Skip to content

Latest commit

 

History

History
998 lines (952 loc) · 82 KB

MODLOG.md

File metadata and controls

998 lines (952 loc) · 82 KB

MODLOG — Rewind Together, session of 2026-10-14

Journal for this session. PROJECT.md is the brief; VERSION.txt the changelog.

Goal (agreed with the user)

"Ready for Nexus": no config editing, friend clicks Join Game in Steam, tunnel starts itself, sync/interaction verified, docs for strangers.

Paths

  • Game: D:\SteamLibrary\steamapps\common\RetroRewind (AppID 3552140)
  • Deployed mod: ...\RetroRewind\Binaries\Win64\ue4ss\Mods\RewindTogether\Scripts
  • Source (edit here): %USERPROFILE%\OneDrive\Desktop\RewindTogether\Mods\RewindTogether\Scripts
  • Deploy: sh dev/deploy.sh (keeps the deployed config.lua; --with-config to overwrite)
  • Logs: ...\Binaries\Win64\RewindTogether.log, ...\ue4ss\UE4SS.log
  • Saves: %LOCALAPPDATA%\RetroRewind\Saved\SaveGames (Player_Save1 = main, Player_Save2 = small)

Backups / restore

  • Saves: %USERPROFILE%\.universal-modder\backups\retrorewind-saves\20261014-140052.zip (um backup restore retrorewind-saves)
  • Config folder: ...\backups\retrorewind-config\20261014-140052.zip
  • Deployed config.lua: ...\ue4ss\Mods\RewindTogether-config.lua.pre-claude-backup
  • UE4SS-settings.ini: ...\ue4ss\UE4SS-settings.ini.pre-claude-20261014

State found at start

  • Code is 1.6.1 per VERSION.txt, but Net.VERSION still says "1.5.8" (rt.status lies).
  • Last two-player test: 2026-08-05 (1.5.8). 1.6.x never tested with a guest.
  • Oct 9 run was solo; no errors.
  • Deployed config.lua is an older file with PeerSteamIds set (a friend's).
  • Unverified guesses: Config.SyncedState property names, the interaction function name list in sync.lua.

Dev tooling added

  • devbridge.lua — only active if Binaries\Win64\RewindTogether.devbridge exists. Slot A = first instance, B = second. py dev/bridge.py A rt.status, py dev/bridge.py A "lua return UE.NetModeName()".

Log

2026-10-14 — two-instance testing on one PC (works)

  • Instance A: Steam launch, hosts. Instance B: RetroRewind-Win64-Shipping.exe RetroRewind started directly with env SteamAppId=3552140 (otherwise Steam just refocuses A). py dev/coop_test.py up|down|log. Bridge slots A/B; traces in Win64\RewindTogether.trace..
  • Clicks only work at the window's native 1920x1080 client size; the FIRST click after focusing is eaten (activation) — click twice. Menu points: PLAY (242,675), slot 1 (750,140).
  • Joining from the main menu works (OpenLevel 127.0.0.1 → Client in ~2 s). No save needed.
  • CXX header dump: %USERPROFILE%\RetroRewind-decomp\CXXHeaderDump (outside repo). Money = Core_Gamemode_C.Money (int64 cents); Level/Experience/Lifetime Experience/Store Value. Clock = WeatherSystem_C (Day, Hour, Minute, Year, Season, DayTime, Current Weather, Day of the week); its minute timer: "Stop or pause Timer - Add one minute". HUD = UI_Player_Interface_C (Change Money Value / Change Time / Change Week Day / Change XP Value), built only by Core_Gamemode_C ("UI Game Interface" property). Interactions: IInteraction_C::Interaction(bool& Accept Interaction); player side Player_Character_C "Trace from Camera for Interaction", input events InpActEvt_Input_Use_*. Hazard stripes = StoreExpansion_C walls of unbought expansions (Unlocked, "Unlock Expansion"(bInstant)). Store sign = StoreName_C (Name FText, Color, Size, "Set StoreName Visual"). The game itself has unfinished MP scaffolding: GameInstance "is Multiplayer"/"ExpectedPlayers", GameMode "Look for Multiplayer", "Player Controller logged".

Crashes found and fixed (1.7.0)

  1. Host died the instant a guest joined: rpcchan passed a Lua string for ClientMessage's FName param → UE4SS access violation (0x70). Reproduced deterministically on a spare instance; FName("...") form is fine. Smoke test now rejects non-FName.
  2. Host/guest died 40–145 s into sessions inside Lua's GC (garbage pointers like 0x6b636a6d): LoopAsync runs Lua on UE4SS's async thread concurrently with game-thread Lua. Switched to LoopInGameThreadWithDelay (present in this UE4SS build). Soak: 6 min, 0 crashes (was 3/4). Keybind callbacks now only ExecuteInGameThread; the always-on LEFT_MOUSE_BUTTON bind removed.
  3. Engine's own spectator ServerSetSpectatorLocation calls during login were replayed as guest interactions → now marked with Roll=77.25.
  4. localgm stripes: Safety.guard returns one value → nil compare every guest join.
  • UE4SS marshalling gotchas (each measured, each a native crash or refusal):
    • Lua string -> FName param: CRASH (use FName("x")).
    • Lua string -> TSoftObjectPtr param (WeatherSystem "Load Level for Season"): CRASH.
    • Out params (bool&): UE4SS 3.0.1 does not advance past the out-param table, so any UFunction with an out param FOLLOWED by more params cannot be called (LevelStreamingDynamic::LoadLevelInstance). Empty "" also refused for an FString.
    • Struct fields come back keyed by the FName table's casing: FRotator "pitch" (lowercase).
  • Season sub-level: host streams /Game/VideoStore/asset/level/Season/L_Store01_Season- as LevelStreamingDynamic at identity; a guest never loads it (no GameMode) → no snow, different fog/colour. Snow Accumulation 0.5 vs 0.
  • Ghost-mode stripes: NavCollisionZone_C + M_GhostMode_NavBlocker; "Toggle Ghost Mode Blockers"(false) on the guest hides all (fixed in worldsync.lua).

