ToBePlayed-Core/readme.md

# To Be Played – Core Service

Business service of **To Be Played**, a Steam backlog tracker. It syncs the user's Steam library, stores the backlog
(status, rating, notes, priority), tracks playtime per day, syncs the wishlist and adds time-to-beat data from IGDB.

To Be Played is split into three repositories:

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

## Features

- **Steam sync**: owned games, playtime and achievements through the Steam Web API. Manual sync plus a nightly
  sync. Resilience4j rate-limits every Steam call and tries the owned games up to three times (never again on a 4xx;
  the wishlist and the calls a page waits for once); Caffeine caches game counts and player summaries. A sync commits its claim (`RUNNING`, with
  `sync_started_at`), its writes and its finish apart, and calls Steam and IGDB outside any transaction: a second
  sync sees `RUNNING` and skips, a failed one stores `FAILED`. A run older than `app.library.running-stale-minutes`
  can be taken over; its own finish then changes nothing.
- **Backlog**: status (`BACKLOG`, `PLAYING`, `FINISHED`, `COMPLETED` = 100 %, `DROPPED`), rating 1–5, notes, priority,
  filters and sorting. "Pile of Shame" summary with a random suggestion of what to play.
- **Picker**: up to 10 random backlog games for a time budget (IGDB normal time to beat) and a mood, HIGH priority
  first, then MEDIUM, then the rest. Games already shown can be excluded for a reroll. Moods `STORY`, `COOP`,
  `CHILL`, `ACTION`, `SCARY` come from IGDB genres, themes and game modes; `SPEEDRUN` from speedrun.com.
- **Speedrun games**: a backlog game counts as a speedrun game when speedrun.com has a game with the same name (and
  a release year at most one year off, when IGDB knows it) with at least one verified run. The game detail returns
  the speedrun.com link as `speedrunUrl`. speedrun.com has no Steam or IGDB link and no key, so one queue works for
  all users, with up to 3 requests per game and 1 request per second; a result is kept 30 days.
- **Achievements**: after every library sync a background job fetches the achievement schema of each owned game
  and the user's unlocks of each game whose playtime changed. One Steam call per second, so user requests keep half
  of the rate limit; it stops after 3 Steam errors in a row or when Steam hides the achievements. A Steam error never
  overwrites stored data. Summary: unlocked total, perfect games, average completion over started games, the 5
  games closest to 100 %.
- **Playtime history**: one row per game and day, for the activity view and the "most played in the last two weeks"
  dashboard banner.
- **Finished from achievements or playtime**: `EndingAchievement` finds the one achievement of a game that means
  "story end" (English text rule, Steam's global unlock rate between 1 and 80 %, the easiest variant wins, more than
  3 candidates = a game without one end). When the user unlocks it, never chose the status and never answered
  "Not yet", the achievement sync sets FINISHED, but only for a certain ending: its text says so and no other
  achievement of the game could be the end. A status the user chose is never changed; such games get a hint on the game page instead. A second,
  weaker hint compares the playtime with IGDB's normal time to beat (at least 90 %), only for games IGDB does not
  mark as endless (sandbox, survival, 4X, simulator, MMO). "Not yet" hides a hint per game; a dismissed playtime hint
  still lets the ending hint through.
- **Hide games**: a user can hide a game, Steam or added by hand (e.g. a playtest or a tool). It leaves every list and number: library,
  picker, Pile of Shame, activity, achievement summary, public profile and landing page stats. The sync still
  updates its playtime, but no achievements, time to beat, moods or speedrun data are fetched for it. The library
  lists hidden games with `hidden=true`.
- **Stats**: `UserStatsService` builds the numbers of the stats page per group, so every page asks only for the
  groups it shows: totals and statuses; the pile (time to beat, weeks "at your pace" of the last 90 days, years of
  40-hour weeks, oldest unplayed game); most played and most loyal game (most months with playtime); top IGDB genres
  and moods; moon trips (Apollo 11, 195.3 h), longest day, share under one hour; finished games per year (from the
  status history); games and playtime per platform (`PLATFORMS`: Steam games as "Steam", manual games by platform, trimmed and ignoring case, the most used spelling is the label); the Pile of Shame trend (90 days of `backlog_snapshot`, plus "this month: added / done", where the
  first sync of a Steam account is no "added", see `user_steam_sync.history_since`). Every number counts only listed games (still owned, not hidden: the view
  `listed_user_game`, V24). A linked set (see "Editions of one game") counts once: its most advanced edition (status, then
  playtime, then the older `first_seen_at`; `counted_user_game`, V28), so the set is on the pile only when every owned edition is `BACKLOG`. Playtime
  sums, most played, genres, moods and platforms still use every listed game.
- **Drill-down**: every countable number opens the games behind it. A `StatKey` (library), `CommunityKey` (landing page)
  or a Pile Points track/bonus row names a number; its SQL "game set" is the same text the number counts, so the
  list and the number cannot drift. 40 games per page (at most page 250), a fixed order per key. A profile visitor
  gets a list only when the ticket shares that section, or the `STATS` section and the key's group (`403 NOT_SHARED`;
  `OVERVIEW` keys are for the owner only). The community list has one row per game with `libraries` and `minutes`,
  never a user, and is cached 10 minutes like the numbers. A bad argument is `400 BAD_ARG`.
- **Status history**: a database trigger logs every status change of a library game in `user_game_status_change`
  (old and new status, time, whether the user chose it), whatever code path made it. It starts with V14; older
  changes are not known.
- **Pile Points** (`modules/rewards`): users earn points for playing games off the pile, in an append-only ledger
  (`reward_ledger`, V26; the balance is the sum). No code path pays: a settle (on `GET /api/rewards/wallet` and
  nightly at 02:30, `app.rewards.settle-cron`) derives rows from the status log, playtime days, achievements and pile
  snapshots, once per user, kind and game (`ref`). Rules and numbers: `RewardRules`; a Steam finish needs playtime and one
  proof of fresh play (ending achievement or playtime around the finish, or half of the IGDB time played since the finish; old history pays nothing). The month bonus needs a finished game in that month. Games added by hand pay 15 for
  a finish, at most 3 per ISO week. The settle takes a per-user advisory lock. V26 also pays a one-time starter grant
  for finishes from before the status log. Shop: a whitelist of 57 looks (`ShopItem`; 45 buyable and 12 unlock-only) in eight slots, including animated completed effects, page backgrounds (CSS, SVG and two canvas engines in the Frontend),
  banner backgrounds (V27 widens `reward_equipped.slot`) and three buyable skins (full themes, slot `SKIN`). The unlock-only
  items (price 0, never sold): the settle books an `UNLOCK` row (ref = item code, 0 points) once a
  track reaches the item's target, retroactively and never lost. The five tracks (`Unlocks`): games finished
  (`FINISHED` + `FINISHED_MANUAL` rows + starter grant / 20), games completed, weeks in a row with a finish, months in
  a row with a `MONTH` bonus, months in a row with more Steam playtime than the month before (hidden games count too).
  A `BUY` or `UNLOCK` row proves ownership,
  `reward_equipped` holds what is worn. A platform admin owns every item (the role of the login token, or the `role`
  claim of the profile ticket for visitors).
- **Editions and links**: editions, remasters and ports of one game stay separate library rows, but an AUDITOR can
  link their keys as one game for every user (see "Editions of one game"). IGDB suggests pairs, users can ask for
  one, and one queue holds both. Every link, unlink and dismissal goes to `admin_audit`
  (`GAME_LINK`, `GAME_UNLINK`, `GAME_LINK_DISMISS`); auditors read that log through `GET /api/library/audit-log`, with
  the actor's display name from Auth's account state push.
- **Wishlist**: synced from Steam with store price and discount.
- **Broken image repair**: the frontend reports a Steam cover or achievement icon that does not load. At 01:30
  (Europe/Berlin) a job checks up to 200 reports. A cover that loads again is left alone; else the store's hashed
  header picture (`IStoreBrowseService/GetItems` with assets) goes to `steam_game.image_url`. When a few achievement
  icons do not load, the schema comes fresh from Steam; an empty answer keeps the stored one. A report that cannot be repaired is tried again on the next report, at most 3 times,
  then after 30 days.
- **Time to beat**: story, extras and completionist hours from the IGDB API. Optional: without credentials the
  feature is off and the rest works.
- **Public profile data**: totals, status counts, games in progress and finished games with rating, perfect games
  and average completion (the numbers count a linked set once; the lists keep every edition). Readable only
  with a profile ticket issued by Auth. Never returns notes, priority or the Steam ID. `stats` holds only the stat
  groups the user shares (the ticket's `profileStats` claim, derived in Auth from the profile tiles); none by
  default. A section the user
  hides (the ticket's `sections` claim: `SUMMARY`, `STATUS`, `PLAYING`, `FINISHED`, `STATS`) is null and runs no
  query; a ticket without the claim shows all.

