flavors are not implemented in-process.

http

The descriptor exposes four credential methods. Distinct channels combine — for example, a connection may hold a signed query and secret headers at once — and a connection with none of them is anonymous:

  • bearer — opaque token sent as Authorization: Bearer <token> on every request. Populate bearer_token.
  • basic — HTTP Basic authentication. Populate both username and password; either value may be empty, but not both.
  • signed_query — pre-issued query string appended verbatim to every request URL (leading ? optional). Populate signed_query, and set the signed_query_scope config key to declare the token's scope family.
  • secret_headers — credential-bearing headers in wire form, one Name: Value per line, sent with every request. Populate secret_headers. The separator is a line break, not a comma, so a value may contain commas. This is the secret-bearing counterpart to the default_headers config key, which rejects Authorization, Cookie, and Proxy-Authorization; those names are accepted here, while Host, Range, If-Match, and the framing and hop-by-hop headers are refused.

A signature is a credential with a declared scope, not part of the address. Writing a query into root_url or prefix returns InvalidArgument at instantiate: the object key is appended after the route prefix, and a signature buried in config can be neither scoped nor rotated. signed_query_scope accepts prefix (the token authorizes everything under root_url — an Azure account, container, or directory SAS, a CloudFront signed URL with a custom policy) or object (a per-object presign, which a connection cannot hold and which returns Unsupported; dispatch such a URL as a per-request address). The plugin also refuses a token whose parameters name a per-object signature — X-Amz-Signature, sr=b, a CloudFront canned policy — while signed_query_scope says prefix. Supplying signed_query with no scope, or a scope with no signed_query, is InvalidArgument; so is setting Basic or Bearer credentials together with a secret_headers entry named Authorization. Userinfo in root_url is the legacy Basic channel and likewise conflicts with another Authorization writer, while it may coexist with a signed query or non-Authorization secret header.

All channels are hot-swappable through update_connection_credentials while the connection is live, but the rotation may not change the connection's exact shape: the Authorization method, presence of a signed query, and secret-header names and multiplicity must remain fixed. Remove and re-add the connection to change that shape. signed_query_scope is configuration rather than a credential, so rotating into a signed query without that key (or supplying the key with no token) is InvalidArgument. A replacement is probed before the atomic swap; reads keep using the old snapshot while the probe runs. Credentials over a plaintext http:// route are refused except on loopback, and any secret-bearing connection refuses an https:// to http:// redirect downgrade.

azure

The descriptor exposes four distinct credential methods. The plugin resolves them in this order:

  • account_key — Shared Key signing. Populate account_key.
  • sas_token — pre-issued shared-access signature appended verbatim to request URLs. Populate sas_token.
  • workload_identity — Entra OAuth2 federated workload identity. Populate federated_token_file + client_id + tenant_id.
  • service_principal — Entra OAuth2 client-credentials. Populate client_id + client_secret + tenant_id.

Each credential field's descriptor default is the matching ${AZURE_*} env-var placeholder (${AZURE_STORAGE_ACCOUNT_KEY}, ${AZURE_STORAGE_SAS_TOKEN}, ${AZURE_TENANT_ID}, ${AZURE_CLIENT_ID}, ${AZURE_CLIENT_SECRET}, ${AZURE_FEDERATED_TOKEN_FILE}); the host's bundle-resolution layer expands these before the bundle reaches the plugin. The plugin itself never reads AZURE_* env vars at runtime. Managed Identity (IMDS), az login, VS Code, and PowerShell credential sources from DefaultAzureCredential are not implemented in-process.

huggingface

A single static access token in the token field (hf_…; a read-scoped token is sufficient). The descriptor default is ${HF_TOKEN}, resolved by the host at TOML load — the plugin never reads the environment itself. An empty bundle connects anonymously, which serves public repositories; private and gated repositories answer 404/403 without a token. The token is verified once at add time with a lenient whoami check (only a definitive 401 parks the connection); rotation goes through update_connection_credentials, and reauth answers Unsupported because no interactive flow exists.

