paskia/docs/MultiSite.md

Multi-Site Support

One paskia process on one port serves multiple sites with one combined database. The administrative instance is separated from the WebAuthn RP: organizations and users are global across rp-ids; rp-id is a first-class per-realm object; passkeys remain tied to their rp-id (WebAuthn-enforced); sessions remain host-bound. Motivating case: app1.company.com and app2.com cannot share an rp-id, but user management must be under single common controls.

Terminology: a realm is one rp-id with its associated hosts and origins. A site is any host served by the instance; each host belongs to exactly one realm. The administrative instance is the whole process: global users/orgs, N realms.

1. What is global vs. per-realm

Global (single instance, shared across realms):

Data	Notes
Organizations, Roles, Users	one global collection
Permissions	domain field host-scopes effectiveness
Sessions	host-bound (Session.host, exact match)
Credentials/passkeys	global collection, each stamped with its rp_id
Reset tokens	user-bound; global
Avatars	paskia.data/users/<uuid>/profile.webp
Auth codes, remote-auth manager	in-memory; carry rp-id fields

Per-realm (registry, keyed by rp-id):

Data	Notes
rp_id, rp_name, origins, auth_host	stored combined Config
Passkey instance	per rp-id; ceremonies verify against the origin realm's rp-id
site_url/site_path	runtime derivation, per realm
OIDC provider (keys, clients)	per rp-id — each realm is an independent issuer

db.data() is a plain global singleton. A contextvar is needed only for the current realm (passkey, config, OIDC view) — not for database access.

The architectural rule: authentication establishes identity, not organization — the requested hostname selects the org/permission context after authentication (via Permission.domain host-scoping and session host binding).

2. Login architecture: three composable mechanisms

The realm infrastructure is shared by three mechanisms, alternatives per deployment and composable within one instance.

2.A WebAuthn Related Origin Requests (trusted domain families)

WebAuthn Level 3 lets otherwise-unrelated domains share one rp-id: the canonical RP publishes /.well-known/webauthn listing permitted origins, and those origins may then run ceremonies with the common rp-id locally — no redirects, no cross-domain cookies. Browser support is universal (Firefox included).

Model: realm company.com with related origin https://app2.com. A page on app2.com calls WebAuthn with rpId: "company.com"; the passkey is scoped to company.com; clientDataJSON.origin is https://app2.com, which the backend validates against the realm's related origins.

Server side: paskia's Passkey passes expected_origin=<the pre-validated origin> and expected_rp_id=self.rp_id; the webauthn library string-compares origin and rp-id separately. The frontend never chooses rpId client-side — ceremony options arrive from the server over the WS. On top of that:

• Origin rule: origins and related_origins are separate fields. An in-domain origin (rp-id or subdomain) is valid unless the realm's origins allow-list is set, in which case it must be listed there. An origin on another domain is valid only when listed in the realm's related_origins — explicit related listing is the trust boundary.
• GET /.well-known/webauthn on the canonical rp-id host serves {"origins": [...]} from the realm's related origins (404 when there are none).
• Dispatch resolution treats a Host matching a configured related-origin hostname as belonging to that origin's realm (exact match only — www.app2.com does not follow app2.com).

Deployment constraint: the browser fetches https://<rp-id>/.well-known/webauthn from the canonical apex directly — if paskia does not host the apex, publish the JSON there statically (the admin realm dialog links to the document for copying).

Constraints (from the WebAuthn WG): implementations must support at least 5 registrable origin labels — this is for a small family of same-trust domains, not hundreds of customer domains. Sharing an rp-id merges the security boundary: a weakly protected marketing domain should not share the realm of the admin application. Config validation enforces a cap (default 5) on related origins per realm.

Re-enrollment note: passkeys never move between rp-ids (WebAuthn-enforced). A host family that first deploys separate realms (2.B) and later consolidates to Related Origins re-enrolls: authenticate against the old realm (or via 2.C), register a new credential under the common rp-id, retire the old one. The per-credential rp-id badge (§9) makes this visible. There is no automated credential migration.