## Authentication

Core never issues tokens. It verifies Auth's EdDSA (Ed25519) JWTs with public keys only. It loads them from Auth's
JWKS by the token's `kid` and loads again when a new `kid` shows up (at most every 10 seconds). Each load replaces
the key set: a key that left the JWKS stops working. `JWT_PUBLIC_KEY` is used only until the first JWKS load works.
The token comes from Auth's `tbp_access` cookie or a `Bearer` header, whichever verifies first. Every token names its kind in `aud`
(`tbp-login`, `tbp-public-profile`, `tbp-internal`, with `iss=tbp-auth`); a login's `sub` is the user id. Core reads
`steamId` and the active `role` (missing or unknown = USER) from the claims and grants that role and every lower one
(`USER < AUDITOR < PLATFORM_ADMIN`). A token without `aud` is refused. The claims are pinned in
`src/test/resources/contract/auth-core-contract.json`, the same file as in Auth; the tests build tokens only from it.
`POST`, `PUT` and `DELETE` need an `X-TBP-Client` or `Authorization` header (CSRF guard), and CORS allows only the
frontend origin.
A token for another audience (for example a profile ticket) is rejected on normal endpoints; the profile ticket is
accepted only by `GET /api/public/profiles/library` in the `X-Profile-Ticket` header.
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: [{field, message}]`, a duplicate key `409 CONFLICT`, anything unexpected `500 INTERNAL_ERROR` without details
(`shared/api/GlobalExceptionController`).
`/internal/**` has its own filter chain: it accepts only a service token of Auth (`aud=tbp-internal`, with `exp`, sent
as `Bearer`), never a login, also not in the dev profile. Traefik exposes the path, so the signed token is its
only guard.

