Players sign in to the website, and admins to the admin panel, with an OpenID Connect provider (DorkNet:SignIn:*). The first sign-in links the provider identity (issuer + subject) to a DorkNet account with that account's password; after that the provider sign-in is enough. - Cookie sessions per host, keys persisted in the database so redeploys keep people signed in; unlink, deletion or a ban ends the session. - Link attempts limited to 5 failures per 15 minutes per identity and per username. - Admin panel: "Sign in with ..." exchanges the session for the usual admin tokens via GET /api/admin/v1/sso (IsAdmin still required). - Downloads: optional DorkNet:Downloads:RequireSignIn serves the manifest and files only to signed-in players; the Downloads and Quest pages offer sign-in instead. - Website: sign-in in the header, /link and /account pages. The game client's login is unchanged.
23 KiB
DorkNet Admin Page
The admin page is the browser operations console for DorkNet. In the
microservices deployment it lives at https://admin.<domain>/ and is
served by the web service from the built Vite SPA assets.
The admin backend endpoints are under /api/admin/v1. Requests from the
SPA are same-origin, so https://admin.<domain>/api/admin/v1/* is also
handled by the web service route on the admin host. The same
/api/admin/* route family is available as a path-owned moderation slice
for non-admin-host routing, so /internal/routes can show moderation as
the owner for that path family.
Runtime Routing
| Item | Value |
|---|---|
| SPA source | DorkNet.Server/admin-ui |
| Build output | DorkNet.Server/wwwroot/admin |
| Container path | /app/wwwroot/admin |
| Public browser host | admin.<domain> |
| API base used by the SPA | /api/admin/v1 |
| Login endpoint | POST /api/admin/v1/login |
| Same-origin admin-host service | web |
| Path-routed admin API slice | moderation |
DorkNet.Gateway preserves the original Host header when it proxies to
backend services. That matters for the admin page: the admin.<domain>
host route must reach web, and web must contain
/app/wwwroot/admin/index.html.
Authentication
The first account created on a fresh database is promoted to admin so
there is always a bootstrap operator. The login form posts to
/api/admin/v1/login; the SPA stores the returned values in browser
localStorage:
| Key | Contents |
|---|---|
dorknet.admin.token |
DorkNet admin JWT |
dorknet.admin.me |
Admin identity object used by the layout |
With external sign-in
configured, the login page also shows Sign in with …. It signs in with
the provider on the admin host, then GET /api/admin/v1/sso swaps that
session for the same token pair the password login returns. The identity
must already be linked to an account on the website, and the account still
needs IsAdmin = true.
Every admin API request sends Authorization: Bearer <token>. Protected
actions are guarded by AdminOnlyAttribute, which checks the resolved
player row still has IsAdmin = true. A 401 clears the local session
and sends the browser back to login.
If the admin host is public, put it behind Cloudflare Access or an equivalent outer access control. The DorkNet JWT is still required after that outer check.
Navigation
| Section | Routes | Workflows |
|---|---|---|
| Overview | / |
Live ops dashboard, online players, active sessions, quick kick/ban/broadcast actions |
| Moderation | /players |
Player directory, bans, reports, per-player ban/grant/gift/password/avatar/account actions |
| Activity | /activity |
Admin audit log, per-player request logs, QA test cases + GitHub issue linking |
| Content | /rooms, /rooms/:id, /import-room, /content |
Room list/detail, room import, instances, leaderboards, community board, loading tips, 3D Charades word lists |
| Operations | /broadcast, /settings |
Server broadcast, server toggles (signups, everyone-is-friends, profanity filter, imported-room version clamp), signup codes, weekly challenges, Play menu tags, Rec Center doors, game config values |
Several older admin URLs are kept as redirects:
| Old route | Current route |
|---|---|
/bans |
/players?tab=bans |
/reports |
/players?tab=reports |
/gift, /passwords, /grants |
/players |
/audit |
/activity?tab=audit |
/logs |
/activity?tab=logs |
/rr-originals, /instances, /leaderboards |
/rooms |
/community |
/content?tab=community |
/loading-tips |
/content?tab=tips |
/signup-codes |
/settings?tab=signup |
/import-room-legacy still exists for the legacy room importer.
3D Charades word lists (/content?tab=charades)
The March 2023 client fetches a charades deck at card-box spawn from
GET api/activities/charades/v1/words/{source}, where {source} is one
of three baked CardBox.cardSource slots — Charades,
CharadesAprilFoolsDay, and Icebreakers (verified in the 2023.03.21
il2cpp dump). The response is a JSON array of
{ "EN_US": "<phrase>", "Difficulty": <int> } (Difficulty is the client
CNMMMNJJDMM enum: 0 easy, 1 hard, 10 very hard, 20 icebreaker).