2.B Multiple rp-id realms under one administrative instance

For domains that should not share an rp-id: rp-id is a first-class object (realm), not an instance attribute. Users are global identities; credentials carry rp_id:

Instance
├── Orgs / Roles / Users (global)
└── Realms
    ├── company.com   (origins [...], credentials scoped by rp_id)
    ├── app2.com      (origins [...], credentials scoped by rp_id)
    └── customer.net  (origins [...], credentials scoped by rp_id)

Alice can hold both a company.com and an app2.com passkey; sessions stay host-only.

2.C Remote authorization + opportunistic local enrollment

For a realm where the user has no credential, the remote-login mechanism provides a federation-style flow: unauthenticated device requests, authenticated device permits, a short-lived single-use opaque exchange code (60s CookieCode) is redeemed by the requesting host, which sets its own host-only cookie. No shared cookies, no reusable tokens in URLs.

• Cross-realm permits are allowed: a device authenticated at company.com may authorize a session for app2.com; the request's realm is recorded and shown to the approver; the target host is registry-validated.
• Opportunistic local enrollment: after a cross-realm remote login, the profile view offers "Add a passkey for " — registration runs locally under the new realm's rp-id, stamping Credential.rp_id. This makes remote login primarily bootstrap/recovery, while everyday authentication stays local.

Policy summary (how deployments choose)

Situation	Mechanism
Few closely related, equally trusted brand domains	2.A Related Origins — one passkey
Independent / customer / lower-trust domains	2.B separate realms — passkey per realm
User lacks a credential for the current realm	2.C remote authorization, then enroll locally

3. Configuration model

3.1 Stored config

class RealmConfig(msgspec.Struct, omit_defaults=True):
    rp_id: str
    rp_name: str | None = None
    auth_host: str | None = None      # this realm's dedicated auth host
    origins: list[str] | None = None  # allow-list of in-domain sign-in sites
    related_origins: list[str] | None = None  # cross-domain ROR origins (§2.A)

class Config(msgspec.Struct, omit_defaults=True):
    realms: list[RealmConfig]         # at least one; first entry is the default realm
    listen: list[str] | None = None   # process-global

• "At least one realm" is enforced by validation (not expressible in msgspec).
• The first entry is the default realm, used only where a default is genuinely needed (bootstrap reset-link URL, startup box ordering, master-admin entry point) — never for dispatch.
• Origin validation — two separate concerns: origins entries must be within the rp-id domain (an allow-list; unset = the rp-id and all subdomains may authenticate). related_origins entries must be outside it, are capped (default 5), and must not collide with another realm's rp-id/auth-host/related origins nor fall inside another realm's domain. Misfiled entries (cross-domain in origins, in-domain in related_origins) are rejected. These rules are enforced both at startup and at admin write time. Origins are never implicitly cross-domain.

3.2 CLI: bootstrap (paskia init) vs. serve (paskia)

The CLI is split so that realm options exist only at bootstrap time — they can never mix with runtime configuration of an already-configured instance:

• paskia init — creates paskia.kantadb in CWD and seeds it:
  • --rp-id: repeatable/comma-separated, default ["localhost"]. Multiple values create multiple realms at once (useful for devserver/e2e); the first is the default realm.
  • --rp-name: applies to the default realm. Its purpose is that the very first admin registration ceremony already shows the correct RP name; afterwards rp-names are edited via the admin interface.
  • --auth-host, --origin: apply to the default realm. Further realms' hosts are configured via the admin interface.
  • --listen: stored into Config.listen (process-global).
  • Seeds the admin user + registration reset link (link URL from the default realm) and prints the link. Refuses to run if paskia.kantadb already exists, or if an unconverted legacy *.paskiadb is present (paskia migrate converts it first).
• paskia migrate — converts a legacy <rp-id>.paskiadb database (§10) to paskia.kantadb. With several legacy candidates, --rp-id selects <rp-id>.paskiadb by name; the others are left in place.
• paskia — serve. Takes no realm options; only --listen (per-run override of stored Config.listen, never persisted). Startup: open paskia.kantadb → validate the stored realm set cross-realm (rp-ids distinct; auth hosts distinct from each other and from every rp-id; related origins capped and collision-free) → build the realm registry → serve. The serve command never converts databases: with no paskia.kantadb, the startup error points at paskia init, or at paskia migrate when legacy *.paskiadb candidates are present.

Nested rp-ids are allowed (longest-suffix dispatch determinism). Adding a child rp-id moves no data — users are global; only new ceremonies stamp the child rp-id.