Native helper (RewindNative.dll) — works

  • UE4SS's C++ SDK needs the private Re-UE4SS/UEPseudo repo (Epic-linked GitHub only) → not available. Instead: package.loadlib (enabled in this UE4SS) loads native\RewindNative.dll; exports are lua_CFunctions that use UE4SS's exported RC::LuaMadeSimple::Lua wrapper (ctor/get_integer/set_integer/...; dtor does NOT close the state — LuaMadeSimple.cpp @0196ef29).
  • Player-0 redirect: swap UFunction::Func (+0xD8 on this build, found by scanning for the first pointer into the exe's .text) of GameplayStatics GetPlayerController/Character/Pawn/ CameraManager with thunks that call the original, then replace host→guest results while a window is open. rt.native test PASS live. DLL pins itself (Lua 5.4 FreeLibrary()s on reload).
  • Build: native\RewindNative\build.bat (VS 2022 Community has MSVC 14.44; no install needed).

Guest interactions — pick-up works

  • FINDINGS (%USERPROFILE%\RetroRewind-decomp\FINDINGS.md, bp*.txt): RMB _23 = interact (Trace from Camera for Interaction from self.Camera, box (1,2,2), dist = Move Tool 200), LMB _20 = per-tool trace (hand: physics / place when holding), E _1 = hand tool, V _2, B _3, K _0. 410/414 GetPlayer* calls pass 0. Item/hand code works on Self. No BP RPCs.
  • guestinput.lua: guest sends (code, camera loc/rot) via ServerSetSpectatorLocation with Roll = 100+code; host sets the guest pawn's Camera to that view, then in a Native.asPlayer window calls Look At Test + the input event on the guest's pawn. MEASURED: guest picked up "Rain Woman" (host: guest Object Hold = videotape, host holds nothing; guest sees it in hand).
  • HUD calls go to the GameMode's single widget (the host's) → during a replay the GameMode's "UI Game Interface" is set to nil and restored (host HUD showed the guest's tape before).
  • Host disables the guest pawns' actor tick (their Look At Test highlighted guest targets on the host's screen; the game never checks IsLocallyControlled).

NPC looks — guest crash hunt

  • Guest died 3× on a Foreground Worker (UE task thread), EXCEPTION_ACCESS_VIOLATION reading 0xffffffffffffffff, same top frames (+3bbbb73, +343e969, +2573f5a, +259ca39), 2–3 min after NPC looks were applied. NPC sync off: clean 5 min. Looks off/state on: clean 4.5 min.
  • Fixes: (1) never CreateDynamicMaterialInstance under a fixed name (same-name NewObject destroys the live MID in place) and don't re-apply unchanged looks; (2) body swap via SetSkinnedAssetAndUpdate(mesh, true) — SetSkeletalMeshAsset is SetSkeletalMesh(..., false) and keeps the old pose buffers.
  • Minidump thread names via the ThreadNamesStream (dev: scratchpad md2.py).
  • ROOT CAUSE FOUND (soak 8: 8 min clean with looks on, was 109–161 s every time): npcsync cached loaded asset objects in a Lua table across frames; GC freed them after their NPCs despawned and the cache handed out dangling pointers. Crash instruction disassembled (capstone on the exe): cmp byte [rax+0x89] after reading [component+0x560] = USkinnedMeshComponent::SkinnedAsset / UStaticMeshComponent::StaticMesh (header dump offsets). Assets are now re-resolved every time. (Queued delivery + SetSkinnedAssetAndUpdate + unique MID names kept as good practice.)
  • Mouse-free host load for tests and the launcher: GameInstance."Save Slot"="Player_Save1" then OpenLevel(L_Player_Store_01) — from FINDINGS 3e. launch.json host may carry "slot".

Put-back on the guest + the lookup cost (2026-10-14)

  • Guest saw a put-back tape floating. Cause: on the guest bReplicateMovement was false on 395 of 396 tapes. Two reasons, both MEASURED: (1) a Blueprint actor copies only a precomputed property list from its class default when it spawns — a tape spawned on the host after its CDO had both flags came out bReplicates=true, bReplicateMovement=false; (2) with the host's CDO patched to true, the flag was never sent (Unreal sends only where a value differs from the class default). Fix: never patch bReplicateMovement on a CDO; host sweep checks every actor's own flags continuously (1 slice/s, 4 classes); guest sweep flips the flag and calls OnRep_ReplicatedMovement (stored ReplicatedMovement non-zero, not attached). Hand-setting the flag on the guest tape made the host's nudge visible at once. Release nudge removed.
  • PERF: UE4SS FindAllOf/FindFirstOf = 15–17 ms per call (hit or miss, any class); StaticFindObject 1–2 ms. Wrapped and counted live for 10 s: host 60% of wall time in lookups (GetAllPlayerControllers 22/s), guest 53% (patchGuestMovement retrying non-existent bare class names every tick). Engine route (ue.lua): GameInstance (one walk per process) + GameplayStatics GetNumPlayerControllers/GetPlayerController/GetGameMode/GetGameState (µs); KSL MakeSoftObjectPath→Conv_SoftObjPathToSoftObjRef→Conv_SoftObjectReferenceToObject = 0.02 ms (null path → wrapper around nullptr, check IsValid); GetAllActorsOfClass (out param last: works) 479 tapes in 4 ms; GetActorOfClass 0.008 ms. Conv_SoftObjPathToSoftClassRef is NOT callable. FindAllOf("BlueprintGeneratedClass") returns nothing in this UE4SS. Per-tick cache closed by UE.endFrame(). Objects re-found by path (UE.pathOf/findObject) — HUD widget 0.03 ms.
  • First engine-route build: host sweep patched nothing and expansions read 0 — GetAllActorsOfClass out-table elements are RemoteUnrealParam (need :get()). UE4SS source (0196ef29 LuaUObject.cpp call_ufunction_from_lua): params live in DynamicUnrealFunctionData{uint8_t data[0x200]} on the stack, never destroyed → engine allocations made for params/returns leak (out arrays: 8 B/element; UE.findObject's FString path: ~0.6 KB/call measured, 100k calls +124 MB vs control +19 MB). Struct returns are converted to plain Lua tables.
  • After fix: host 727 actors patched at session start; expansions 6; guest tapes 397/397 with bReplicateMovement; guest put-back verified both screens (rack slot 1966,-1190,139, nothing near the hand). Lookups 2.4% of wall on both (was 60/53%). Launcher: 170/170 tests.
  • Leak budget: findObject memoized per tick, native classes cached for good, NPC sync lists AI_Base_Character_C once instead of 4 classes. Sweeps: host 1 slice/2 s, guest 1 slice/2 s.
  • Second leak, worse: UE4SS make_ref()s every out-param table into the Lua registry and never unrefs it → every actor list (+ its RemoteUnrealParam wrappers) lives forever. Registry grew ~4/s; guest live Lua heap 31 MB and rising. Fix: UE.releaseOut(t) (nil the registry entry whose value is t; newest refs are at the end of the array part), UE.withHit for K2_Set* FHitResult.
  • Guest neon red again on one run: the sign lights were applied in the second the guest joined. Re-running "Update Materials for Lights"/"Update for Lights" later turned it blue → re-run them on each world message (2 s) for 30 s after the first application.

