跳转至

Console and authentication

Operator surface (/console/*), permission catalog and roles, system settings, and the two authentication modes. For why openKMS splits identity, operations, and data access, see Security design. For the data security model (RBAC vs resource ACL, sharing, inheritance), see Data security.

Console (admin)

  • Entry: header user menu Console (when outside console routes) opens /console for users with the admin role, permission all, or console:access. Inside /console/*, the header shows Exit Console back to home (/). The console shell and route gate still use canAccessPath against frontend_route_patterns from GET /api/auth/permission-catalog for individual pages (plus feature toggles where applicable).
  • Console overview (/console): Introduces console sidebar tools only—system health, permissions, data security, data sources (each tagged by layer — data or ontology), Kubernetes clusters, object storage, users & feature toggles, settings; cards link when canAccessPath allows; quick links to Permissions and Access groups when those permissions apply; optional nudge when the catalog still has only all (until openkms_permissions_onboarding_dismissed is set).
  • System health (/console/health): requires console:access (or admin). GET /api/admin/health-status probes core dependencies: API (implicit), PostgreSQL (SELECT 1), object storage (S3/MinIO head_bucket when configured; otherwise skipped), background jobs (Procrastinate procrastinate_jobs table present), Ontology Function Service (GET {OPENKMS_ONTOLOGY_FUNCTION_SERVICE_URL}/health), and Langfuse when LANGFUSE_BASE_URL is set on the backend: GET {base}/api/public/health (skipped if unset; skipped with a short note if LANGFUSE_HEALTHCHECK is false). Tracing credentials on the server (LANGFUSE_SECRET_KEY / LANGFUSE_PUBLIC_KEY) are not sent to the browser; the detail text only states whether they appear configured. When the caller also has console:data_sources, each registered data source is connection-tested (PostgreSQL / Neo4j; same logic as POST /api/data-sources/{id}/test). The page shows overall status, per-component latency, last-checked time, and a Refresh control in the page header (top right). Route/API patterns for console:access include /console/health and GET /api/admin/health-status (migration r8s9t0u1v2w3 refreshes defaults). Public GET /health (no auth) remains a minimal liveness probe for load balancers and smoke tests—not the Console page.
  • Permission management (/console/permission-management): Permission catalog is stored in security_permissions; the page loads rows from GET /api/admin/security-permissions (includes id for edit/delete). Under Roles, All selects catalog-only mode (add/edit/delete permission rows). Choosing a named role shows checkboxes to draft which keys that role receives; Save role permissions calls PUT /api/admin/security-roles/{id}/permissions once—no auto-save on each toggle. Switching roles with unsaved changes prompts to discard. Migrations seed all when the catalog table is empty and backfill default pattern rows for every hinted operation key (a2b3c4d5e6f7); admins may add keys via Add permission, Add missing suggested keys (from operation_key_hints on GET /api/admin/permission-reference), or POST /api/admin/security-permissions, using the in-page Route & API reference (and Operation keys tab) for path patterns. Roles may only assign keys that exist in security_permissions. The built-in all row cannot be edited or deleted. Migration seeds the admin role with all; member is created on first non-admin local sign-in (also starts with all). You cannot remove all from a role that still has only all in one step—add another permission, save, then remove all. Local: user_security_roles synced from is_admin. OIDC: JWT realm_access.roles match security_roles.name; realm admin bypasses permission checks.
  • Data security (/console/data-security/issues, /console/data-security/groups, /console/data-security/groups/:groupId): requires console:groups. Issues lists sharing misconfiguration with inline fix; Access groups uses a master–detail layout (group list, members tab, shared-access tab). Legacy /groups/:id/members redirects to /groups/:id. Per-resource sharing is documented in Data security.
  • Data Sources (/console/data-sources): console:data_sources (or admin). Kubernetes clusters (/console/kubernetes, /console/kubernetes/{id}): console:kubernetes — register existing clusters with encrypted kubeconfig, test connectivity, and browse namespaces / Deployments / Pods (read-only); see Kubernetes clusters. Object storage (/console/storage): console:storage — paginated bucket browser (S3 list_objects_v2); create folder placeholders; multi-select move prefixes/objects; metadata only (no presigned download). Connectors (/connectors): connectors:read / connectors:write — see Connectors. Datasets (/datasets): data:write to register/ingest and manage; data:read to view (filtered by dataset ACL); data:edit to edit row data in editable datasets. The data layer (datasets / data-model) is gated only by data:*; the Data Insights has its own insights:read / insights:write (a story is not the same grant as the data behind it — and a dataset-backed finding is still checked against that dataset's ACL), separate from ontology:* and console:*. Object types / link types (/ontology-manager/object-types, /ontology-manager/link-types): console:object_types / console:link_types or ontology:read / ontology:write as applicable; System Settings, Users & Roles, Feature Toggles remain console:* (or admin).
  • Users & Roles (/console/users): requires console:users. Local auth: list users, toggle is_admin (syncs security role links), delete/add users. OIDC auth: read-only list of oidc_identities (users who have signed in); roles remain in the IdP.
  • System settings (/console/settings): console:settings (or admin). GET /api/system/settings / PUT /api/system/settings load and persist system_settings (Postgres singleton row): system_name, default_timezone, api_base_url_note (optional note only; SPA API URL remains build-time). A PUT whose trimmed system_name would be empty is stored as openKMS. GET /api/public/system returns { "system_name" } without authentication (strict middleware allowlist); the value is the trimmed DB field and may be "". The sidebar title stays blank until that response arrives, then shows openKMS when the name is empty or whitespace (otherwise the returned name); on fetch failure it shows openKMS. After saves, notifySystemSettingsUpdated triggers the same fetch (custom event). Migration seeds row id=1 and attaches GET/PUT patterns to console:settings for strict API enforcement.
  • Feature toggles: evaluations (default off), connectors (default on), agents (default on), and media (default off) – persisted in PostgreSQL (feature_toggles table), shared across all users/devices. When agents is off, the SPA hides /agents and /projects/*, and /api/projects/* plus /api/user/git-credentials return 404. When media is off, the SPA hides /media/* and /api/media* plus /api/media-channels* return 404.
  • GET /api/feature-toggles (authenticated) returns current toggle state
  • PUT /api/feature-toggles requires console:feature_toggles (or JWT admin)