Every service method a controller calls carries one role marker, on the method or on its class (the method wins):
`@RequiresRole(Role.X)`, `@Public` (no login, e.g. landing stats; the public profile checks its ticket) or
`@Internal` (only Auth's service token). A Spring Security method interceptor checks it while a request is served;
scheduled jobs and background work run without a user and are not checked. That stays on purpose: the sync
worker calls `SteamApiService` and the import worker `ManualGameService`, both `@RequiresRole(USER)`, without a
request. Auth denies instead, because it has no such caller. A denied call answers `401` without a
login, else `403`. 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 (the check would never run), and a service
method calling another service's method that needs more.

**Mirrored in Auth.** These files exist in both services with small differences; a change in one needs the same
change in the other: `ClientHeaderFilter`, `RoleAuthorizationManager`, `MethodSecurityConfig`, `Role`, `RequiresRole`,
`Public` (only Core has `Internal`), `UserPrincipal`, the login token check in `JwtService.parseLoginToken`, the
ArchUnit role rules (`RoleAnnotationRules`) and `src/test/resources/contract/auth-core-contract.json`
(`ToBePlayed-Deployment/scripts/check-contract.sh` compares the two).

## API

| Method | Path | Auth |
|---|---|---|
| `GET` | `/api/rewards/wallet` (balance, catalog, owned, equipped, last 50 ledger rows with the game name, unlock `progress` per track) | JWT |
| `POST` | `/api/rewards/purchases` (`{item}`; `200` wallet, `409 ALREADY_OWNED` / `NOT_ENOUGH_POINTS` / `NOT_FOR_SALE`) | JWT |
| `GET` | `/api/rewards/progress/{track}/games?page=`, `/api/rewards/history/{kind}/{ref}/games?page=` (the games behind a track or a `MONTH` / `STREAK` bonus; `400 NOT_A_GAME_LIST`) | JWT |
| `PUT` | `/api/rewards/equipped/{slot}` (`{item}`, null = take off; `200` wallet, `409 NOT_OWNED`, `400 WRONG_SLOT`) | JWT |
| `GET` | `/api/library/games` (`hideDuplicates=true` leaves out the lesser editions of a linked set), `/api/library/games/{appid}` | JWT |
| `PATCH` | `/api/library/games/{appid}/backlog` (only the sent fields; `null` clears one) | JWT |
| `PUT`, `DELETE` | `/api/library/games/{appid}/corrections/{field}` (`{value}`; `name`, `image`, `speedrun`, `moods`, `endless`, `genres`, `time_to_beat` = `{normallyMinutes, completelyMinutes}`; `204`, `400 INVALID_CORRECTION`) | JWT, AUDITOR |
| `GET` | `/api/library/catalog?q=&page=&size=` (every game some user still has, by name part, `PageDto` of `{appid, name, imageUrl, platforms}`; `q` at most 100 characters; size 20, 40 or 80) | JWT, AUDITOR |
| `GET` | `/api/library/catalog/{appid}` (what the correction dialog reads: name, image, speedrun, moods, endless, genres, timeToBeat, corrections; `404 GAME_NOT_FOUND`) | JWT, AUDITOR |
| `POST` | `/api/library/games/{appid}/merge` (`{otherAppid}`; `200 {appid}` = the key that stays, `400 SAME_GAME` / `STEAM_GAMES`, `404 GAME_NOT_FOUND`) | JWT, AUDITOR |
| `POST` | `/api/library/merges/{mergeId}/undo` (`mergeId` = the `id` of a `GAME_MERGE` audit row, within 30 days; `204`, `404 MERGE_NOT_FOUND`, `409 UNDO_NOT_AVAILABLE` / `NEWER_MERGE`, `410 UNDO_EXPIRED`) | JWT, AUDITOR |
| `POST`, `DELETE` | `/api/library/games/{appid}/link` (POST `{otherAppid}`, `204`, `400 SAME_GAME`, `404 GAME_NOT_FOUND`, `409 ALREADY_LINKED`; DELETE takes the game out of its set, `204`, `404 NOT_LINKED`) | JWT, AUDITOR |
| `GET` | `/api/library/audit-log?action=&page=&size=` (the newest `admin_audit` rows, `PageDto` of `{id, actorUserId, actorName, action, appid, gameName, targetAppid, targetGameName, detail, createdAt, undoable}`; `action` is `GAME_MERGE`, `GAME_MERGE_UNDO`, `GAME_LINK`, `GAME_UNLINK`, `GAME_LINK_DISMISS`, `GAME_CORRECT` or `GAME_CORRECTION_RESET`; size 20, 40 or 80) | JWT, AUDITOR |
| `GET` | `/api/library/link-queue?page=&size=` (`PageDto` of `{a, b, igdb, requests, createdAt}`: both games, most requested first; size 20, 40 or 80) | JWT, AUDITOR |
| `DELETE` | `/api/library/link-queue/{appid}/{otherAppid}` (dismiss; `204`, `404 PAIR_NOT_OPEN`) | JWT, AUDITOR |
| `POST` | `/api/library/games/{appid}/link-requests` (`{otherAppid}`, both in the caller's library; `204`, `400 SAME_GAME`, `404 GAME_NOT_FOUND`, `409 ALREADY_LINKED` / `PAIR_DISMISSED`, `429 TOO_MANY_REQUESTS`) | JWT |
| `PUT` | `/api/library/games/{appid}/status` (status only, keeps rating, notes, priority) | JWT |
| `PUT` | `/api/library/games/{appid}/finished-hint` (`{dismissed: ENDING\|PLAYTIME}`, "Not yet"; `204`) | JWT |
| `PUT` | `/api/library/games/{appid}/hidden` (`{hidden}`; Steam and manual games) | JWT |
| `POST` | `/api/library/imports` (`{source: TEXT\|CSV\|SEARCH\|GALAXY\|GROUVEE\|HLTB\|PLAYNITE\|BOOKMARKLET, rows: [{title, platform, playtimeMinutes, status, igdbId}]}`, 1–1000 rows; `202 {id}` in `QUEUED`, `409 IMPORT_RUNNING`, `503 IGDB_OFF`) | JWT |
| `GET` | `/api/library/imports` (`{imports}`, the last 20), `/api/library/imports/{id}` (status, progress counts, `ahead` while it waits in a queue, `attempts`; the rows with their candidates while `READY`) | JWT |
| `POST` | `/api/library/imports/{id}/save` (`{rows: [{lineNo, igdbId, platform, status, playtimeMinutes}]}`; `202`, then poll until `SAVED`; `409 IMPORT_NOT_READY`) | JWT |
| `POST` | `/api/library/imports/{id}/retry` (a `FAILED` import; `202`, `409 IMPORT_NOT_FAILED` / `IMPORT_RUNNING`) | JWT |
| `GET`, `DELETE` | `/api/library/imports/{id}/undo-check` (`{games, changed}`), `/api/library/imports/{id}/games` (undo; `204`, `409 IMPORT_NOT_SAVED`) | JWT |
| `POST` | `/api/library/sync` | JWT |
| `POST` | `/api/library/broken-images` (`{appid, kind}`, kind `COVER` or `ACHIEVEMENT_ICONS`; `204`) | JWT |
| `GET` | `/api/library/picker?maxMinutes=&mood=&count=&exclude=` | JWT |
| `GET` | `/api/library/stats?groups=` (`OVERVIEW`, `PILE`, `TOP_GAMES`, `TASTE`, `HABITS`, `FINISHED`, `TREND`, `PLATFORMS`; none = all) | JWT |
| `GET` | `/api/library/stats/games?key=&arg=&page=` (the games behind a stats number; `400 BAD_ARG`) | JWT |
| `GET` | `/api/library/achievements/summary` | JWT |
| `GET` | `/api/library/sync-status`, `/summary`, `/activity`, `/wishlist`, `/steam-profile`, `/games-count` | JWT |
| `GET` | `/api/public/profiles/library` | profile ticket |
| `GET` | `/api/public/profiles/library/stats/games?key=&arg=&page=` (`403 NOT_SHARED`, `400 BAD_ARG`) | profile ticket |
| `GET` | `/api/public/profiles/rewards` (`{equipped}`, the worn looks) | profile ticket |
| `GET` | `/api/stats/public` (+ games per platform, top genres, finished per year) | public |
| `GET` | `/api/stats/public/games?key=&arg=&page=` (the games behind a landing number, per game `libraries` and `minutes`; `400 BAD_ARG`) | public |
| `POST`, `PUT` | `/internal/accounts/deleted` (`{userIds}`; POST adds, PUT replaces the whole set; `204`) | service token |
| `POST`, `PUT` | `/internal/accounts/states` (`{accounts: [{userId, steamId, displayName, blocked}]}`, `displayName` optional, at most 100 characters, kept in `account_name` for the audit log; POST updates, PUT replaces the whole set; `204`) | service token |
| `POST` | `/internal/accounts/purge` (`{userIds}`), `/internal/accounts/prune` (`{knownUserIds}`, not empty) | service token |
| `GET` | `/api/version` (`{commit, builtAt}` of the running image, from the CI build; `dev` locally) | public |