The admin tab exposes:
- A library of unlimited named word lists (
CharadesWordListEntityrows), each with a per-card difficulty. Paste-import accepts one phrase per line with an optional| easy|hard|veryhard|icebreakersuffix. - Live card slots — three dropdowns binding each client slot to any
library list. Switching a slot just repoints its binding
(
ServerSettingsEntity.CharadesSlotBindingsJson); it takes effect on the next card-box refresh (room rejoin). An unbound slot falls back to the built-in list seeded for it.
Three built-in lists (Default / April Fools / Icebreakers) are seeded on first boot and bound to their slots. Seeding is idempotent — admin edits survive restarts.
Profanity filter toggle (/settings)
ServerSettingsEntity.ProfanityFilterDisabled gates the server-side
api/sanitize/* filter. When on, every sanitize route returns input
unchanged and treats all text as clean, so room/invention names and chat
are never censored. Off by default; checked per request so it takes effect
immediately.
Imported-room version clamp (/settings)
ServerSettingsEntity.RoomBlobVersionClampDisabled gates the CDN's
PersistedRoomData version clamp
(RoomDataBlobService.ClampVersionsFor2023). Rooms imported from modern
RecNet zip exports carry room-data blobs whose top-level version varints
(field 1 DEPRECATED_RoomPersistenceVersion, field 30
PersistedRoomVersion) are far past what the March-2023 client knows —
a Sep-2025 save stamps version=131 against the client's maximum of 16
(RoomDataBlobService.Client2023MaxPersistedRoomVersion; the proto's
LatestVersion=19 came from a newer build's dump and this client rejects
it) — and the client rejects the whole room with its "update Rec Room to
visit this room" gate before spawning anything.
This clamp only addresses the room-header version gate. A room whose
CircuitsV2 chip graph (circuit_data, field 18) was saved in a newer Rec
Room is rejected by a separate CircuitsV2 version gate regardless — check
the client's Player.log, where the stack frame under "Booting player to
dorm" names the failing subsystem (CircuitsV2Manager = the circuit
graph, not the header). That case is not fixable by this clamp.
While the clamp is active (the default), the CDN serve path rewrites the
two varints down to the 2023 maxima on the way out for .room / .meta /
.dat blobs. The rewrite is wire-surgical: every other byte is preserved,
already-compatible blobs pass through byte-identical, and nothing stored
in S3 is modified. Toggle endpoint:
POST api/admin/v1/settings/room-blob-version-clamp with
{"Disabled": bool}. Clamped responses are served with a 60s edge TTL so
flipping the toggle lands within a minute.
CircuitsV2 limits (/settings)
ServerSettingsEntity.CircuitsV2LimitsJson backs the
UGC.CircuitsV2.Limit.* gameconfig keys the client reads into
RecRoom.CircuitsV2.Config.LimitConfig. The panel's headline setting is
the network budget: retail allows a circuit 5 RPCs per frame and
1024 bytes per 60 frames, and silently drops every send past
that, which is the usual cause of a busy CV2 room appearing to desync.
DorkNet ships both budgets at 1,000,000,000.
Two things about this are counter-intuitive enough to be worth stating:
- The budgets are spent over a rolling window measured in frames, so widening a window makes the limit stricter, not looser. Raise the budget, not the window.
- The byte budget is checked before the RPC budget, so raising
Rpcson its own leaves circuits blocked. Both have to move together. InkPerChipis a cost charged against a 30000-ink room cap that is hard-coded in the client, so lowering it is what gives circuits more room. DorkNet ships it at 0.
The room-size limits ship at about 10x retail: Compute 500000, Chips 30000, CloudValues 1000, and UnculledChips/VisibleChips 3000. CloudValuesAutosaveSeconds stays at 300.
Values are clamped on save, and most of the clamps are crash guards
rather than taste. Rpcs of 0 divides by zero in the client's CV2 debug
subscreen. A time window allocates one array slot per frame, so an
unbounded window is an out-of-memory kill at circuit init. Chips sets
the size of four list preallocations. InkPerChip × CloudValues is an
unchecked int multiply. Compute is the client's only runaway-loop
guard, so it has to stay finite. Full
reasoning and the disassembly it came from are in
docs/recroom-2023-circuitsv2-limits.md.
Endpoints: GET/POST api/admin/v1/settings/circuits-limits. The client
builds its rate limiter once when circuits start, so players must
rejoin their room before a change takes effect.