Permission catalog (canonical keys)

Defined in backend/app/services/permissions/permission_catalog.py; seeded into security_permissions by migration. Roles may only assign keys that exist in the catalog.

Family Keys
Console console:access, console:users, console:groups, console:permissions, console:settings, console:feature_toggles, console:data_sources, console:kubernetes, console:storage, console:datasets, console:object_types, console:link_types
Connectors connectors:read, connectors:write
Documents documents:read, documents:write
Document channels channels:read, channels:write
Articles articles:read, articles:write
Knowledge bases knowledge_bases:read, knowledge_bases:write
Evaluation evaluation:read, evaluation:write
Wiki wikis:read, wikis:write
Ontology ontology:read, ontology:write
Data layer data:read, data:write, data:edit (Datasets / Data Model apps; data:edit = edit dataset row data)
Media media:read (browse the library), media:write (upload, edit, generate)
Data Insights insights:read (open the app, read findings, re-run one), insights:write (create/update/delete stories) — plus the bound dataset's ACL
Knowledge map knowledge_map:read, knowledge_map:write
Agents projects:read, projects:write (projects, skills registry, agent chat; feature toggle agents)
Catch-all all (built-in admin key; cannot be edited or deleted)

Authentication

  • OIDC mode (default): any OIDC IdP – Authorization Code + PKCE in browser (oidc-client-ts); RP-initiated logout when the IdP exposes end_session_endpoint. On login, oidc_identities stores sub → preferred_username / email from the access token (callback, POST /sync-session, GET /api/auth/me) for sharing UI and ACL alias resolution—no IdP Admin API and no personal API key required.
  • Local mode (OPENKMS_AUTH_MODE=local): sign-up when OPENKMS_ALLOW_SIGNUP (exposed as allow_signup on GET /api/auth/public-config); sign-in with username or email + password; users stored in PostgreSQL; HS256 JWT + session cookie; no built-in admin password (first signup only becomes admin). The UI uses public-config so it stays aligned with the server even if VITE_AUTH_MODE differs.
  • openkms-cli: OIDC client credentials (Bearer) or, in local mode, HTTP Basic (OPENKMS_CLI_BASIC_*)
  • Worker / scheduler (heartbeats on /internal-api/process-heartbeat): separate OPENKMS_WORKER_OIDC_* or OPENKMS_WORKER_BASIC_* on the backend and those processes (client id must be in OPENKMS_INTERNAL_SERVICE_CLIENT_IDS for OIDC)
  • Profile (/profile): authenticated users see display name, email (if present), administrator yes/no, realm roles, and resolved permissions (local users: DB keys such as all or granular console:*; OIDC IdP admins receive the full catalog); data from GET /api/auth/me. Linked from the header user menu.
  • Settings (/settings): API keys — create, list, and revoke personal API keys for assistants and scripts (POST / GET / DELETE /api/auth/api-keys); the full secret is shown only once at creation. Linked from the header user menu (Settings).
  • OIDC + API traffic: the token provider used by getAuthHeaders() returns the access token only (no setUser / no POST /sync-session + GET /api/auth/me on every token read). Session sync and profile refresh run from OIDC init and UserManager user events. The SPA permission-catalog effect depends on a stable username / roles / permissions key so equivalent user objects do not refetch in a loop.
  • Idle / expired session on API calls: authAwareFetch treats structured 401 auth failures (see architecture Invalid JWT on API calls) as recoverable: one silent signinSilent + sync-session (OIDC) or cookie /me check (local), then a single retry with refreshed headers; if that still fails, the app clears session, shows a short toast, and sends the user to sign-in again (local /login, OIDC interactive redirect).
  • Protected routes: under MainLayout, all except home (/) require auth; /login and /signup are separate routes. Unauthenticated users on / see the static marketing home; on any other path they see "Authentication Required". Authenticated users without JWT admin / all must match the union of frontend_route_patterns from GET /api/auth/permission-catalog for their keys (paths / and /profile are always allowed); otherwise "Access denied" with a link home.
  • GET /api/auth/public-config (unauthenticated): auth_mode and allow_signup only—no secrets—so the SPA matches deployed auth mode.
  • Backend accepts Authorization: Bearer, session cookie (after POST /sync-session in the browser), and in local mode Authorization: Basic for CLI (minted local-cli service JWT).
  • Strict API patterns (optional): OPENKMS_ENFORCE_PERMISSION_PATTERNS_STRICT — when true, authenticated /api/* must match security_permissions backend_api_patterns unless bypassed (all, realm admin, local-cli, or allowlisted paths). See Configuration.