## Tech stack

- Java 25, Spring Boot 4.1 (Web MVC, Security, Data JPA, Validation, Cache with Caffeine), virtual threads
- PostgreSQL 17, Flyway migrations
- Resilience4j: retry and rate limiter for Steam, rate limiters for IGDB and speedrun.com
- JJWT with Ed25519 public key
- Spring Boot Actuator + Micrometer (Prometheus)
- Tests: JUnit 5, Testcontainers (PostgreSQL), ArchUnit
- CI: Forgejo Actions, multi-stage Docker image

## Architecture

A modular monolith. Each module under `core_service.modules` is a vertical slice that could move to its own
service later.

```mermaid
flowchart LR
    Browser["Frontend in the browser"]
    subgraph CoreService["ToBePlayed-Core"]
        Security["configuration.security<br/>JwtAuthenticationFilter"]
        subgraph Library["library"]
            LibApi["api"] --> LibService["service"] --> LibRepo["repo"]
        end
        Stats["stats"]
        Version["version"]
        Account["account"]
        Rewards["rewards"]
        Steam["steam_adapter"]
        Igdb["igdb_adapter"]
        Speedrun["speedrun_adapter"]
        LibService --> Steam
        LibService --> Igdb
        LibService --> Speedrun
    end
    Browser -->|"tbp_access cookie"| Security
    Rewards --> Db
    Security -->|"JWKS by kid"| Auth["ToBePlayed-Auth"]
    Steam --> SteamApi["Steam Web API"]
    Steam -->|"HEAD"| SteamCdn["Steam CDN"]
    Igdb --> IgdbApi["IGDB API"]
    Igdb -->|"OAuth token"| Twitch["Twitch"]
    Speedrun --> SpeedrunApi["speedrun.com API"]
    LibRepo --> Db[("PostgreSQL tbp_core")]
    Stats -->|"listed_user_game, counted_user_game, game_mood, user_game_status_change, deleted_user, blocked_user"| Db
    Auth -->|"service token<br/>/internal/accounts"| Account
    Account -->|"per-user rows, deleted_user, blocked_user, account_name, user_seen_at"| Db
```