Guest HUD: what you hold and look at (2026-10-14, late)

  • The plan in PROJECT.md 0 (answer GetGameMode with the HUD widget) would have done nothing: the widget's UI_Interface_C events are all empty stubs (bp UI_Player_Interface__game_ui @1025-1031); only Core_Gamemode_C's events forward to "UI Game Interface".
  • Census of the dump: 78 Blueprint GetGameMode calls, each followed by a cast of the result - 51 Cast<Core_Gamemode_C>, 24 ToInterface<UI_Interface_C>, 1 each to Ghost_C, OptionApply_C, MainMenu_GameMode_C. The HUD display calls (product in hand, context menu, cursor style, look-at detail, cursor visibility) all live in the UI_Player_Interface_C function library (core/blueprint/function - same short name as the widget in core/widget/game_ui).
  • RewindNative v2 (rt_hook_null, rt_standin_arm/disarm/stats): GetGameMode answers null with Core_Gamemode_C's CDO only for VM frames (FFrame Code non-null) of that library whose next statement is EX_Let/EX_LocalVariable/EX_ObjToInterfaceCast(UI_Interface_C). Liveness of the CDO and the widget is checked on each answer through UE4SS's exports (IndexToObject/GetUObject, UObjectBase GetInternalIndex/GetObjectFlags/GetClassPrivate/GetOuterPrivate).
  • MEASURED live: armed on its own (GetOffset_Internal works; CDO frame present); Lua's UE.GetGameMode() still invalid on the guest; ~1000 Blueprint GetGameMode calls/s on a guest; look-at gave the hand cursor and the "USE" hint on the guest's HUD.
  • Product in hand did NOT appear from the stand-in alone: the host's replay picks up, the guest sees the tape on its hand (attachment replicates), but the guest pawn's own Object Hold stays empty, so Set Object Hold (which calls "UI - Set Visible Product in Hand") never runs there. Calling the pawn's Set Object Hold with what is attached to its "AttachObject Socket" made the panel appear ("THE MAILBOX'S BURNING POINT MOVIE ACTION"). hud.lua now does that at 4 Hz while the stand-in is armed. Spawned tapes have different names on host and guest (videotape_C_2147469324 vs _2147475982), so the guest reads its own socket.
  • End to end: pick up (panel), swap at the RESERVED rack (panel follows), throw to the floor (panel gone), host HUD never shows the guest's tape, rt.leave (disarmed, CDO released), rejoin (re-armed on the new widget). 19,823 answers, 0 stale, no crash, no breadcrumbs.
  • Tooling: dev/coop_test.py up stops at client_point() when dev/gw.py has no WinDrivePid.ps1 (RT_WINDRIVE); the bridge steps it runs work by hand. py dev/gw.py ps is not sorted by start.

Guest actions: till, build mode, skin tool, selling, end of day (2026-10-14/15)

  • Read off the Blueprints first (bp/MembershipMachine, CheckOut, Wall, AC_MeshSkin, Core_Gamemode).
  • Every guest action had arrived with pitch 0: UE4SS fills an FRotator param by the name table's casing ("pitch"). UE.rotArg. A floor-level standee was out of reach until then.
  • replication.lua wrote bReplicates before SetReplicates -> runtime spawns (build-mode boxes) never reached guests. Toggling SetReplicates on the host showed the box at once; fixed order.
  • Guest walls/floors/ceilings/counter were the level defaults: only the save loader applies AC_MeshSkin. skinsync.lua (Load Asset(true) resets the skin to 0 - measured; use false with "Savegame data loaded" = false). Guest matched the host 2 s after joining.
  • Till: entered for the guest by the host (Walking Mapping Context is host-only), glide stopped
    • finished as the guest, guest's game told to follow, W -> action 7, Space -> action 9, drawer
    • display mirrored (it stayed shut on the guest), exit spot fix (navmesh is host-only: the guest landed at 0,0,92), host hint cleared via the pawn's Init Cursor and Interface. Card customer (+$75.25) and cash customer ($25 for $21, $4 change) both done by the guest.
  • Skin tool: Hud.withStandInFor(class) arms the stand-in for the clicked object's flow class for one call; the game's picker opens on the guest; UI_SkinMode "Material Selected_Event" hooked -> action 8 -> host Set new Skin/Mesh -> skinsync. Real mouse clicks re-skinned carpet + shelf.
  • Selling: worked; host HUD missed the money (HUD hidden during replays) -> refreshed after.
  • End of day: no travel, guest stays connected through summary into day 2; guest presses ignored while "End of Day Interface" is in the viewport (the reopened door would start a 2nd one).
  • A guest crashed once in Bodies.tick after a hot reload re-placed it beside the host while its till collision was off (it fell out of relevancy and came back). The HUD stand-in was then switched off by its probation file; rt.safety reset brought it back.