3.3 Runtime accessors

• The realm registry is built in the FastAPI lifespan after kanta.open(), from db.data().config.realms — realm data does not travel through PASKIA_CONFIG. Per-realm site_url/site_path are computed at registry-build time (priority: auth_host > origins[0] > PASKIA_VITE_URL for the localhost realm > http://localhost:port > https://rp-id), using the effective listen endpoints for the localhost fallback.
• PASKIA_CONFIG carries only process-global serve parameters (the effective listen endpoints) so the derivation inside the server process can resolve the localhost-port fallback.
• Admin realm writes persist the combined Config and rebuild the registry in place, so dispatch sees auth_host and related-origin changes immediately.

4. Credentials carry an rp-id

• Credential.rp_id: str is stamped at registration from the ceremony's rp-id. With Related Origins the stamp is always the realm's canonical rp-id regardless of which origin the ceremony ran on — the credential genuinely is a company.com passkey.
• authenticate_chat filters the raw_id scan by c.rp_id == ceremony rp-id — prevents wrong error semantics and a cross-realm oracle ("no credential" vs "verification failed" would leak which rp-id a credential belongs to).
• exclude_credentials (registration) and reauth allow_credentials are filtered by the ceremony's rp-id (User.credential_ids_for(rp_id)) — users are global, so their credential id sets are cross-realm.
• Cascades are uuid-keyed; deleting a user removes their passkeys across all realms (correct: users are global).

5. Dispatch and realm context

• paskia/realms.py: Realm { config, passkey, ... } and a registry keyed by rp-id, built in the lifespan from the stored combined Config and rebuilt on admin realm writes. No per-realm Kanta/DB.
• Host resolution (resolve(host)): normalize (lowercase, strip port and trailing dot), then exact rp-id → exact auth host → exact related-origin hostname → longest-suffix rp-id. Unknown → None. (Order safe because startup and admin-write validation forbids collisions between these sets.)
• A pure ASGI dispatch middleware, outermost, handles "http" and "websocket" scopes. Unknown Host → 421 Misdirected Request (WS: pre-accept close). Sets the current_realm contextvar + scope["state"]["realm"].
• WebSocket resolution follows the Origin, not the connection Host: in auth-host mode the login page is on the app host, the WS connects to the auth host, and Origin names the host being logged into. So:
  1. the middleware resolves the origin realm from the Origin hostname — including related-origin hostnames;
  2. the connection Host must be a valid WS endpoint for that realm: the realm's effective auth host (§6), or the origin host itself when the realm has no auth host at all — else pre-accept reject;
  3. validate_origin runs endpoint-side against the origin realm's Passkey (post-accept JSON errors preserved);
  4. current_realm = origin realm for the WS handler's duration. The ceremony rp-id is always the origin realm's rp-id — exactly what the browser enforces for the page's origin under both classic and related-origin rules.

6. Auth host: per-realm values with global fallback

A realm without its own auth host falls back to the first configured auth host (realm-list order):

effective_auth_host(realm) = realm.auth_host or first_configured_auth_host or None

Own vs. effective auth host are distinguished everywhere:

• Follow the realm's OWN auth host: UI mode detection (minimal- profile decision in App.vue, via own_auth_host in settings), the redirect middleware, ui_base_path, and reset_link_url. A realm with no own auth host keeps its full UI on its own hosts — otherwise reset/registration pages on app2.com would redirect to auth.company.com, where the ceremony's Origin resolves the owner realm and stamps the wrong Credential.rp_id, breaking 2.B onboarding and 2.C local enrollment.
• Follow the EFFECTIVE auth host: WS endpoint selection only (passkey.js builds the WS URL from settings). The fallback auth host serves WS + restricted APIs for foreign realms.
• Settings (ApiSettings) exposes both fields (own_auth_host alongside the effective auth_host) so the frontend makes the mode decision correctly.
• Root mode (site_path == "/") applies only on a realm's own auth host, so it is never ambiguous: a realm's UI lives on its own hosts; the fallback auth host serves the owner realm's UI plus WS for the rest.

