To Be Played – Auth Service

Identity service of To Be Played, a Steam backlog tracker. It handles registration, login, user settings and the public profile, and it signs the JWTs that the Core service trusts.

To Be Played is split into three repositories:

Repository	Role
ToBePlayed-Auth (this one)	Users, login, settings, signs JWTs
ToBePlayed-Core	Steam library, backlog, wishlist, stats; verifies JWTs
ToBePlayed-Frontend	React single-page app

Features

• Registration with an invitation code, login with email and password (BCrypt). Emails are stored trimmed and in lower case (V19, unique on lower(email)); a password has 8 characters and at most 72 bytes (BCrypt's limit).
• Brute-force protection in memory: 5 failed logins per email or 20 per IP, and 10 failed registrations per IP, block for 15 minutes with 429 TOO_MANY_ATTEMPTS and Retry-After. Unknown emails cost the same BCrypt time as a wrong password, and every failed login gets the same error.
• JWT signed with EdDSA (Ed25519). Auth holds the private key, Core only the public key, so Core can verify tokens but never create them. Tokens carry a kid header, and GET /.well-known/jwks.json publishes the public key as a JWK, so Core finds the right key after a rotation.
• Session in two HttpOnly cookies, so no script on the page can read a token. The access token lives 15 minutes in tbp_access (sent to Auth and Core). The refresh token is random, lives 30 days from its last use in tbp_refresh (sent to /api/auth only) and is stored only as a SHA-256 hash. Each refresh replaces it; a replaced token that comes back revokes all tokens of that login. Logout revokes them too.
• CSRF guard: POST, PUT and DELETE need an X-TBP-Client or Authorization header, and CORS allows only the frontend origin. A Bearer header still works for API clients and tests.
• User settings: display name (unique, case-insensitive), language, dark mode.
• Steam account linked through "Sign in through Steam" (Steam OpenID 2.0, no API key), never typed in. Steam proves the SteamID64, and Auth asks Steam to confirm the answer (check_authentication) before it stores the ID as verified. A one-time state plus a binding cookie tie the answer to the browser that started the link, so a foreign Steam login link cannot move someone's Steam ID into another account. IDs typed in before this stay, but unverified: a verified link from another account takes them away. Unlinking is possible.
• Appearance settings synced across devices: accent color, corner style, wide layout, background and dashboard banner from the user's own game art. Only whitelisted presets and Steam app ids are stored, never raw CSS or URLs.
• Opt-in public profile at /u/<display name>. Auth answers with a short-lived profile ticket (EdDSA JWT, 10 minutes, typ=public-profile) that Core accepts for the public library data only. The profile has its own look (accent, corners, banner and background from own game art) and a tile layout (profile_tiles: the shown tiles in order, each with a size S, M or L), apart from the app appearance; the public answer sends both as style. The ticket's claims come from the layout: sections names the shown parts (STATS when any STATS_ tile is shown), profileStats the stat groups of the STATS_ tiles. The user's dashboard layout (dashboard_tiles, NULL = never changed) is stored the same way. The ticket is not a login: every token names its kind in aud (tbp-login, tbp-public-profile, tbp-internal, iss=tbp-auth), and both services check it. A login's sub is the user id; the email is in no token. Until item 29 the tokens also carry the claims of before item 26 (userId, typ), so an older Core still reads them. The claims are pinned in src/test/resources/contract/auth-core-contract.json, the same file as in Core (ToBePlayed-Deployment/scripts/check-contract.sh compares both).
• Errors: every error answer of a controller has the body {code, message} (a 401 or 403 of the security filters has none); a refused request is 400 VALIDATION_FAILED with fields, a duplicate email or display name 409 EMAIL_ALREADY_IN_USE / DISPLAY_NAME_ALREADY_IN_USE, anything unexpected 500 INTERNAL_ERROR without details (shared/api/GlobalExceptionController).
• Roles: one per user, a ladder USER < AUDITOR < PLATFORM_ADMIN (users.role, the first admin is set by SQL). The access token carries the active role in the role claim. View as: an admin switches the own login down to USER or AUDITOR and back (POST /api/auth/view-as); the choice is stored per login in refresh_tokens.active_role and never goes above the stored role, so a demoted admin loses the rights at the next refresh. Core checks the same claim.
• Role markers on services: every service method a controller calls carries @RequiresRole(Role.X) or @Public, on the method or on its class (the method wins). A Spring Security method interceptor checks the active role while a request is served (scheduled jobs and @Async work run without a user and are not checked); a denied call answers 401 without a login, else 403. @Public skips only the role check: login, refresh, "view as" and the Steam callback prove the caller themselves. The URL rules in SecurityConfig stay as the outer guard. ArchUnit fails the build for an unmarked method, two markers, a controller calling an @Async method, and a service method calling another service's method that needs more.
• Admin users page (/api/admin/**, active role PLATFORM_ADMIN): search and page all users, block and unblock (users.enabled; blocking ends every login and, within 30 s, the access token at Auth; Core accepts the token until it expires, max 15 min; the login then answers 403 ACCOUNT_BLOCKED, but only after the right password), delete (the same flag and nightly cleanup as "Delete my account"). No action on the own account or on another admin. Every action and every "view as" is a row in admin_audit.
• Dashboard hints: a PLATFORM_ADMIN writes a hint for all users (plain text in English, German optional, start and end time, type DISMISS or ACCEPT). The dashboard shows the running hints one at a time, newest first. "Do not show again" on a DISMISS hint or "OK" on an ACCEPT hint is a row in hint_done, so the hint never comes back for this user on any device. Creating and ending a hint are rows in admin_audit.
• Delete my account with the password (5 wrong tries per 15 minutes, then 429). It only sets users.deleted_at and revokes every refresh token: from then on no login, no refresh, no access token, no public profile, no Steam link, not in the landing stats. Email and display name stay taken; the Steam ID is freed, so an undo needs a new Steam link. A nightly job (04:15, Europe/Berlin) removes users flagged 30 days ago, in Core first and then in Auth. It calls Core's /internal/accounts/* with a service token (EdDSA JWT, 60 s, aud=tbp-internal). app.account-cleanup.dry-run (default true) only logs.