Guest NPCs: walking, hair, faces, thought bubbles (2026-10-15)

  • Same NPC side by side (the store's employee, probe of both copies): guest Character Scale 0 vs 0.996, AnimBP play rate 0.0 vs 0.80, Hair/Facial-Hair attached to the capsule (inside the body) vs HeadTop_EndSocket, mouth/eyes plain MICs vs MIDs (Frame, Eyelid), Widget3D_QuestRequest shown vs hidden, badge "Text Name" vs "Adam", actor tick off vs on.
  • Cause (bp/AI_Base_Character): BeginPlay -> Bind Init casts GetGameMode to Core_Gamemode_C; on a guest it fails, delays 1 s and retries forever, so "Step 2 of creation Skin" (Init Default Variable, Adjust Scale, Init Graph, tick on) never runs. AnimBP_Base: play rate = Max Speed Walk Scale * (1 / Character Scale); divide by zero answers 0.
  • npcsync.lua: the look now carries size, walk rate, bubble height, eyelid height, hair/beard materials, all 8 body slots (mouth and eyes too) and the staff name; the state carries the bubble and its icon. Guest: SetWorldScale3D, the NPC's own "Attach Child to Skeletal", the mouth frame + Surprised/Angry morphs per emotion (AI Emotion's table), bubble shown/hidden and turned to the guest each tick. The NPC's own tick stays off: Limit Walk Speed by rotation would turn every NPC toward GetControlRotation = 0 (no AI controller on a guest).
  • Found on the way: the guest's own AI_Director (started by ActorGatherer on every machine) spawned 12 walk-bys in a few minutes (Role Authority, no controller, frozen at spawn points, bubble up). K2_ClearTimer(director, "Event - Timer") + K2_DestroyActor on Role Authority NPCs.
  • Verified: feet swing through a stride on the guest (GetSocketLocation RightFoot, camera on the NPC - bones only refresh when drawn), hair and face match in screenshots, an angry face shows on both (Frame 0.72, Angry 1.0), a popcorn bubble raised on the host shows on the guest turned to the guest and closes with it, a re-armed director is stopped and its spawn removed.
  • Not mirrored: eye direction (animated every frame on the host), the NPC's look-at-player head turn (each copy's AnimBP does it for its own player). Player bodies (bodies.lua) are still the bare NPC body: bald, grey.

The "Bodies.tick" guest crash, breadcrumb blame, hot-reload re-placement (2026-10-15)

  • Dump UECC-...1919342548779CCA2F2A5BAEF49442F2: GameThread, AV at UE4SS+0x22ba30. RAX = FNV-1a basis, R11 = FNV prime, R9 = the faulting address, R8 = 0x20, R10 = 0x3ff. Disassembly (pefile + capstone, .pdata for function bounds): MSVC _Forced_rehash of a static unordered_map<string_view, wstring> + shared_mutex at UE4SS .data 0xf41118, the memo UE4SS+0x228570 that UE4SS's UObject member access (0x22b590, via the __index/__newindex thunks 0x22a2f0/0x22a310/0x22acd0) uses for names. Keys come from 0x848050 = lua_tolstring + strlen + lua_remove: pointers into Lua's own strings. 4 nested pcalls below UE4SS's callback = Log.guard -> Safety.guard(dress) -> setRelLocation -> UE.withHit -> comp:K2_SetRelativeLocation.
  • Log: the guest (pid 37636) had been hot-reloaded at 23:40 and 23:51 (Ctrl+R reloads BOTH copies - the "two loading lines"), crashed at ~23:52:3x dressing the host's re-created pawn.
  • Blame: both copies shared RewindTogether-inprogress.txt; the host's guarded calls (bodies on the guest's lingering pawn every 2 s, rpcchan sends) deleted the guest's breadcrumb, leaving only the stand-in's probation file.
  • RewindNative 3: signature (lea r12,[lock] ... mov rbx,FNV basis; then "cmp [rip+size], rax" naming lock+0x18), layout validated under the SRW lock (pow2 buckets, mask, 2 iterators per bucket, ring walk == count), unlink all nodes and leak them (DLL is /MT). test_namecache.cpp: real MSVC map, keys in VirtualAlloc'd pages, freed -> the 513th insert faults (the dump); after the purge 2000 inserts are fine; installed UE4SS.dll: exactly one match at +0xF41118.
  • Lua: native.lua repairs at boot (before any module) and from a __gc sentinel at lua_close; UE.TryGet/TrySet/TryCall pin names, UE.pin for others, dev bridge keeps compiled chunks.
  • safety.lua: RewindTogether-inflight..txt per running copy (token = addresses of the CoreUObject/GameplayStatics CDOs, beat, phase, probation + its phases, guard stack), written in place and flushed; slot = own token (hot reload) or first record not written for 10 s. Blame: innermost guard > probation feature if phase idle/its own > report the phase only.
  • spawn.lua: no first placement for a pawn older than 10 s (GetGameTimeSinceCreation); placeGuest refuses a pawn with collision off or MovementMode 0.
  • Live, two copies: name cache found on both (207/185 names, 512 buckets); 4 reloads, 9 repairs per process, 670-710 names dropped; the host never re-placed the guest ("made 61/110/289/320s ago"); a till-like guest (collision off, MOVE_None) not moved by a reload or rt.bring. Out of range needs the guest to MOVE: the server's view of an idle client's camera stays where it was (host pawn kept 24 s at -39,868; gone 3 s after a W press). Three cycles out of range and back: host pawn ..474695 -> ..470808 -> ..463792 -> ..459517, dressed each time through K2_SetRelativeLocation, no crash. Guest hung inside a guard (stand-in put on probation), killed, restarted: blamed only the guard; host's record untouched; rt.safety reset cleared it.