| Module | Content |
|---|---|
| `library` | Sync, backlog, manual games, imports, playtime history, wishlist, achievements, public profile data |
| `stats` | Public numbers for the landing page |
| `version` | `GET /api/version`: commit and build time of the running image |
| `account` | Driven by Auth: tombstones in `deleted_user`, purge and orphan prune of the per-user rows; blocks in `blocked_user`, display names in `account_name`, Steam ID changes stop the nightly sync |
| `steam_adapter` | Steam Web API client, availability check of Steam CDN pictures |
| `igdb_adapter` | IGDB client (Twitch OAuth): time to beat, game tags for moods, release years, game search |
| `speedrun_adapter` | speedrun.com client: game search by name, verified runs |

**One shared schema.** All modules use one database `tbp_core`, and some read or write another module's tables in
SQL, which ArchUnit cannot see: `stats` reads `listed_user_game`, `counted_user_game`, `steam_game`, `game_mood`, `user_game_status_change`, `deleted_user` and `blocked_user`; `account`
deletes the per-user rows of the library tables (`AccountCleanupRepository`; a new library table with `user_id` goes
into its list) and writes `user_seen_at` and `last_synced_at` of `user_steam_sync` (`AccountStateRepository`); `library`
reads `deleted_user` and `blocked_user`. A module that moves to its own service takes these reads along as calls.
The listed rule ("still owned, not hidden") is `UserGameRepository.LISTED` for JPQL and the view
`listed_user_game` for native SQL; V24 says what a migration of `user_game` must do with the view. The numbers that
count games once per linked set read `counted_user_game` (V28), built on `listed_user_game`: a migration that changes
a column of `user_game` drops this view first, then `listed_user_game`.

Inside a module: `api` (controllers, DTOs), `service` (business logic), `domain` (JPA entities) and `repo`
(repositories). `src/test/java/core_service/architecture/ArchitectureTest.java` checks these rules on every build:

- Modules do not depend on each other. Only `library` may use `steam_adapter`, `igdb_adapter` and
  `speedrun_adapter`.
- No cycles between `modules`, `shared` and `configuration`.
- `api` never uses `repo`, and `repo` never uses `api`.
- `domain` never uses `api`, `service`, `repo` or an adapter: the import stores IGDB hits as the own record
  `IgdbGame`, so a change of the adapter's `IgdbApiService.Game` cannot break stored imports.
- `service` never uses `api`, except the adapters' response DTOs and a fixed list of 8 services that still build
  their module's response DTOs. The list only gets shorter.
- Every class in `repo` ends with `Repository`.
- `shared` never uses `modules`, and `modules` never use `configuration`.
- A `*Service` class in a module's `service` package has `@Service`. A `*Controller` class in `api` has
  `@RestController`, `@ControllerAdvice` or `@RestControllerAdvice`.

### Sync flow

1. `GET /api/library/games`, `/games-count` and `/picker` call `LibrarySyncService.syncIfStale`. It starts a sync
   when the last one is older than 24 hours, older than 15 minutes after a failure, or was for another Steam ID.
   `POST /api/library/sync` starts one at once, but at most every 60 seconds (else `429` with `Retry-After`).
2. A nightly job (23:55, Europe/Berlin) syncs every account used in the last 14 days, one after the other. A
   request without a `steamId` claim (the user unlinked the Steam account in Auth) takes the user out of it.
   Afterwards it writes every user's pile size of the day to `backlog_snapshot`, also for users without Steam, and
   dates finished games by the unlock of their ending achievement
   (`FinishedEstimateService`, `user_game.finished_at_estimate`).
3. `LibrarySyncWorker` runs on the background pool. `user_steam_sync` lets only one run per user start, and none for
   a user in `deleted_user` or `blocked_user`.
4. The worker loads the wishlist and the owned games from Steam. It updates `steam_game` and `user_game`, adds
   the playtime growth of the day to `user_game_playtime_day`, and soft-deletes Steam games the user no longer
   owns. The growth counts at most Steam's playtime of the last two weeks, so a game seen for the first time brings
   no old playtime, and one game gets at most 24 hours per day. A BACKLOG game with new playtime becomes PLAYING, unless the user chose its status. When Steam returns no
   games, the library stays as it is.
5. `TimeToBeatService` and `MoodService` ask IGDB for missing or old time-to-beat, mood and genre data, for Steam
   games and manual games. Hidden games only get their playtime updated.
6. After the commit, `LibrarySyncedEvent` starts `AchievementSyncService` and `SpeedrunService` (backlog games only)
   in the background. The achievement sync also fetches Steam's global unlock rates with each schema, and sets
   FINISHED for a game whose ending achievement is unlocked while the user never chose its status.

The frontend polls `GET /api/library/sync-status` until the run ends.

At 01:30 `ImageRepairService` repairs the pictures reported through `POST /api/library/broken-images` (table
`broken_image`, one row per game and kind). Reports only count for Steam games the user owns.