Per-sub-room max players (/rooms → room detail)
Player caps are two separate settings, and conflating them is the easy mistake here:
| Setting | Column | Scope |
|---|---|---|
MaxCapacity |
RoomEntity.MaxCapacity |
What the ROOM advertises; flows into RoomInstance.MaxCapacity for matchmaking and into the room-level MaxPlayers of the details payload. |
MaxPlayers |
RoomSceneEntity.MaxPlayers |
What ONE sub-room caps itself at; appears per entry in the details SubRooms[]. |
The room-level figure must not be read off sub-room 0 — that collapses the two into one and makes an admin edit to a sub-room silently move the room's advertised capacity.
Sub-room caps have two write paths onto the same column:
- In game (room owner only) — the "Max Player Count in This Subroom"
slider issues
PUT rooms/{id}/subrooms/{sub}/maxplayers. Unlike most of its siblings this one does not form-encode:NLDBPDCNNCF.MNBPCGJNLNP(Int64, Int64, Int32) builds aBLNOGFGHIIFwhose single JSON key ismaxPlayers, with verb 3 = PUT. It reads the reply back asFGCPNAACHIK, so the response is the full room-details shape and has to carry the new value. - Admin (any room, ownership not required):
GET api/admin/v1/rooms/{id}/subrooms— every sub-room with its cap, matchmaking flag, sandbox flag and current data blob. Row ids go out as strings; the SPA is JavaScript and mangles integers past 2^53.PUT api/admin/v1/rooms/{id}/subrooms/{subRoomId}/maxplayerswith{"MaxPlayers": int}.subRoomIdis the sub-room's order index — the id the client and the rest of the wire use — not the row's primary key.
Both paths clamp to 1..80, matching the room-level cap in
rooms/{id}/props: 0 would make a sub-room unjoinable, and a shared range
keeps the two settings from disagreeing about what is a legal value.
Enforcement caveat carried over from RoomEntity.MaxCapacity: this is the
advertised cap. The client never sets Photon RoomOptions.MaxPlayers, so
hard enforcement still needs a ClientMod Photon patch.
RRO room data blob — Maker Pen and chips in a Rec Room Original (/rooms → room detail → Sub-rooms)
Baked Rec Room Originals (Rec Royale, Paintball, Laser Tag, …) are seeded
with an empty RoomSceneEntity.DataBlobName and empty
RoomEntity.CurrentDataBlobName. Every details builder and matchmaking
resolver (RoomsController.CurrentOrSyntheticDataBlobName,
GoToController.ResolveInitialDataBlobAsync,
MatchPlayerController) turns that into an empty DataBlob, and the 2023
client reads an empty blob name as "this room has no room data": it skips
the persistence download entirely, so there are no room roles (no Maker
Pen despite the CDN's RRO role overlay), no CV2 room-settings chips, no
room mood engine and no saves. Real Rec Room served these rooms a blob —
Rec Royale's game-rule chip (RecRoyaleSolos_GameRuleWrapperCircuit /
RecRoyaleSquads_GameRuleWrapperCircuit, prefabs in resources.assets)
is delivered that way; the Frontier scene has no baked game rule.
The Sub-rooms tab shows, per sub-room, the blob the client will actually
be told to download (effectiveBlobName) and, for baked originals, two
actions:
- Enable Maker Pen blob —
POST api/admin/v1/rooms/{id}/subrooms/{subRoomId}/blob/enable. Stamps the synthetic default name (room_{id}_dorknet_v8.datfor the entry scene,room_{id}_dorknet_v8_sub{n}.datotherwise) on the scene and, for sub-room 0, on the room. No resolver changes are involved: they already return a non-empty name verbatim, andCdnControlleralready answers that name for AG rooms with the RRO-editable blob (permissive room roles, no objects) while nothing is saved.IsRROis derived fromIsAGRoom, so the client keeps loading the baked map and only additionally fetches the blob. Idempotent; a no-op on a sub-room that already has a blob. Seeded originals such as Rec Royale have noRoomScenesrow at all ("Scenes 0"); the list synthesises a virtual "Home" entry (isVirtual: true, max players = room capacity) and enable creates the real row with the game-side defaults (RoomsController.GetOrCreateSceneForMutationAsync). What the CDN hands back for that name while nothing is saved is a per-room stub built byRoomDataBlobService.BuildRroEditableBlob: the skeleton of a real save (Rec Center'sroom_100_v1.datwas the reference) —version16 (the client's ceiling; the old roles-only stub was stamped 19 and the client fetched it and silently loaded the room with no roles, i.e. no Maker Pen),activity_id= the room'sLocationReplicationId,last_save_time, empty connectable graph / object model, scene settings and the fourroom_mood_*blocks, plus the permissive role collection. The miss path now runs the same 2023 version clamp as the S3-hit path. It also carries the room's own three game-role objects (a default role named after the room, plus the built-in "In-Game" and "Eliminated" roles, well-known GUIDs shared by every room save — Rec Center and Stunt Runner both have exactly these). The 2023 client grants Maker Pen off the player's game role (ICreatorRole.CreatorRoleCanUseMakerPenonGameRole), so a blob with no role objects leaves even a co-owner with nothing to hold — that was the second reason "enable + co-owner" still showed no Maker Pen. The default role setscan_use_maker_pento true, matching the permissive room-role overlay; the built-in roles mirror the reference saves. For the two Rec Royale rooms the stub also carries the room's game-rule chip (RecRoyaleSolos_GameRuleWrapperCircuitprefab id983db81a…,RecRoyaleSquads_…857f7500…, read from the client's spawnable-tool table; a saved Squads chip is header + prefab id + tag + node id + an empty wrapper payload), so the match logic exists on first join without anyone spawning it by hand. - Upload save —
POST api/admin/v1/rooms/{id}/subrooms/{subRoomId}/blob/upload?mode=full|basics, multipartfile= a realPersistedRoomDatasave from ANY client version (.binpb/.room/.dat, e.g. a RecNet export of a custom Rec Royale room). The server rewrites it for this room (RoomDataBlobService.PrepareUploadedRoomBlob):activity_id→ the room'sLocationReplicationId, dropsoutput_log,sub_room_idand every field the 2023 schema doesn't know, clamps the header versions to the 2023 maxima, and inbasicsmode keeps only the game-role (field 50) and game-rule (fields 40/41) objects while stripping the circuit payloads (circuit_data, connectable graph,circuit_v2_data— a newer CircuitsV2 graph is rejected by the client's own CV2 version gate no matter what). Stored under a freshroom_{id}_upload_*.dat, recorded inRoomDataBlobs, and the sub-room (and room, for sub-room 0) is pointed at it. On serve the CDN applies the RRO role overlay and clamp exactly as for Rec Center's saved blob. Response carries asummary(viewsIn/viewsKept/roleViews/gameRuleViews/versionIn/versionOut). - Reset to baked scene —
POST api/admin/v1/rooms/{id}/subrooms/{subRoomId}/blob/reset. Clears both pointers. This is the escape hatch if a save left the room unplayable. Saved blobs stay inRoomDataBlobs, so the in-game restore list can bring one back later.
Both are audited (enable_subroom_blob, reset_subroom_blob). Who can
then save the room in game is the normal rule in SaveDataCore: owner,
any admin, or an accepted co-owner (RoomRoles.Role == 0). Grant a
co-owner role from the Roles tab if a non-admin should be able to save.
Procedure that makes Rec Royale playable AND editable:
- Enable the blob on the Solos room's sub-room 0 (and the Squads room's).
- Join as an admin; the Maker Pen is available. Spawn the room's game-rule
chip (
RecRoyaleSolos_GameRuleWrapperCircuitin Solos,RecRoyaleSquads_GameRuleWrapperCircuitin Squads) — via the palette's Game Rules section, or the debug console'sDebugSpawnToolif it is not listed — then save. Without that chip there is noBattleRoyaleManagerin the scene and the match never starts. - From then on the saved blob (with the game-rule chip and whatever room-settings chips were placed) is what everyone loads. If a save breaks the room, Reset to baked scene.
Tests: RroRoomBlobTests.
Why Rec Royale still has no Maker Pen after all of the above
The Maker Pen permission check (AGMBHDNGOEH.IJEIEBGJJLB) first runs a
room precondition (LALJEMHEIBH) that, via MLKAFHHHEGA.GHPIBJPNOFF →
PNPLLCBBMJP.JGPBNEHFLJA, looks the current room up in
AGRoomRuntimeConfig — a table baked into the client's
resources.assets — and returns a per-original boolean. Rec Center and
Stunt Runner allow creation there; Rec Royale does not. No blob, role or
server response changes that table, and it is consulted only for rooms the
client considers originals (IsRRO = our IsAGRoom).
Server-side experiment: clone Rec Royale, flip the clone's Rec Room
Original (IsRRO) checkbox off on the General tab
(rooms/{id}/props IsAGRoom: false), upload a real save to it, join. If
the client still loads the Frontier scene for a non-original, Rec Royale
becomes an editable template; if the scene fails to load, the table gate
has to be bypassed client-side (ClientMod postfix on GHPIBJPNOFF).