Host-only visuals on a guest: titles, signs, register, change, preview, end of day (2026-10-15)

  • Titles: ShelveTitle_C values (genre, tag, empty, visible) from the host, matched by place and yaw (a double-sided shelf has two titles on one spot at yaw +-90). "Default Text" was really TextSign_C's "User Text" (6/6 with words on the host, none on the guest): sent as "signs", set with UE.setText then "Update Text". After: 73/73 titles, 6/6 signs the same.
  • FText writes: UE4SS writes 0x18 into a 0x10 FText. End-of-day rows vanished (Font Size 4.98e+20), a TextSign's "UI Text Command" pointers were garbage. UE.setText = KismetSystemLibrary SetTextPropertyByName (FText last). worldsync's store sign used the bad write too.
  • Register: "Add Product Scanned to the list"(name, nil tags, price, type) gave $0.00 at type FURNITURE and the guest hung ~30 s later (killed; records kept in the session scratchpad). The guest now creates UI_Screen_ProductScan_C (WidgetBlueprintLibrary Create) and fills it with the *ByName setters, then AddChild. Money on the table: Replication.patchActor per actor in "Array of Spawned Money on table ref" ($10/$5 bills seen on the guest's counter).
  • Preview: ReceiveTick hook on Player_Character_C (runs after the BP), pawn matched by address, Ghost Tick with the furniture class's CDO for 0.35 s windows (the box's "Objects in Inventory" is not replicated). First live run: green on the guest at a spot where the host threw the box. Cause: Look At Test asks "Trace from Camera for LookAt Ghost" (reach by tool + 50 when looking at something) before Ghost Tick. Added; after: green at 1320,-1700 and the host placed the shelf at 1320,-1700 yaw -90; aimed close at a wall/shelf the guest showed no preview and the host threw the box (the box at 2426,-23 on both).
  • End of day: UI_EndOfTheDay_C on the guest (buttons disabled), statistics texts, XP texts, ProgressBar percent, WidgetSwitcher page and three visibility digits (txt_LevelUp and two SizeBoxes - without them the guest showed "LEVEL UP!" over an empty box). Then UI_NewDay_C.
  • User report mid-session: customers handed the guest a basket instead of money. bpdump: BTTask_Checkout-Pay "Setup Pay With Money" spawns Money_NPC-Stash_C into the hand on the host (also the membership-card and console-rent tasks); AI_Base "Toggle Visiblity Basket" only sets Has Basket, AI_Client's override hides the basket's StaticMesh - and npcsync only set Has Basket. Now: patchActor on each NPC's "Money Stash", and the guest calls Toggle Visiblity Basket on change.
  • Live, after a restart of both: customers' baskets show/hide on the guest with Has Basket (vis false at the till, true while shopping). A first watcher crashed the GUEST: it named stash:GetAttachParentActor() with :GetFName() the instant the cash replicated, before the attachment had (AV reading 0x18 = null's NamePrivate, under UE4SS+0x229dd7). Probe bug, not the mod; noted in PROJECT.md. Then, with Adam idle at 22:00, the guest took the till, talked to the customer (damaged-tape fee), and she held out cash: Money_NPC-Stash_C at 1617,-1510,114, attached to the same customer on both (role 3 / 1), basket hidden on both, bills in her hand on both screenshots.

Crash record: a clean exit is not a crash (2026-10-15)

  • Game quit paths (bp): the store's ESC menu opens L_Boot (map load -> Hud.reset -> stand-in trusted); only UX_Boot_Menu calls QuitGame. The bad cases were window close / Alt+F4, Steam's Stop and killed copies, which left the same record as a crash.
  • Measured (one copy, WM_CLOSE via taskkill without /F): the process is gone in 2 s; no Lua finalizer runs (UE4SS does not lua_close at exit) and a RegisterHook on GameInstance:ReceiveShutdown never fires. The Safety "closed" finalizer therefore only ever fires on hot reloads.
  • RewindNative: rt_crashes_since(t, pid) (report folders under %LOCALAPPDATA%\RetroRewind\Saved
    Crashes since t whose CrashContext.runtime-xml matches), rt_pid, rt_pid_alive (a running process with this exe's name). Native tests cover all three.
  • safety.lua: records carry the pid; EVERY ended record is judged (not only the claimed slot) and cleared; uncertain verdicts (probation / phase-only) wait for Safety.settle(crashesSince, pidAlive), called by main.lua right after the native load: no report -> not a crash; report -> blamed; unknown -> blamed as before unless the record was closed in order. A fresh record whose pid is dead is judged too (a copy killed and relaunched within 10 s - measured 7-8 s).
  • Live: window closed with the stand-in on probation -> "stopped without crashing"; killed and relaunched 8 s later -> same, killed record cleared; real crash (string for ClientMessage's FName, AV at 0x70, report pid 63400) -> reported against its bridge phase, stand-in not blamed. A crashed process that hangs after writing its report keeps counting as running.

Posters, NPC speech, level-up unlocks on a guest (2026-10-15)

  • Posters: PosterFrame_C (+ _Standing, _Small) "Specific SKU" is a plain BP int, set by Load or the SKU menu; Apply Material for Film based on SKU (@4081) calls Film_Data Create Film Data, which with SKU 0 rolls a random film. The skins' Materials Loaded runs it on a guest with 0. skinsync now sends "posters" (place;yaw;sku) beside titles/signs. Live: 4/4 frames the same SKU and title on both; a guest frame set to 0 rolled 10463373 and was back to 1794141 within the 10 s re-check; "They Came for the Inmate" on both screenshots.
  • Speech: Widget3D_AiSpeakInWorld (screen space, under Head_Socket) filled only by AI_Base "AI Talk about a problem" / "AI Toggle a permanent problem" (callers: AI tasks, CheckOut, BlackMarket, flyers). Host RegisterHook on both (re-hooked when the UFunction's address changes), queue, "npct" from the tick; guest: widget's Set Text and Show Message (or the NPC's own event when a sound path comes along), Set Text And Show Message Permanent / Start Animation Hidden Message. No natural line in ~2 min of open store; the host called a customer's own event (nil sound, as the replay passes nil object params): the guest's copy at 1568,-1260 showed the same words at opacity 1.0, both screenshots.
  • Unlocks: UI_EndOfTheDay "Add Widgets For Level Unlock and Unlock it" runs from the XP animation (GameMode-only). dayend sends u=system.movie.game per child; the guest Creates UI_Statistic_Levelup_Unlock_C, SetBytePropertyByName x3, AddChildToHorizontalBox; a movie genre also gets UI_GenreUnlock_C at z 12. Live: XP +300 (7112 -> 7414, level 11 -> 12 at 7300), clock to 19:58, sign closed, door Interaction once customers left, UX_Button_Base event -> level 12 "Cotton Candy Machine" + "Repair Desk" (4.0.0, 15.0.0) on both; second day +750 -> level 13 "Kid" (0.12.0), genre popup made on both and gone on both by the next look.
  • User report mid-test: "there is also no guest pawn". Host's copy of the guest at 2495,41,94 in MovementMode 3 (the join placement), the guest at 1356,-1352: aim.py teleports the client only, and with the server not correcting the client nor taking its position, nothing converges. Moved the host's copy by hand for the test; window key input never reached the guest copy, so bServerAcceptClientAuthoritativePosition could not be tried with real moves. Proposed as a follow-up.

