feat: gateway fork — compute-driver-socket + credentials-driver-socket flags, OIDC

[pdettori]

Summary

Implements kagenti/kagenti#1353 — the gateway-side plumbing for the MVP gateway-per-tenant architecture.

• Cherry-pick OIDC auth (upstream PR feat(auth): add OIDC/Keycloak authentication with RBAC and scope-based permissions NVIDIA/OpenShell#935): OIDC/Keycloak JWT authentication with RBAC, scope-based permissions, and e2e tests (10 commits)
• --compute-driver-socket: New External compute driver variant that connects to a pre-existing Unix domain socket (for sidecar-based compute drivers)
• --credentials-driver-socket: Credentials driver client that delegates credential resolution to an out-of-process driver over gRPC/UDS, with credentials_driver.proto contract

New CLI flags

Flag	Env var	Purpose
--compute-driver-socket <path>	OPENSHELL_COMPUTE_DRIVER_SOCKET	Connect to external compute driver via UDS
--credentials-driver-socket <path>	OPENSHELL_CREDENTIALS_DRIVER_SOCKET	Connect to credentials driver via UDS

Proto contract

proto/credentials_driver.proto defines the CredentialsDriver gRPC service:

• ResolveCredential(name) → {token, expires_at_ms, token_type}
• ListCredentials() → [{name, scopes}]

This is the canonical location for the contract (gateway defines it, drivers implement it). See design discussion on the issue.

Architecture context

Part of the MVP implementation plan — gateway-per-tenant with sidecar compute and credentials drivers communicating over Unix domain sockets.

Test plan

• CI passes (cargo build, cargo test, cargo clippy, cargo fmt) — 1,269 tests, 0 failures
• OIDC e2e tests from PR feat(auth): add OIDC/Keycloak authentication with RBAC and scope-based permissions NVIDIA/OpenShell#935 pass — 10/10 passed on Kind cluster with Keycloak (RBAC, scopes, client credentials)
• ComputeDriverKind::External parses correctly — config::tests::compute_driver_kind_parses_external ... ok
• configured_compute_driver accepts External — tests::configured_compute_driver_accepts_external ... ok
• External compute driver errors when socket path missing — validated by configured_compute_driver + runtime check in build_compute_runtime()
• Credentials driver connects and resolves over UDS — integration test with mock driver (follow-up)

OIDC e2e test details (Kind cluster validation)

Validated on local Kind cluster with Keycloak (openshell realm), gateway deployed with OIDC enabled:

Test	Result
TestRbac::test_admin_can_create_provider	PASSED
TestRbac::test_user_cannot_create_provider	PASSED
TestRbac::test_user_can_list_sandboxes	PASSED
TestRbac::test_unauthenticated_request_rejected	PASSED
TestRbac::test_health_does_not_require_auth	PASSED
TestScopes::test_sandbox_scoped_token_can_list_sandboxes	PASSED
TestScopes::test_sandbox_scoped_token_cannot_list_providers	PASSED
TestScopes::test_openshell_all_grants_full_access	PASSED
TestScopes::test_no_openshell_scopes_denied	PASSED
TestClientCredentials::test_ci_token_can_list_sandboxes	PASSED

Fix applied: Updated Provider construction in tests to use ObjectMeta(name=...) wrapper matching current datamodel.proto schema (commit cbb79cf7).

Closes kagenti/kagenti#1353

Assisted-By: Claude (Anthropic AI) noreply@anthropic.com

[pdettori]

15 commits, all signed-off. Rust, Python, Markdown, and DCO checks passing; container builds and license-headers check still pending at time of review.

The OIDC auth implementation is thorough — well-tested (10 e2e tests), good separation of authentication (oidc.rs) from authorization (authz.rs), proper anti-spoofing header stripping, and clean handling of all three auth modes (unauthenticated, sandbox-secret, and Bearer). The new --compute-driver-socket and --credentials-driver-socket flags are minimal and well-scoped.

No blocking issues. Three comments inline (one suggestion, two nits).

[pdettori]

Suggestion: The comment on line 119 says audience "should match the server's --oidc-audience, NOT the CLI client ID" but the default value here is openshell-cli. This is self-contradictory and will confuse operators configuring a non-Keycloak OIDC provider.

Either drop the "NOT the CLI client ID" clause (if openshell-cli is intentionally the right default for the test Keycloak realm), or change the default to "" and update the comment to clarify when to use the CLI client ID vs a dedicated resource-server audience.

[pdettori]

Nit: The dummy TCP address "http://[::]:50051" is required by Endpoint::from_static but is silently ignored because the connect_with_connector closure overrides transport with the Unix socket. A one-line comment (// Dummy address required by tonic; actual transport is the UDS connector below.) would prevent future readers from wondering if this port is significant. Same applies to :50052 in credentials.rs.

[pdettori]

Nit: connect_uds duplicates the 100×100ms retry loop from compute/external.rs::connect — same structure, same last_error accumulation. Fine for MVP scope. Worth a follow-up to extract a shared uds::connect_with_retry(path, attempts, interval) helper so the retry policy lives in one place.