The Xbox app ships the same version of Halo Campaign Evolved as Steam (CU4, 1.112.1610.0), built for the WinGDK platform. MJOLNIR's Lua mods and cooked containers run on it unchanged; the native halves did not until 2026-10-10, because they call into the exe at addresses measured on the Steam build.
What differs
| Steam | Xbox app | |
|---|---|---|
| Install root | steamapps\common\Halo Campaign Evolved | XboxGames\Halo- Campaign Evolved\Content |
| Binaries | Meteorite\Binaries\Win64 | Meteorite\Binaries\WinGDK |
| Exe | HaloCampaignEvolved.exe, PE timestamp 0x8a03f777 | HaloCampaignEvolved.exe, PE timestamp 0x3d94a571 |
| Simulation DLL | timestamp 0x6a7a740a | timestamp 0x6a7a741d |
| PlayFab / Party | PlayFabMultiplayerWin.dll, PartyWin.dll | PlayFabMultiplayerGDK.dll, Party.dll |
| Online subsystem | Steam | GDK (Xbox Live) |
| Containers | pakchunk<N>-Windows.* | pakchunk<N>-WinGDK.* |
- The simulation DLL is the same code. The two copies differ in 29 bytes (timestamp, checksum, debug id) plus Steam's signature. Every simulation RVA and byte patch holds for both.
- The exe is a separate link of the same engine code. Engine functions match instruction for instruction at other addresses. Platform code differs: the Xbox app's exe has no Steam subsystem.
- The exe cannot be read on disk. The Xbox app's file protection refuses
even an administrator. The DLLs beside it can be read. To study the exe,
dump it from the running game:
OpenProcess+ReadProcessMemoryoverSizeOfImagefrom the module base. The image comes out relocated to the load base, so for Ghidra rewrite each section's raw pointer and size to its virtual ones, and setImageBaseto the load base. - The game is a packaged app. It starts through
shell:AppsFolder\Microsoft.198377053870B_8wekyb3d8bbwe!AppHaloCampaignEvolvedShipping. The launcher derives that id from the install'sappxmanifest.xml. Noms-xbl-<product id>://link is registered for it.C:\Program Files\WindowsApps\Microsoft.198377053870B_…is a junction to theXboxGamesfolder, so files installed there are what the game sees.
Porting addresses
native/map_registry keeps one address set per exe (exe_profiles, keyed by
the PE timestamp). The Xbox app's set was found from Steam's with
native/signatures/port_rvas.py:
python -P native/signatures/port_rvas.py <steam exe> <dump> --dump 36D29F0 ... gD349040 ...
All eight functions matched whole. The three globals were read back from 5
to 8 functions each. FModuleManager::Get needs a longer pattern than the
script's first try, because it differs only in a thread-local slot index
(0x154 on Steam, 0x13c on the Xbox app). Do the same after a game
update.
| Steam | Xbox app | |
|---|---|---|
FModuleManager::Get | 0x36D29F0 | 0x32EBF10 |
FModuleManager::GetModule | 0x36D3540 | 0x32ECA60 |
| FName hash | 0x3709650 | 0x33214B0 |
| FName from a view | 0x36FCC60 | 0x3316080 |
FCoreDelegates::MountPak | 0xD349040 | 0xC31C940 |
HandleMountPakDelegate | 0x4694C00 | 0x4239420 |
FCoreDelegates::OnUnmountPak | 0xD349248 | 0xC31CB48 |
HandleUnmountPakDelegate | 0x4694CC0 | 0x42394E0 |
GMalloc | 0xD4B7428 | 0xC47DF60 |
levels glue Get | 0x7B7DFC0 | 0x7749E60 |
| levels glue row insert | 0x7BAC140 | 0x7777F20 |
The AssetRegistry's GetFirstPackageByName is slot 30 on both builds.
native/lobby finds everything by code pattern. On the Xbox app every
pattern matches once, as on Steam, except two. One is the Steam join hook,
replaced as described below. The other is the presence session's 4-player
literal, which does not exist on the Xbox app, so the lobby DLL searches for
it only on Steam. The fireteam still shows 16 slots there.