### Background work

Background work runs on two pools (`AsyncConfig`; every `@Async` names one, ArchUnit checks it):

- The default pool (4 threads, queue 100) runs the syncs and the time-to-beat and mood refreshes a page view
  starts. A page view queues only games without a fresh row. When the queue is full, the task is dropped with
  a warning, and the next view or sync asks again. Known limit: a dropped sync still answers `202`; the user
  starts it again.
- The follow-up pool (2 threads, no limit on the queue) runs the achievement runs after a sync and every speedrun
  refresh. They hold a thread for minutes because of the Steam and speedrun.com pace, so they do not take the page
  views' threads. A task is small (a user or a few appids), and none is dropped.

Imports have their own two threads, and `game_import` is their queue. Scheduled jobs (nightly sync, image repair,
import watchdog) run on the scheduler.

```mermaid
flowchart LR
    View["page view, add, merge"] -->|"stale games only"| Pool["default pool<br/>4 threads, queue 100"]
    Sync["sync request"] --> Pool
    Synced["sync, after commit"] --> FollowUp["follow-up pool<br/>2 threads"]
    View -->|"speedrun"| FollowUp
    Pool --> Apis["Steam, IGDB, speedrun.com"]
    FollowUp --> Apis
    Import["POST /imports"] --> Lanes["GameImportWorker<br/>2 own threads"]
    Cron["scheduler"] --> Jobs["nightly sync, image repair, import watchdog"]
```

Core runs as one instance. The pools, the speedrun queue, the per-user achievement runs and the achievement locks
live in memory: a second instance would ask Steam and IGDB twice, and the database would stay consistent (sync
claim, upserts). A restart drops what waits in the pools; it is all refresh work and is asked again. Imports survive
a restart.

### Corrections by an auditor

An AUDITOR (active role) corrects a game's name, header image, speedrun link, moods, endless flag, genres and time to beat for
every user (`game_correction`, V20). The value is also written into the cached row (`steam_game`, `game_speedrun`,
`game_mood`), so no read joins the corrections. The refresh writers skip a corrected field: the Steam sync (name),
the image repair (image), the IGDB add (name, image), the mood, speedrun and time to beat upserts (SQL). `SteamGame` has no
setter for name and image: `applyCorrection` is the only Java writer, a refresh uses the guarded updates in
`SteamGameRepository`. "Back to source" (`DELETE`) writes back the source value from the first correction and
lets the next refresh ask IGDB or speedrun.com again. An image must be on a host of the frontend's CSP.
The time to beat is one field for both times (`normallyMinutes`, `completelyMinutes`, each empty or 1 to 60000,
completely not below normally) written into `game_time_to_beat`, so pile hours and stats use it; the IGDB `hastily`
time is not correctable. Every correction and every reset writes `admin_audit` (`GAME_CORRECT`,
`GAME_CORRECTION_RESET`). `GET /api/library/catalog` lets an auditor find and correct a game they do not own.
Known limit: a mood or speedrun refresh that waits on the row lock while an auditor saves can still write the
source value once (its guard sees the state from before the save); the auditor saves again. `save(new SteamGame(...))` on a
stored key also overwrites a corrected name; the sync and the IGDB add do it only after `findById` found no row.

### Merging duplicates

One game can have two keys: a Steam appid and minus its IGDB id, or two IGDB ids. An AUDITOR merges them for every
user (`GameMergeService`, V21). The key off Steam always goes into the other one; two Steam appids are refused,
since the sync would bring both back. In one transaction:

- A user with only the old row: the row moves to the new key with its status history (FK `ON UPDATE CASCADE`).
- A user with both rows: one row stays on the new key. Status, rating, notes and priority come from the row whose
  current status the user chose (both: the one chosen last); a gap is filled from the other row. Source, platform and playtime stay, unless the
  Steam copy is no longer owned. Both status histories stay; the merge itself logs no status change.
- The shared rows of the old key (catalog row, moods, speedrun, time to beat, corrections, broken images) are
  deleted. So are its queue pairs, its requests and its `game_edition` row; its set row goes to the kept key (see
  "Editions of one game").
- `game_alias` points the old key to the new one (kept flat), so a later add of the old IGDB game lands on the new
  key. `admin_audit` gets one `GAME_MERGE` row: who, both keys, both names and how many rows moved or were combined.

Merges made since V29 can be undone for 30 days (`POST /api/library/merges/{mergeId}/undo`, AUDITOR). The merge saves
a copy in its own transaction: `game_merge_undo` (the deleted shared rows as JSON per table, the link set after the
merge, the flattened aliases) and `game_merge_undo_user` (per user: both `user_game` rows as JSON, the history ids and the
reward ledger ids the merge moved). The copy holds notes, so it goes with the account. A nightly job
(`app.library.merge-undo-purge-cron`) deletes copies older than 30 days; older merges answer `410`.

```mermaid
stateDiagram-v2
    [*] --> MERGED: POST /games/{appid}/merge (copy saved)
    MERGED --> UNDONE: POST /merges/{id}/undo (30 days, newest first)
    MERGED --> FINAL: nightly purge after 30 days
```

The undo, in one transaction (the newest merge of a kept key goes first, else `409 NEWER_MERGE`):

- Shared rows (catalog row, moods, speedrun, time to beat, corrections, broken images, `game_edition`) come back;
  `game_alias` of the old key is removed and the aliases the merge flattened point to the old key again.
- A user with only the old row: the row moves back if it is still on the new key, still added by hand (no Steam
  takeover) and the user has no old row. It keeps its current status, rating and notes.
- A user with both rows: both rows are back if the kept row still exists. If the user did not change status, rating, notes or
  priority of the kept row since the merge (compared with a copy taken right after the merge), each row gets its own old
  values. If they did, both rows get the current values. The kept row gets its Steam ownership back when the merge
  had replaced it and nothing took the row over since. The old status history goes back to the old key; the log
  entries the undo itself causes are removed.
- Ledger rows the merge moved go back to the old key (so does what a moved user earned on the row after the merge), only when the user has the old row
  again. Nothing is paid twice. A game the user hid after the merge stays hidden on the restored rows.
- Links: when the kept key's set is still as the merge left it, the old sets come back with their `set_id`s; if
  links changed since, they stay and the old key comes back unlinked (the audit detail says "links kept").
- Users who deleted their account are not restored. `admin_audit` gets a `GAME_MERGE_UNDO` row; the `GAME_MERGE` row stays.