7. Login flows and in-memory stores

7.1 Auth codes

• OIDCCode and CookieCode carry rp_id, verified at redemption — defense in depth.
• Stamping source matters: codes are stamped with the realm of the session they will redeem — not naively with the current realm at issuance. Remote-completion codes are minted inside the permit handler (permitting realm's context) but redeemed by the requesting device on its own host, so they are stamped with RemoteAuthRequest.rp_id — stamping them with the permitter's realm would break every cross-realm remote login. Registration-flow and OIDC codes stamp from the current realm (issue and redeem sides always match). The host re-check at set-session independently binds CookieCode to the host; the rp_id check complements it.

7.2 Remote authentication — cross-realm permits allowed

• RemoteAuthRequest.rp_id records the requesting device's origin realm.
• Permit side may differ from the request side (mechanism 2.C): the permitting device authenticates with its realm's passkey, and the session_host=request.host override creates the session for the requesting host. The override must resolve to a configured realm (registry check) — arbitrary-host session binding is refused. The request's rp-id is shown to the permitting user ("device at app2.com requests login").
• No policy flag: cross-realm remote login is how the product works (users are global).
• Exchange codes stay single-use, 60s, host-bound at redemption.

8. OIDC: per-realm providers in one DB

• DB.oidc is dict[str, OIDC] keyed by rp-id. Each realm is an independent provider: own signing key (oidc.<rp-id>.key in the transaction log shape), own clients.
• The util/oidjwt.py key cache is keyed by rp-id.
• Issuer stays per-request-Host — each realm host is an issuer alias sharing the realm's key. Session carries two fields: issuer (stamped from the WS Origin, scheme included, at OIDC-session creation and re-stamped at refresh — stamping from the WS connection Host would be wrong, that is the effective auth host, not the authorize/discovery host the RP validates against); and rp_id — the owning realm, needed by every path that runs without request context:
  • backchannel logout uses session.rp_id to select the realm's key and session.issuer as iss;
  • session listings resolve the client under the session's own realm, not the request's.
• Admin OIDC-client CRUD operates on the current realm's OIDC entry.
• Permission domain validation accepts a subdomain of any configured rp-id, any related-origin hostname, or any realm's client UUID.

9. Admin API and UI

• GET /auth/api/settings: per-request-Host realm values (rp_id, rp_name, effective auth_host, own_auth_host, site URLs).
• Realm management (master admin only, auth:admin) — how rp-ids are managed after bootstrap:
  • GET/POST /auth/api/admin/realms/ and PATCH/DELETE /auth/api/admin/realms/{rp_id}. Writes require recent authentication (5 minutes).
  • Create: rp_id + optional rp_name (defaults to the rp-id), auth_host, origins, related_origins; full §3.1 validation (cap, cross-realm collisions); registry rebuilt immediately, including the realm's Passkey instance and OIDC provider entry.
  • Update: same validation against the would-be combined config. Changing a realm's rp-id itself is not supported (it would orphan every credential stamped with the old rp-id) — delete and recreate instead.
  • Delete: refused for the last remaining realm and while any credential carries the realm's rp-id (re-enroll or delete those credentials first); cascades nothing else (users/orgs are global).
  • The admin UI has a Realms section with a table (rp-id, name, effective auth host, sign-in site and related domain counts), per-row edit/delete and an add-realm dialog. The dialog edits the in-domain allow-list and the related domains as two separate lists with their own explanations. Its connectivity probe fetches <origin>/auth/api/settings and compares the returned rp-id against the edited realm — a related origin served by this instance answers with the realm's rp_id. Connectivity/mismatch results are warnings; malformed entries, misfiled entries (cross-domain in the allow-list, in-domain in related domains), and an auth host outside the rp-id domain block saving.
• Credential listings: Credential.rp_id serializes automatically into user-info and admin user detail responses; the frontend shows an rp-id badge only when credential.rp_id !== settings.rp_id — single- realm installs never show a badge.
• Enrollment prompt (2.C): when the user has no passkey for the current realm, the profile view offers to add one (a fresh remote-auth session satisfies the recent-auth requirement of registration).
• Bootstrap/reset links use the default realm's URL.
• Bootstrap check: the "admin has no credentials" startup check tests for an admin credential under the default realm's rp-id — with global users an admin may have passkeys only under another realm, and the printed link must still be usable for the default realm.