Joins by connection string
On Steam a join is the task manager's own GameRichPresenceJoinRequested_t
handler, called from OnlineTick (docs/multiplayer_servers.md). The Xbox
app's counterpart is the invite handler that FOnlineSessionGDK registers
with XGameInviteRegisterForEvent:
- The handler (
0x663bf40,(session, const char *uri)). It readsconnectionString,sender(orjoineeXuid) andinvitedUser(orjoinerXuid) from the URI's query:name=up to the next&, with no URL decoding. It then queues the join on the subsystem's next tick. That path is taken while a console variable reads true (it does in CU4). The other path reads an MPSDhandle. - The invited user must be a signed-in user. The queued task looks the
XUID up with
XUserFindUserById. A miss startsXUserAddByIdWithUiAsyncand waits for it on the game thread. - The queued task stores a pending invite (sender, the user handle, the connection string), which the session acts on later. The sender XUID is only stored.
mjolnir_lobby delivers the request from the GDK subsystem's own tick:
FTSTickerObjectBase::Tickatsubsystem+0x70(0x6686270), whose first act is to service the invite's task queue.- The session is at
subsystem+0xF8, and points back at+0x2D8. - The player's XUID is the first live key of the subsystem's XUID →
XUserHandlemap at+0x208(0x38-byte elements).
The URI it sends is
ms-xbl-mjolnir://inviteHandleAccept/?invitedUser=<xuid>&sender=<id>&connectionString=<string>.
PlayFab connection strings (cv2:<lobby>|<n>|kv1:<key>) contain no &.
Why a delivered invite went nowhere
Delivering the invite was not enough. Every join from the Xbox app stopped
after join: delivered, with no PFMultiplayerJoinLobby and no error, and
the player saw nothing happen. The GDK subsystem does its whole part, all in
one tick: the task parks the invite (0x664bb80), the tick builds a search
result and hands it on (0x664bd60), and the session raises
OnSessionUserInviteAccepted (IOnlineSession vtable +0x348) with
success. Two things then lose it:
- Nobody listens there. The GDK session's only listener is its Online Services adapter, and nothing subscribes to the join event that adapter re-raises. The game listens on the other session interface, PlayFab's. Its adapter is bound with the same delegate type, and its join event has one handler.
- The key is the GDK one. The result carries the connection string as
CUSTOMJOININFO. PlayFab's join readsCONNECTIONSTRING. Raised on the PlayFab session with the GDK key, the game left its fireteam and showed FAILED TO JOIN without calling PlayFab.
So for a join it delivered (within 30 s), mjolnir_lobby hooks the GDK
session's trigger slot and does two things instead. It renames the setting
in place. That is safe while the settings map has a single hash bucket
(fewer than four settings), where a lookup walks every element. Then it
raises the invite on the PlayFab session. It finds that session by its
adapter's delegate instance: a heap scan for the vtable of the GDK
session's own listener, once per game. The adapter is the instance's
+0x18, and the session is the pointer the adapter holds whose invite
listeners include that instance. Every other invite reaches the GDK session
as shipped.
A real Xbox invite raises the same event on the same session, so it probably goes nowhere in the shipped game either. That is untested.
Verified 2026-10-10 on the Xbox app with a dummy connection string: the game
left its fireteam, called PFMultiplayerJoinLobby with the string (returned
0) and showed its own FAILED TO JOIN when PlayFab refused the fake key. That
is what the Steam build does with the same string.
Verified 2026-10-10 between two real players: a Steam player joined a game hosted on the Xbox app.
Crossplay
Steam and Xbox app players share one PlayFab lobby service, so either can host and the other joins from FIND GAMES or a link. The game refuses a player from another platform when crossplay is off for either side. On the Xbox app that follows the Microsoft account's "You can join cross-network play" privacy setting. A Steam joiner dropped a few seconds after connecting to an Xbox app host points at that setting.