`GET /api/library/games/{appid}` follows `game_alias` when the user has no row on the asked key, and answers with the
kept `appid` (the frontend then replaces the URL). After an undo the alias is gone. The game search shows games already in ToBePlayed first, with their catalog name and image, and
games other users added by hand that IGDB did not return; `inCatalog` marks both (only those can be merged).

### Editions of one game

IGDB tells which library keys are one game: `game_edition` (V28) holds a "same game" key per library key, refreshed
like `game_mood` (stale check, nightly sync, add, detail view; `app.library.edition-ttl-days`). The key is
`version_parent` for an edition, else `parent_game` when `game_type` is 9 (remaster) or 11 (port), else the game's own
IGDB id; remakes (8) keep their own id. `null` when IGDB does not know the game.

- `game_link(appid, set_id)`: the keys an auditor linked, for every user. A set is every row with one `set_id`.
  An unlink takes one key out; a set of one is dropped. A merge puts the kept key in the merged key's set; when both have a set, the sets join.
- `game_link_pair` is the auditor queue (`OPEN`, `DISMISSED`, `LINKED`). IGDB pairs are inserted when the queue is
  read; a row in any state keeps the pair out for good. An unlink dismisses the pairs of that key with the former
  members. `game_link_request` holds who asked for a pair (the queue shows only the count). A user asks for a pair of
  two games of their own library; at most 20 open requests per user (a per-user advisory lock makes the cap hold for parallel calls), a dismissed pair answers `409 PAIR_DISMISSED`.
- Pile, stats, snapshots and the dashboard count a set once (`counted_user_game`, see "Stats"); the library
  filter `hideDuplicates=true` hides the lesser editions. Pile Points are unchanged: `STARTED` and `FINISHED` stay per row,
  and the `MONTH` bonus reads `backlog_snapshot`, which counts a set once.
- The game detail has `linked` and `editions` (the user's other listed games of the set); a game row has `editions`
  (size of the set in the user's library, `null` below 2).

### Deleted accounts

Auth drives it (see Auth's readme). `POST /internal/accounts/deleted` adds a tombstone in `deleted_user`: the user
gets no more syncs and leaves the landing stats and the pile snapshots at once. Auth's nightly job sends the whole
flagged set (`PUT`), then `purge` for users flagged 30 days ago (only tombstoned ones), then `prune` with all its
user ids, which removes the rows of users Auth does not know (refused for more than
`app.account-cleanup.max-orphans`, default 5). `app.account-cleanup.dry-run` (default `true`) only counts the rows.
The purge also removes the user's `account_name` row and link requests (`game_link_request`).

### Blocked users and Steam ID changes

Auth sends a user's state (`POST /internal/accounts/states`) after a Steam link, an unlink, a Steam ID another account
took by a verified link, a display name change, a block and an unblock; its nightly job sends every user (`PUT`). A blocked user is in
`blocked_user`: no nightly sync, no new sync run, not in the landing stats. The access token still works in Core until
it expires (max 15 min). A new Steam ID or no Steam ID takes the user out of the nightly sync; the next visit syncs
the ID of the token (Core's copy changes only then, so that sync sees the change and drops the old account's games). A visit with an access token that still has the old ID (max 15 min) can sync it once
more and put the user back on the nightly list; the nightly `PUT` at 04:15 takes the user off again, so at most one
nightly sync (23:55) runs with the old ID. Unblock and the nightly `PUT` lift the block.

### Manual games next to Steam games

A game the user adds by hand comes from IGDB (`ManualGameService`). Steam games and manual games share
`steam_game` and `user_game` and one key, the column `appid` (see `GameKeys`):

- `appid > 0`: the Steam appid. A manual game gets it when IGDB knows the game on Steam.
- `appid < 0`: minus the IGDB id, for a game that is not on Steam. Steam appids are always positive, so the two
  never meet.

`user_game.source` (`GameSource`) says who owns the row. The sync never changes or removes a `MANUAL` row, with
one exception: when the user later owns the game on Steam under the same appid, the sync takes the row over: it becomes
`STEAM`, and status, rating and notes stay. Only `MANUAL` rows can be edited (`PUT /api/library/games/{appid}/manual`)
or deleted (`DELETE /api/library/games/{appid}`). Time to beat uses the Steam appid for positive keys and the IGDB
id for negative keys. Search and add check both keys of a game. A game added under minus its IGDB id can still get
a second row when IGDB learns its Steam appid later and the user owns it on Steam; an auditor then merges the two
(see "Merging duplicates").