10. Storage

• Fixed CWD-relative path: paskia.kantadb — a single kanta JSONL file. Kanta rotation siblings (paskia@<timestamp>.kantadb) are unaffected. There is no environment override; CWD selects the deployment.
• User files (avatars) live in the fixed sibling directory paskia.data/users/.
• Legacy conversion: paskia migrate converts a legacy *.paskiadb database — a directory containing main.db, or a legacy single-file database — into paskia.kantadb: main.db (or the single file) becomes paskia.kantadb, users/ becomes paskia.data/users/, and the old directory is renamed aside to <name>.converted-bak. A lone candidate converts without options; with several candidates --rp-id <rp-id> selects <rp-id>.paskiadb by name and the rest are left in place (e.g. a *.bak.paskiadb backup does not block conversion). Empty directories are ignored. Conversion is an explicit operator action, never a serve side effect — read-only opens never trigger conversion or writes.
• The legacy database's structs live in a separate module (paskia/db/legacy.py). There is no multi-database merging.
• The startup box prints per-realm lines.

11. Lifespan and background tasks

• One Kanta for paskia.kantadb, opened once in the lifespan; one background cleanup task (DB is global).
• The kanta bootstrap hook only ever fires for a database created by paskia init or paskia migrate; the serve command never bootstraps.
• The registry is built from the stored Config after open; per-realm Passkey instances constructed (each realm's origins validated at startup — fail-fast, including related-origin cap checks).
• The admin-credential check runs at serve startup and reprints a usable registration link when the admin lacks a credential under the default realm (§9).
• oidc_notify fire-and-forget tasks need no realm context for DB access (global DB); issuer comes from the session (§8).
• The dispatch middleware is the only place current_realm is set for requests; admin realm writes rebuild the registry.

12. Development

• scripts/devserver.py: bootstraps via one-shot paskia init when no database exists (multi --rp-id, and --rp-name/--auth-host/ --origin for the default realm), then runs plain paskia serve. Caddy dev origins iterate all bootstrap rp-ids plus the auth host and explicit origins.
• PASKIA_AUTH_HOST (consumed by frontend/vite.config.js) is a comma-separated list of bare hostnames; the vite dev proxy forwards /.well-known/openid-configuration and /.well-known/webauthn to the backend.
• The example caddy/auth/setup snippet forwards both well-known paths to paskia so a static /.well-known/* handler does not shadow them.
• E2E: e2e/tests/global-setup.ts runs paskia init --rp-id localhost,test.localhost in the test-data dir (which doubles as the server CWD) and serves; e2e/tests/50-multirealm.spec.ts exercises host dispatch, the well-known endpoint via the admin realm API, and a cross-realm remote login (request at test.localhost, permit at localhost, session valid on test.localhost) including the enrollment prompt UI. Related Origins have no browser e2e: a genuine related origin needs a non-subdomain host over HTTPS, and the browser fetches the well-known document itself — server-side coverage is in pytest (tests/test_realms.py).

13. Security model

• Dispatch: unknown Host → 421 before any router/DB access (direct-IP and unconfigured-name access does not work; trailing dots normalized).
• Related Origins boundary: cross-domain origins are valid only when explicitly configured and capped; the well-known document is served only for the canonical realm and only lists configured origins. All origins sharing an rp-id share one security boundary — do not mix trust levels within a realm.
• Realm administration: realm create/update/delete is gated on auth:admin — deployment-wide by design; validation runs on every write, not just at startup.
• Passkeys: rp-id binding browser-enforced and server-recorded; ceremonies, credential scans, exclude/allow lists all scoped to the origin realm's rp-id. No cross-realm oracle in the scan.
• Sessions: host-bound, exact match. Cross-realm sessions arise only via (a) a ceremony at the origin realm (incl. related origins), or (b) a remote permit by a device holding a valid session at its own realm, with registry-validated target host.
• Users/orgs global: deleting a user/org cascades across all realms — intended. auth:admin is deployment-wide. Permission domain host-scopes effectiveness per host.
• Cross-realm permit transparency: requesting realm/host shown to the approver; both sides logged.
• Secret hygiene in logs: the OIDC signing-key censoring matches the oidc.<rp-id>.key path shape, so realm keys never print in plaintext in the JSONL transaction log.
• OIDC: per-realm keys/issuers; logout tokens carry the stored issuer.