API

Method	Path	Auth
POST	/api/registration/new	public
POST	/api/auth/login	public, sets both cookies
POST	/api/auth/refresh	refresh cookie, replaces both cookies
POST	/api/auth/logout	refresh cookie, clears both cookies
GET	/api/users/me	access token
PUT	/api/users/settings	access token
POST	/api/users/steam-link	access token, returns Steam's login URL and sets the binding cookie
DELETE	/api/users/steam-link	access token, removes the Steam ID and sets a new access cookie
POST	/api/auth/view-as ({role}, null = own role; new cookies; 403 ROLE_NOT_ALLOWED above the own role)	refresh cookie
GET	/api/admin/users?query=&page=&size= ({content, page, size, totalElements, totalPages})	PLATFORM_ADMIN
PUT	/api/admin/users/{id}/blocked ({blocked}; 409 SELF_ACTION, TARGET_IS_ADMIN, ALREADY_DELETED)	PLATFORM_ADMIN
DELETE	/api/admin/users/{id} (409 SELF_ACTION, TARGET_IS_ADMIN)	PLATFORM_ADMIN
GET	/api/hints ({results}: running hints not closed for good, newest first)	access token
PUT	/api/hints/{id}/done (204; 404 HINT_NOT_FOUND)	access token
GET	/api/admin/hints ({results} with status PLANNED/ACTIVE/ENDED by the server clock and doneCount, newest first)	PLATFORM_ADMIN
POST	/api/admin/hints ({textEn, textDe, kind, startsAt, endsAt}; 201 {id}, 400 INVALID_HINT)	PLATFORM_ADMIN
PUT	/api/admin/hints/{id}/end (204; a planned hint never starts)	PLATFORM_ADMIN
DELETE	/api/users/me ({password}; 204 and both cookies cleared, 403 INVALID_PASSWORD, 429)	access token
GET	/api/steam-link/callback	public, Steam's redirect back; binding cookie; redirects to the frontend settings
PUT	/api/users/settings/appearance	access token
PUT	/api/users/settings/profile-style ({accent, corners, bannerAppid, backgroundAppid}; 204)	access token
PUT	/api/users/settings/dashboard-tiles ({tiles: [{id, size}]} in order, at most 20; {tiles: null} resets; 204). Auth checks only the id shape [A-Z_]{1,30}: Core and the Frontend skip ids they do not know, and Core never shares the overview or the activity	access token
PUT	/api/users/settings/favorite-platform ({platform}, free text up to 32, e.g. PS5; null clears it; 204)	access token
PUT	/api/users/settings/profile-tiles ({tiles: [{id, size}]}, e.g. SUMMARY, STATUS, PLAYING, FINISHED, STATS_<GROUP>; 204)	access token
GET	/api/public/profiles/{displayName}	public
GET	/api/stats/public	public
GET	/api/version ({commit, builtAt} of the running image, from the CI build; dev locally)	public

Endpoints use Spring's API versioning with a query parameter: ?version=1.0 (default 1.0).