Obtaining an API token

Docs examples use Authorization: Bearer $TOKEN. How you get $TOKEN depends on OPENKMS_AUTH_MODE:

  • Local mode — POST /api/auth/login with JSON { "login", "password" } returns access_token (HS256). Prefer env vars + jq over embedding the password in the shell:
TOKEN=$(curl -sS -X POST "${API%/}/api/auth/login" \
  -H 'Content-Type: application/json' \
  -d "$(jq -n --arg login "$OPENKMS_LOGIN" --arg password "$OPENKMS_LOGIN_PASSWORD" \
        '{login:$login,password:$password}')" \
  | jq -er '.access_token // empty')

Default lifetime: OPENKMS_LOCAL_JWT_EXP_HOURS (168 hours). Long-lived tokens are convenient for dev scripts but widen blast radius if leaked.

  • OIDC mode — use the IdP user access token as Authorization: Bearer … (verified via JWKS). Operation permissions for non-admin users come from realm_access.roles matched to security_roles.name—the token must carry those roles. /internal-api/models/... accepts only sub=local-cli or OIDC clients in OPENKMS_INTERNAL_SERVICE_CLIENT_IDS; human tokens and personal API keys get 403.

  • CLI (local mode) — OPENKMS_CLI_BASIC_USER / OPENKMS_CLI_BASIC_PASSWORD → Authorization: Basic … → internal local-cli JWT. Dev/trusted networks only.

  • Worker / scheduler — OPENKMS_WORKER_BASIC_* (local) or OPENKMS_WORKER_OIDC_* (OIDC client credentials); same rules as CLI for /internal-api (local-cli or allowlisted azp).

  • Personal API keys — Settings → POST /api/auth/api-keys; plaintext okms.{uuid}.{secret} shown once. Local: permissions follow user_security_roles. OIDC: roles are snapshotted at creation—recreate the key after IdP role changes.

GET /api/auth/me confirms the token and lists resolved permission keys. Hardening themes: Tech debt — API tokens.

Credentials (do not commit)

Variable Purpose
OPENKMS_OIDC_CLIENT_SECRET OIDC confidential client
OPENKMS_SECRET_KEY Session cookie + local JWT signing
OPENKMS_CLI_BASIC_PASSWORD CLI Basic secret (local mode)
OPENKMS_WORKER_BASIC_PASSWORD / OPENKMS_WORKER_OIDC_CLIENT_SECRET Worker + scheduler internal API auth
AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY S3/MinIO — env only, never CLI args
OPENKMS_DATABASE_PASSWORD PostgreSQL

Full list: Configuration. Use .env.example as a template; keep .env out of git.

Production checklist

  1. Strong OPENKMS_SECRET_KEY (32+ random bytes).
  2. HTTPS for frontend and backend; OIDC redirect URIs locked down; OPENKMS_ALLOW_SIGNUP=false in local mode if registration should be closed.
  3. Review IdP realm roles vs security_roles.name; prefer granular catalog keys over permanent all.
  4. Restrict database and object storage to trusted networks; keep the VLM URL internal.
  5. Keep dependencies updated (pip install -U, npm audit).

Static home (landing page)

  • Public landing page for non-authorized users (HomeStaticLanding, i18n landing)
  • Pain points: knowledge scattered; evidence can’t travel with the story; manual parsing work
  • Benefits: centralized knowledge hub; grounded answers and domain data; research that becomes a multi-slide analysis briefing
  • Functionalities sections: document management, articles, knowledge bases (including semantic search when pgvector is configured), ontology & graph (datasets, object/link types, optional Neo4j), agents & analysis (installable skill, cross-surface research, interactive analysis presentations), pipelines & automation (jobs, per-channel pipelines, model linkage)