Conversations, NPC shaking, a guest who could not walk (2026-10-15, overnight)

  • Conversations: talksync.lua. TalkSystem_C is spawned on the host for GetPlayerPawn(0) during a guest's replayed click (= the guest's pawn); BeginPlay draws UI_Talk_C into the HOST's viewport and runs a 0.5 s camera timeline on GetPlayerPawn(0) (the host). Hook on ReceiveBeginPlay (post): Force Stop LookAt + RemoveFromParent; host reads txt_talk/txt_NameClient/Zone_ButtonAnswer and sends "talk"; guest builds UI_Talk_C + UI_Talk-Answer_C (Answer Type/Button Text/Number Text), UI input + cursor; hook on UI_Talk-Answer_C:Button Clicked Event -> action 11 (Pitch = type); host clicks the hidden widget's "Clicked on Answer Button Event" as the guest. End Talk System hook / actor gone -> "talk e=1". Guest gone -> answer 1 (Quit). Live: Adam's 4 answers on the guest only, host camera untouched, "Never mind" closed both; guest killed mid-talk -> the host's talk ended once its controller went, Adam back to task 2/3.
  • Sampling NPCs per frame from the dev bridge: StaticFindObject by full path inside a ReceiveTick hook took the guest to 2.5 fps (and leftover hooks keep running - use a new global per sampler). Resolving once inside the hook and reusing: ~70 fps.
  • NPC shaking: a customer at (915,-1735) on the guest: capsule z 90 <-> 116..119 every other frame, mesh z up to 101 - stepping onto something the host did not have. Actors near: the guest had two sets of the snack display's SnackBox_Snack_C / UISelect triggers / rotating meshes; the extra set (role SimulatedProxy, the host's replicated child actors, their children spawned locally) had trigger collision on. "SnackBox" in Config.ReplicatedClasses; all 4 instances are child actors. Hand-disabling the extra set: a customer 10 units from a trigger walked level. Fix: never replicate child actors (switch off if replicating). SetReplicates(false) did NOT remove the guest's existing copies; a fresh session has none (4/8/4 on both).
  • "No guest pawn" -> a guest that could not walk: the guest's own pawn had role 1 (SimulatedProxy) on the guest, RemoteRole 1 on the host. UE 5 SetReplicates(true) calls SetRemoteRoleForBackwardsCompat(SimulatedProxy) every time; the sweep called it on the already- replicating pawn. AddMovementInput and window W did nothing; LaunchCharacter moved it (simulated). RemoteRole = 2 written on the host (SetAutonomousProxy is not reflected): the guest walked and the host's copy followed. With bServerAcceptClientAuthoritativePosition too, the host's copy ended exactly on the guest (1300,-1312). That broke the join placement (both left at 2367,177,-408): server-side moves never reach a trusted client - Replication.beforeHostMovesGuest untrusts for 1.5 s, the engine's ClientAdjustPosition carries the move, restoreTrust re-trusts.
  • Final build, regression on two copies: join (autonomous, placed 2495,41 on both, trust back), walk 311 units + jump (same spot on both), both bodies visible, rt.bring, till enter/exit (exit fix on the guest's own pawn, same spot on both), tape off a shelf and back, build mode (placed at the preview spot), conversation, customers past the snack display (mesh steady).
  • Sprint check (guest's own position every frame via a ReceiveTick hook, MaxWalkSpeed = Run Speed 550 on the guest only, AddMovementInput each frame, 150 frames): both flags on -> moved 994, 2 reversals under 1 unit, host's copy at the guest's spot (1501,67); control with the host's corrections on (both flags false) -> moved 744, 33 reversals up to 84.9 units.
  • Review fixes (code review of all of the above): guest conversation closes after 10 s without word from the host (host resends every 3 s); a departed guest's talk is quit only while the host is walking (End Talk System resets player 0 = the host); "talk ended" retried until sent; NPC speech not queued with no guest connected, permanent lines tracked regardless; trust window 5 s (corrections only answer a guest's move); no class-default patch for a class with child-actor instances; a failed unlock fill is retried; conversation answers are taken during the summary.
  • Last sanity run (final code): join placement on both, trust back, walk followed exactly, conversation open past 5 s, answered, closed. Then a dev probe (FindFirstOf("TalkSystem_C") on a destroyed talk + HasAnyFlags) crashed the host - probe only; the mod finds talks by path and checks UE.alive (object-array flags) first. Records deleted, saves unchanged (6677069f).
  • Launcher one-click setup (user: "make the launcher handle all the stuff for people so installing rewind together is stupid simple"). Core/SetupPlan.cs decides UE4SS / install-or-update / switch on; MainViewModel.SetUpAsync runs it when the launcher opens, when the game closes, before HOST/JOIN, and from the Setup line's single button. A running game: asked once, closed with CloseMainWindow (never killed). The release zip now carries the pinned UE4SS zip (Mods/UE4SS/, sha-checked by package.py and again by the launcher) and its MIT LICENSE; the port tools moved to Extras/ (mod messages updated). 181 launcher tests pass. Checked through UI Automation on fake game folders only: fresh install with no download, 1.6.1 -> 1.7.0 keeping config.lua, switched-off left alone then switched on, stand-in game closed by the launcher and by hand, exe run alone from %TEMP% -> "Unzip the download first". The real game was not touched.
  • Real-game close check (user OK; one copy via Steam; saves backed up and byte-identical after, GameUserSettings.ini restored from .pre-claude, UE4SS-settings.ini restored byte-identical). WM_CLOSE from outside: game quits in 1.6 s (main menu) / 1.4 s (in the store, Sun 8:00), DLLs free at once, crash guard record "idle (between ticks)", nothing disabled. Found: with UE4SS's GUI console shown (1.6's installer set GuiConsoleVisible = 1; the launcher keeps that file), Process.MainWindowHandle is the console, and CloseMainWindow closed only the console - the game kept running. GameProcess.CloseAsync now posts WM_CLOSE to the visible "UnrealWindow" of the process (fallback CloseMainWindow). Verified through CloseAsync itself on the real game with the console in front: closed in 2.4 s. Stand-in fallback rechecked on the final exe. Zip rebuilt. Note: the pinned UE4SS's default ini has no console at all (ConsoleEnabled = 0, GuiConsoleEnabled = 0), so friends on a fresh launcher install have no window for rt.* commands.