Tech stack

• Java 25, Spring Boot 4.1 (Web MVC, Security, Data JPA, Validation, Cache with Caffeine), virtual threads
• PostgreSQL 17, Flyway migrations
• JJWT with Ed25519 keys
• Spring Boot Actuator + Micrometer (Prometheus)
• Tests: JUnit 5, Testcontainers (PostgreSQL), ArchUnit
• CI: Forgejo Actions, multi-stage Docker image

Architecture

Package by feature (modules/login, modules/registration, modules/user, ...), and inside each feature three layers:

Layer	Package	Rule
API	api	REST controllers, request/response records, ApiMapper. Calls services, never repositories.
Domain	model, service, exception	Plain Java models and business logic. Uses repositories, never API classes, entities or EntityMapper.
Persistence	persistence, repo	JPA entities, Spring Data repositories, EntityMapper; repo holds JdbcTemplate queries. Entities never reach the API layer.

user owns the tables users and user_settings through the one JPA entity UserEntity. login and registration read and write users with JdbcTemplate (refresh tokens stay JPA). Every query for "a user who may log in and is shown to others" uses shared/account/ActiveUser. A settings change is one UserRepository.update(userId, change): load the user, change the model, write back the display name and the settings. A new setting needs no new repository method. Three Caffeine caches in CacheConfig: users (by userId) and publicStats expire 10 minutes after the write, activeUsers (the filter's "blocked or deleted?" check per request) after 30 seconds. Every write evicts what it changes. A request that read the database just before a block or delete can put the old answer back, so the short TTL is the real limit: a blocked or deleted user keeps a valid access token for up to 30 s. profile and stats only read data. They have no entities and use JdbcTemplate queries in a repo package. version has only an api package: GET /api/version answers the commit and build time of the running image. account (delete and nightly cleanup) sets the flag with JdbcTemplate in a repo package too, and calls Core through CoreAccountService. admin (users page, audit) works the same way. It deletes through the port shared/account/AccountDeleter, and the login module writes "view as" to the audit table through shared/audit/AdminAuditLog, because modules never use each other. hint (dashboard hints and their admin API) uses JdbcTemplate in a repo package too, and audits through shared/audit/AdminAuditLog.

ArchUnit enforces these rules in src/test/java/auth_service/architecture/ArchitectureTest.java, so a layer violation fails the build. It also keeps the modules independent of each other and free of cycles. Test base classes (AbstractUnitTest, AbstractWebMvcTest, AbstractPersistenceTest, AbstractIntegrationTest) live in auth_service.shared.test. Integration and persistence tests share one PostgreSQL container (TestPostgres), and the base classes carry every annotation, so each kind starts one Spring context. A test with its own annotations starts another one.

Login flow

sequenceDiagram
    participant B as Browser (Frontend)
    participant A as Auth
    participant C as Core
    B->>A: POST /api/auth/login
    A-->>B: Set-Cookie tbp_access (15 min), tbp_refresh (30 days, path /api/auth)
    B->>C: API call with tbp_access
    C->>A: GET /.well-known/jwks.json (new kid only)
    A-->>C: Public keys (Ed25519)
    C-->>B: 200, or 401 when the access token expired
    B->>A: POST /api/auth/refresh with tbp_refresh
    A-->>B: New tbp_access and tbp_refresh, old refresh token rotated
    B->>C: Same API call again
    Note over B,A: View as (admin only)
    B->>A: POST /api/auth/view-as {role} with tbp_refresh
    A-->>B: New tbp_access (role claim) and tbp_refresh, active role stored for this login

1. LoginService checks the password. LoginController sets both HttpOnly cookies (AuthCookies), so no script can read them.
2. Core verifies the access token with Auth's public keys only. JwksController publishes them; Core loads them again only for an unknown kid.
3. On a 401 the frontend calls POST /api/auth/refresh once and sends the request again. RefreshTokenService replaces the refresh token. A replaced token that comes back revokes the whole login.

Steam link flow

sequenceDiagram
    participant B as Browser (Frontend)
    participant A as Auth
    participant S as Steam
    B->>A: POST /api/users/steam-link (logged in)
    A-->>B: Steam login URL, Set-Cookie tbp_steam_link (state, 10 min)
    B->>S: Sign in on steamcommunity.com
    S-->>B: Redirect to /api/steam-link/callback?state=...&openid.*
    B->>A: GET callback with tbp_steam_link
    A->>S: check_authentication (openid.* sent back)
    S-->>A: is_valid:true
    A-->>B: Redirect to /settings?steam=linked
    B->>A: POST /api/auth/refresh (new steamId claim for Core)

SteamLinkService checks the state against the binding cookie before it asks Steam anything, and uses each state once. SteamOpenIdService checks mode, endpoint, return_to and the signed fields, then asks Steam. UserService.linkVerifiedSteamId stores the ID and takes it from an account that only typed it in.