opendal

Auth depends on the chosen service (one of fs, s3, webdav — the only services compiled in). fs requires no credentials. The OpenDAL s3 driver wants access_key_id + secret_access_key (bare names, not aws_*-prefixed — these are NOT interchangeable with the first-party s3 plugin's credential shape). webdav uses HTTP Basic credentials (password, plus username passed through config_json). The adapter accepts only SecretValue::Bytes, SecretValue::OAuthToken, and SecretValue::File for these fields; SecretValue::SystemIdentity and SecretValue::MtlsCertPair are rejected as Unsupported. Ambient-identity flows (Kerberos ticket cache, etc.) are not threaded through. See plugin-opendal.md for per-driver detail.

nucleus

Omniverse Nucleus's auth shape predates the unified OAuth pattern used by the Omniverse Storage Service and the cloud backends. The descriptor exposes three user-selectable credential methods:

  • sso — single sign-on (browser). Recommended. Binds no credential fields. The plugin (establish_interactive_auth) calls Tokens::subscribe to register a one-shot login intent, receives a server-generated nonce, and surfaces AuthEvent::OpenBrowser with the Nucleus login URL annotated with ?nonce=<nonce>. The host UI opens the URL; once the user completes the form, the second response on the same subscription carries the authentication result. The plugin never touches raw passwords or 2FA challenges on this path.
  • userpass — username and password. Binds username and password; drives Nucleus's Credentials::auth flow. For service accounts and headless deployments.
  • api_token — API token. Binds a single api_token credential. Drives Tokens::auth_with_api_token; takes precedence over username / password. The Auth envelope's username is recorded as the authenticated principal. The recommended automation path.

Successful authentication via sso produces a refresh token that the plugin persists in the OS keyring keyed by the server hostname (not the host-minted, unstable ConnectionId). On subsequent connects, warm-continuation reads refresh_token from the keyring, rediscovers tokens_url, and exchanges the refresh token without re-prompting; on a service-call Expired / InvalidToken, the plugin invokes Tokens::refresh once per token-generation epoch (concurrent callers share a single refresh under a per-shared mutex). AuthRequired / AuthExpired / PermissionDenied clear the keyring entry; Transient preserves it for retry. SSO (SSO::*) and DeviceFlow (DeviceFlow::*) are declared as target / deferred flows; they are not wired today.

broker

The broker plugin authenticates to an upstream ovstorage-broker daemon; the broker itself holds the long-lived cloud credentials and runs its own authn substrate against whatever listener-authn mode operators configured. From the calling host's perspective the client-side credential surface is intentionally minimal:

  • client_credentials — OIDC client credentials grant. Binds client_id + client_secret. Only valid when the configured address is a discovery URL (not a direct gRPC endpoint): the plugin fetches /api/v1/auth-config, follows the published openid_configuration URL to the IdP's standard OIDC discovery doc, and drives a non-interactive client-credentials grant. Suited for service accounts and headless deployments.

Interactive flows (PKCE loopback for Browser, RFC 8628 device flow for Headless) are driven by Factory::authenticate against the same broker-published auth-config. The host signals which mode is available through the gRPC x-ov-iauth: browser|headless|none metadata header attached by the broker-client SDK on every RPC; the broker reads this on its streaming Auth RPC, threads the parsed InteractiveAuthCapability into the upstream-OAuth flow, and emits AuthEvent::OpenBrowser / DeviceCode / Progress / Succeeded back to the host. None mode causes the broker to emit a terminal Failed { AuthRequired } envelope without opening a browser tab or device prompt — the fail-fast shape render workers and CI hosts need.

Three-tier OAuth, broker-held. Successful interactive auth lands a refresh token on the broker via the unary RegisterCredential round-trip; the broker persists it through its OAuthCredentialProvider

  • secret_tokens row + SecretStore (per-(BackendId, PrincipalView) slot). The host process holds only the short-lived broker bearer — never the upstream cloud credentials. Workers and a coordinator that share a PrincipalView at the broker hit the same cache slot, which covers the multi-tenant SaaS and render-worker shapes without any host-side coordination.

The broker daemon's listener authn mode (jwt_verify, trusted_unsigned_jwt, trusted_forwarded_headers, peer_cred, or the reserved mtls) decides how the host's bearer is validated — that's operator-owned config and is not part of the broker-client-side credential schema. See docs/public/broker-operator/README.md for the daemon side.

References