  • No UE4SS window on fresh installs (user chose: no extra window, point docs at the game's console). The pinned UE4SS ini has ConsoleEnabled = 0 / GuiConsoleEnabled = 0, but its ConsoleEnablerMod (on) gives the game's own console on ~, # and F10, and rt.* commands (RegisterConsoleCommandHandler) print into it via the output device. Checked on the real game at the main menu: F10 once = one-line console, twice = full console, three times = closed; rt.status ran and its report showed in the full console. (Typing into the full console through WinDrivePid kept dropping keys - a driver quirk; UE's full console stays open after Enter.) FRIEND guide, README, VERSION.txt, release-notes draft and mod messages (commands.lua header, config.lua comments, spawn.lua rt.fixpawns hint, mainmenu.lua warning) now say "press ~ twice". F6 still only writes to the log while SafeMode is on. Zip rebuilt; saves unchanged, settings restored. Not deployed to the real game's mod folder.
  • F6 visible without the UE4SS window. UConsole reflects nothing (CXXHeaderDump), so the game's console can't be opened from Lua; instead Commands.showStatus -> Hud.notice puts a 3-line status in the HUD's UI_Error box (Display Error(FString), "{br}" breaks, Animation Time raised to 800 ticks = 8 s) whenever the overlay is off (SafeMode default). Overlay.isEnabled added. Smoke: 5 checks (no HUD, notice text/colour, return value, 8 s, overlay on). Real game (user OK; saves unchanged, GameUserSettings.ini restored, bridge files removed): F6 at the menu logs only; in the store the real key shows the centred box; timer measured 268 ticks after ~2.7 s of 800. Deployed hud.lua/overlay.lua/commands.lua (only this session's changes differed). Gotcha in PROJECT.md.
  • Launcher, for 1.7.1 (uncommitted until released):
    • Self-update from GitHub releases (Core\SelfUpdate.cs): banner + Update, SHA-256 checked, swap with rollback, restart keeping --data-dir/--game-dir. End to end against a stand-in GitHub on 127.0.0.1: about 14 s.
    • Mod check before joining (Core\ModFingerprint.cs): the host publishes rt_ver/rt_cmods/ rt_cnames/rt_smods/rt_snames in Steam rich presence. The guest's friend list and JOIN say what differs, with "Join anyway". Unit-tested with simulated Steam data only; not yet tried with two real accounts.
    • Discord status (Core\DiscordPresence.cs): local RPC pipe, application "Rewind Together" (id 1560334255547093155, art asset "logo"). Real Discord accepted the activity (name "Rewind Together", logo resolved, button URL attached) from a throwaway console using the launcher's own classes; the status was cleared afterwards. Set for the game's pid. Not yet seen with the real game running: whether Discord then hides its own "Playing Retro Rewind" is unconfirmed.
    • Logo: launcher\tools\make-wordmark.py draws the REWIND TOGETHER sticker (RT pink/cyan, purple die-cut outline) to dist\discord\logo-wordmark.png.
    • 223 launcher tests, 5 clean runs in a row.
  • Launcher: Nexus Mods browser (uncommitted). Mods tab > NEXUS MODS lists Retro Rewind's mods from Nexus's public GraphQL API (game retrorewindvideostoresimulator / 8977, 248 mods, 233 without adult ones), with search (nameStemmed MATCHES), sorts, thumbnails, and adult mods hidden by default. Downloading for a player needs their Nexus account (free: the site's Mod Manager Download; Premium: the API), and public apps may only sign players in after Nexus registers them. So Download opens the main file's page (?tab=files&file_id=N) and a Downloads-folder watcher installs the file when it lands. The user asked for "Download" in place of "Get it on Nexus". Records keep nexusModId/nexusVersion (badges); the host publishes rt_nexus, so a guest gets "Get their mods". .7z/.rar now open through Windows' tar.exe. Checked live on a fake game: listing, search, Download opening the file page, and a stand-in file under Employee Mod's real Nexus name installed as nexus-53 1.5 with an INSTALLED badge. Found on the way: a 1.7.1 launcher (the user pressed Update in the test window, which self-updated it for real) rewrote installed.json without the Nexus fields. Fixed: Load restores the id from "nexus-", and unknown fields survive saves. 250 launcher tests. Phase 2 (sign in, one-click installs) needs Nexus to register the app: the email draft is in dist\nexus\registration-email.txt.
  • Live test, 15 Oct 2026 (two real players, the host's and the guest's own recordings): the guest got off the office computer and was instantly outside the map behind the store, stuck on black ground. Cause (bp\MachineActor.txt): leaving any MachineActor_C glides the pawn to "Prefered Location where player get out of Machine" projected with K2_ProjectPointToNavigation (@4286) and to Vec(0,0,100) when that fails; a guest has no navmesh. Same bug the till had in 1.7.0. 18 classes are machines (bpdump grep "class \S+ : MachineActor_C"): ComputerScreen, Catalogue, Return, StaffBook, Calendar_Machine, CandyDispenser_Base, CheatScreen, ConsoleRental_Mat, CottonCandyMachine, CustomMovieStation, PopcornMachine, RepairStation, SidewalkSign, SlushMachine, SnackShelf, StoreExpansion_Machine, StoreLevel_Machine, StoreName. Fixed for all of them (guestinput.lua tickMachines: "In Machine Mode" true -> false, then > 300 cm from the exit spot -> put on it). The host was not dragged when the guest sat at the computer (host recording). Also seen: the guest's computer order pages show the widget's design placeholders (it reads a GameMode) - separate feature, listed as known.
  • Guests had no pause menu at all (Player_Character Input_OpenESC casts GetGameMode first). pausemenu.lua: Escape on a guest opens UI_ESC-Menu_C itself; every pause menu gets "Back to the store" (a UX_Button_ESC_C above "Return to menu", click via UX_Button_Base's bound event @312). Guest: collision/walking back on, action 12 -> host Spawn.placeGuest beside the host. Host: by the register's exit. rt.unstuck, Config.PauseMenuBackToStore (missing = on). Smoke 388 PASS (15 new). Live, two copies on one PC (user's OK; their open game closed first, saves byte-identical afterwards, GameUserSettings.ini unchanged, bridge/trace files removed, new scripts left deployed): guest sat at the office computer, real W -> its glide went to (0, 0, 178), the mod put it at the exit (2377, -1738), host saw the same spot. Real Escape on the guest opened the pause menu with BACK TO THE STORE above RETURN TO MENU; a second Escape closed it. Guest stranded at the origin + a real click on the button -> host log "(0, 0, 92) -> (2495, 41, 132)", beside the host. Host's own menu has the button too (above the last-save line); clicking it while paused put the host at the register's exit (1718, -1486) and unpaused.