Delete account flow

sequenceDiagram
    participant B as Browser (Frontend)
    participant A as Auth
    participant C as Core
    B->>A: DELETE /api/users/me {password}
    A->>A: users.deleted_at = now, revoke refresh tokens
    A->>C: POST /internal/accounts/deleted (service token)
    A-->>B: 204, both cookies cleared
    Note over A: an old access token gets 401 within 30 s (activeUsers cache)
    Note over A,C: nightly 04:15
    A->>C: PUT /internal/accounts/deleted (all flagged users)
    A->>C: PUT /internal/accounts/states (every user)
    A->>C: POST /internal/accounts/purge (flagged 30+ days)
    A->>A: DELETE users (only when Core answered and no dry run)
    A->>C: POST /internal/accounts/prune (all user ids)

A failed call to Core does not fail the delete: the next night sends the whole flagged set again. Undo within 30 days is UPDATE users SET deleted_at = NULL (see the Deployment README).

Steam ID and block sync to Core

Core keeps its own copy of the Steam ID and of the block. After a verified Steam link (also for the account that loses a typed-in ID), an unlink, a block and an unblock, Auth sends the user's state after the commit (AccountStatePublisher); the nightly cleanup sends every user. The call runs on the task executor, so neither the answer nor a database connection waits for Core; a failed call only warns.

sequenceDiagram
    participant A as Auth
    participant C as Core
    A->>A: link / unlink / block / unblock (commit)
    A->>C: POST /internal/accounts/states {userId, steamId, blocked}
    C->>C: blocked_user, Steam ID in user_steam_sync
    Note over A,C: nightly 04:15
    A->>C: PUT /internal/accounts/states (every user)

Run locally

Local development without keys (profile local)

The whole stack runs on your machine with no API keys and no calls to Steam, Twitch/IGDB or speedrun.com. Needs Java 25, Node 24 and Docker. Clone ToBePlayed-Auth, ToBePlayed-Core and ToBePlayed-Frontend next to each other, then start them in this order (Core also starts without Auth and finds it later):

1. Auth: in ToBePlayed-Auth, ./mvnw spring-boot:run -Dspring-boot.run.profiles=local (IDE: active profile local). Port 8080.
2. Core: in ToBePlayed-Core, ./mvnw spring-boot:run -Dspring-boot.run.profiles=local. Port 8090.
3. Frontend: in ToBePlayed-Frontend, npm install, then npm run dev, and open http://localhost:5173 (localhost, not 127.0.0.1, so the browser sends the login cookies to every local port). The development values must be VITE_API_URL=http://localhost:8080 and VITE_CORE_API_URL=http://localhost:8090.
4. Log in with dev@localhost / password123. On the dashboard, start a Steam sync: the library fills with 8 games, achievements, time to beat, moods, a wishlist and one speedrun game from recorded answers.

In Auth, local makes a new signing key at each start (LocalKeysEnvironmentPostProcessor, only in this profile, and never together with prod), so no key is in the repo; after an Auth restart (also a devtools restart), log in again. Never use local on a server: the dev user's password is public. The Flyway callback in db/local adds the dev user (PLATFORM_ADMIN, linked Steam ID). The invite code for more users is local. What Core mocks and how to change the recorded answers: Core readme, "Run locally".

With your own keys (profile dev)

Needs Java 25, Docker and your own keys.

./mvnw spring-boot:run -Dspring-boot.run.profiles=dev

Spring Boot's Docker Compose support starts PostgreSQL from compose.yaml, Flyway applies the migrations, and the service listens on http://localhost:8080. Health: /actuator/health, metrics: /actuator/prometheus.

The cookies are host-only unless APP_AUTH_COOKIE_DOMAIN is set; production sets it to the parent domain, so Core gets the access cookie too. Browsers accept Secure cookies on http://localhost, so local development needs no cookie settings. Open the frontend on localhost (not 127.0.0.1), so the browser sends the cookies to every local port.

Configuration comes from environment variables:

Variable	Purpose
JWT_PRIVATE_KEY, JWT_PUBLIC_KEY	Ed25519 keypair, Base64 of the DER bytes (PKCS#8 / X.509)
JWT_KID	Key id written into the token header
AUTH_INVITATION_CODE	Code required for registration
DB_PASSWORD	Local PostgreSQL password
APP_PUBLIC_URL	Auth's public address, the Steam OpenID realm (default http://localhost:8080)
APP_FRONTEND_URL	Where the Steam link ends (default http://localhost:5173)
APP_CORE_INTERNAL_URL	Core for the nightly account cleanup, required (local http://localhost:8090, Docker http://tbp-core-service:8090)

Generate a keypair:

openssl genpkey -algorithm ed25519 -out private.pem
openssl pkey -in private.pem -outform DER | base64 -w0   # JWT_PRIVATE_KEY
openssl pkey -in private.pem -pubout -outform DER | base64 -w0   # JWT_PUBLIC_KEY

Tests

./mvnw clean verify

Runs unit, WebMvc, persistence and integration tests plus the ArchUnit rules. Needs Docker for Testcontainers.

Roadmap

• Account deletion and data export.

License

All rights reserved. The source is public for reading only.