Test cases ↔ GitHub issues
QA test cases can file and track GitHub issues. Configure with:
| Key | Meaning |
|---|---|
GitHub:Token |
PAT with issues:write on the target repo |
GitHub:Repository |
owner/name, e.g. DorkSquadRR/DorkNet |
GitHub:ReconcileIntervalMinutes |
Sweep interval, default 15; 0 disables sweeps but leaves the manual endpoint working |
Without a token nothing breaks. IGitHubIssues.IsConfigured is false, the
endpoints answer 503 with the reason, and the background reconciler logs once
and idles. A server with no GitHub credentials is a normal deployment.
UI: Activity → Test cases (/activity?tab=testcases). File or unlink an
issue per case, or reconcile them all. The buttons disable themselves with an
explanation when GitHub is unconfigured rather than failing on click.
Sub-room caps live on the room page: Rooms → a room → Sub-rooms.
Endpoints (admin-gated — these are not part of the 2023 client's surface):
POST api/testcasemanagement/v1/testcase/{id}/issue— file and link. Idempotent: a case already carrying a live issue returns that issue instead of filing a duplicate, so sweeping a failing pass repeatedly is safe. The issue body carries the description, key, room, test pass and tester comments; labels areqaplus the case's own tags.DELETE api/testcasemanagement/v1/testcase/{id}/issue— unlink only. The issue itself is left alone; closing someone's issue as a side effect of tidying a QA link would be the wrong call.POST api/testcasemanagement/v1/issues/sync— reconcile now. The background service runs the same code path on its interval, so manual and automatic cannot drift.GET api/admin/v1/testcases[?passId=&status=]— the list the SPA renders, with each case's issue number and link.
Each of the three also answers under /api/admin/v1/testcases/..., because the
admin SPA speaks that prefix exclusively.
Where the link is stored
In TestCaseEntity.JiraBugUrl — the field Rec Room's own QA tooling used for
the bug filed against a failing case. Reusing it means no migration, which
matters more here than the name: the Postgres path is EnsureCreated-only and
never replays migrations, so a new column has to be added twice (migration and
an idempotent Ensure* patch) or it is simply missing in production. A field
that already exists everywhere has neither problem, and the admin UI renders it
already. A leftover genuine Jira link is ignored rather than misparsed — the
issue number is only read from a URL matching the GitHub issue shape.
What the reconciler will and won't change
| Issue | Case status | Result |
|---|---|---|
| Closed | Failed | → Passed |
| Open | Passed | → Failed |
| anything | Claimed / NotYetTested | untouched |
Closed means fixed: the case that was failing on that bug now passes, and reopening the issue undoes it. Only those two states are the reconciler's to move — a case a tester has Claimed, or one nobody has run yet, is never rewritten underneath them.
Build And Deploy
The service Dockerfile builds the admin SPA for service images that need
static assets. In the microservices stack the web image must include:
/app/wwwroot/admin/index.html
/app/wwwroot/admin/assets/*
The admin SPA calls relative URLs, so it does not need a configured API
origin as long as the browser is opened on https://admin.<domain>/.
The native admin mobile app points at the same public admin host; see
../DorkNet.AdminMobile/README.md.
Smoke Checks
Check the static host:
curl -I https://admin.yourdomain.com/
Check that protected admin API routing is alive. A 401 is expected
without a valid admin token:
curl -i https://admin.yourdomain.com/api/admin/v1/stats
Check the built assets inside the running web container:
docker exec <web-container> sh -lc 'find /app/wwwroot/admin -maxdepth 2 -type f | sort | head -50'
Useful logs:
docker logs <stack>-gateway-1 --tail=200
docker logs <stack>-web-1 --tail=200
docker logs <stack>-moderation-1 --tail=200
For normal browser admin traffic, start with gateway and web.
Check moderation when you are testing /api/admin/* through a
non-admin host such as api.<domain>.
Troubleshooting
| Symptom | Check |
|---|---|
admin.<domain>/ returns 404 |
web logs for probe="/app/wwwroot/admin/index.html" exists=False; rebuild/redeploy the web image |
| Admin page loads but actions 404 | Gateway /internal/routes, then whether the request host is admin.<domain> or a path-routed host |
| Login or actions return 401 | Token expired, local session cleared, or the player no longer has IsAdmin = true |
| Browser spins forever | Open the browser network tab, then check gateway and web logs for the stuck URL |
| Upload/import fails near 100 MB | Use the chunked importer; the SPA uses 50 MB chunks to stay below Cloudflare request limits |