### Imports

Many games at once (V22, `GameImportService`): a list, a CSV or several picks of the game search become one
`game_import` with a `game_import_row` per line. `game_import` is also the queue of `GameImportWorker`, which has
two queues, each on one thread for all users:

- **matching**: IGDB searches with typo help. One thread, so an import never takes the whole IGDB budget from the
  interactive search; other users' searches may wait for the limiter meanwhile. Up to 5 candidates per line are
  stored whole (JSONB), so the save needs no IGDB call, and `exact` marks a best one with the line's name.
- **saving**: adds the chosen lines the way the manual add does (same keys and duplicate check) with
  `user_game.import_id` in the same transaction, then loads their time to beat and moods. A save never waits
  behind another user's big matching.

Each line is stored when done, so a stopped try goes on where it stopped. A worker sets `progress_at` when it takes
an import and after every line. A watchdog (every minute) stops a try without progress for 5 minutes
(`app.library.import-stall-minutes`): the import goes to the end of its queue and the queue gets a new thread. A
restart counts as a stopped try too. After 3 tries the import is `FAILED`; the user can try it again. One import
per user can be in the matching queue (a unique partial index).

```mermaid
stateDiagram-v2
    [*] --> QUEUED: POST /imports
    QUEUED --> MATCHING: matching worker takes it
    MATCHING --> READY: every line matched
    READY --> SAVING: POST /save (the review)
    SAVING --> SAVED: games added, time to beat and moods loaded
    SAVED --> UNDONE: DELETE /games (undo)
    MATCHING --> QUEUED: no progress for 5 min or restart
    SAVING --> SAVING: no progress for 5 min or restart (end of the queue)
    MATCHING --> FAILED: 3rd stopped try
    SAVING --> FAILED: 3rd stopped try
    FAILED --> QUEUED: POST /retry (stopped while matching)
    FAILED --> SAVING: POST /retry (stopped while saving)
    FAILED --> UNDONE: DELETE /games (the games added before it failed)
```

The undo deletes the games of the import that are still `MANUAL` (a game the Steam sync took over stays). A removed
row the import brings back gets no `import_id`: it holds rating, notes and history from before, so the undo keeps it;
`undo-check` counts those with a rating, notes, priority or a status change after the save.

## 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.

What the `local` profile does:

- Spring Boot's Docker Compose support starts each service's `compose.yaml`: PostgreSQL for Auth (port 5432) and
  for Core (port 5433), and in Core a WireMock container (port 8089, compose profile `local`) that answers in
  place of Steam, Twitch/IGDB and speedrun.com from `ToBePlayed-Core/local/wiremock`.
- Auth makes a new signing key at each start (`LocalKeysEnvironmentPostProcessor`, only in this profile), so no
  key is in the repo. After an Auth restart (also a devtools restart), log in again.
- Never use `local` on a server: it adds an admin with a public password. Auth refuses `local` together with
  `prod`.
- Auth adds the dev user (Flyway callback in `db/local`): role `PLATFORM_ADMIN`, so user, auditor and admin pages
  can all be tried ("View as" in the menu switches down), and a linked Steam ID that the recorded Steam answers
  belong to. The invite code for more users is `local`.
- Core checks the login with Auth's local JWKS, like in production.
- The browser still loads game pictures from the Steam and IGDB image servers; offline they stay gray.
- "Sign in through Steam" still opens the real Steam page (no key needed). The dev user is already linked.
- A request to WireMock that has no recording answers 404 and shows up in Core's log.
- To see an IGDB link suggestion: add Hollow Knight by hand (game search, PS5). It shares IGDB's "same game" key with the
  Steam copy, so the link queue (auditor) lists the pair with the IGDB badge.

To change the recorded answers, edit the JSON in `local/wiremock/mappings` (which request) and
`local/wiremock/__files` (the answer), then restart Core (it restarts the container). `LocalMocksParseTest` checks
that every answer still fits the adapters' response types. `http://localhost:8089/__admin/requests` lists what Core
asked.

### With the real services (profile `dev`)

Needs Java 25, Docker and your own keys.

```bash
./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:8090`. Health: `/actuator/health`, metrics: `/actuator/prometheus`.

In the `dev` profile a request without an `Authorization` header gets a dummy user, so the API is usable without a
running Auth service.

Configuration comes from environment variables:

| Variable | Purpose |
|---|---|
| `JWT_JWKS_URL` | Auth's `/.well-known/jwks.json` |
| `JWT_PUBLIC_KEY` | Ed25519 public key (Base64 of the X.509 DER bytes), used until the first JWKS load. Set this, `JWT_JWKS_URL`, or both |
| `STEAM_API_KEY` | Steam Web API key |
| `IGDB_CLIENT_ID`, `IGDB_CLIENT_SECRET` | Twitch app for IGDB (optional) |
| `DB_PASSWORD` | Local PostgreSQL password |

## Tests

```bash
./mvnw clean verify
```

Runs unit, WebMvc, persistence and integration tests plus the ArchUnit rules. Needs Docker for Testcontainers.
In integration tests background work runs on the calling thread, so a test reads the result at once.
`AbstractMockedAdaptersTest` mocks Steam, IGDB and speedrun.com, `AbstractMockedLibraryServicesTest` the sync and
the refreshes of the picker and stats pages; each is one shared Spring context. Race tests wait until the other
connection waits for the row lock (`awaitLockWaiter`), not for a fixed time.

## Roadmap

- Sale alerts for wishlist games.
- Move the adapter DTOs out of `api`, then turn on strict module independence in ArchUnit and drop the adapter
  exception of the "`service` never uses `api`" rule.

## License

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