Network conditions, three players, a host crash, and the office computer for guests (2026-10-15)

  • Tools: dev/netsim.py (UDP proxy, lag/jitter/loss, in order unless --reorder), dev/worldcheck.py (host vs each guest: every tape/box/shelf/decoration/poster/standee paired by position, money, level, XP, day, clock, season, weather; --wait), coop_test.py up <address> / `guest-only ` (a third copy takes slot C). Rented-out tapes are hidden at (0,0,0) on the host and never replicate (83 of 479): worldcheck skips hidden actors.
  • 150 ms / 25 ms jitter / 2%: everything matched after the join; a guest pickup and a 7.4 m walk (AddMovementInput from a ReceiveTick hook, real client moves) converged on both. Rejoin and late join (host changed money and clock while the guest was away): store complete 17 s direct, 19 s through 300 ms / +-80 ms / 5%, 18 s again on the same proxy. Three players through the bad link: both guests match. With per-packet jitter that reorders (--reorder) it took ~3 minutes: Unreal drops a packet older than the newest. netsim's first version went deaf after a guest left (Windows proactor loop stops reading a UDP socket after one ICMP port-unreachable); it now runs on the selector loop. Guest bandwidth: 100,000 B/s, server tick 60.
  • Found and fixed: a guest who left holding a tape left it floating at hand height (now dropped the game's way: Toggle Currently in Hand false + Toggle Simulate Physic true). Dropped and thrown things then lay on the host's floor and hung in the air on guests: physics runs on the Mesh, which leaves its parent; the replicated root stays (read off a dropped tape: simulating, attach parent none). thrown.lua keeps the root on the mesh; resting ones checked every 5 s (six resting tapes looked up 5x/s were 27 of 50 path lookups a second). "No other player's pawn ... a replication problem" fired on every join and leave: now only after 20 s with a PlayerState and no pawn.
  • HOST CRASH, twice, three players, idle, 6.5 and 3.5 minutes in, both after a guest had left: EXCEPTION_ACCESS_VIOLATION in UE4SS, the same native stack, trace phases NpcSync.tickHost and World.tickHost. Offsets against UE4SS.dll's export table: UObject::GetOutermost <- UObjectGlobals::FindObject <- ForEachUObject - a walk of every object. UE4SS's StaticFindObject is that walk (measured 2-29 ms a call; only native classes are fast) and both phases call it on timers (speech hooks every 5 s, the sky's material collection). ue.lua now replaces the global with UE.staticFind (soft-path route, 0.1 ms, same answers on ten kinds of path). Also: class learning walks for config names this game lacks (AI_Character, Staff, Customer, Snack_Pack_4x, ConsoleRental_Console-Base: 8+ a minute, forever) give up after 6 tries per world; component classes (Shelve_Container_C) walked at most once a minute. Fine-grained dev trace added (UE.trace -> devbridge "phase > step"). After the fix: 43 minutes with three players through 300 ms / 5% loss, 21 guest departures and rejoins (every 2 min, alternating), no crash and no error in the log (before it: 6.5 and 3.5 minutes). StaticFindObject now ~2 calls/s, the native ones cached; walks 2 a minute (component classes). Left over, older: ~13 per-tick class lookups a second through the soft-path route, each leaking ~0.6 KB in UE4SS.
  • Cleanup: 12 saves byte-identical to the pre-test backup, GameUserSettings.ini identical to .pre-claude, bridge/trace files removed, the released 1.7.3 scripts put back in the game folder (1.7.4 is built in dist/ but not committed or published).
  • Office computer for guests (computer.lua, guest action 13). Research note: side-session-notes/.../office-computer-pages.md. Host payload: Market_C arrays/TMaps read in place, price via Movie_C "Generate Value in Pennies of Movie" (1800 for Skull on Mars), old flag, new-release days left, net worth (screen's "Return Net Worth of Store"), date ("Day and Season" out of Date_C "Format Date in RR"). Guest: cards made with WidgetBlueprintLibrary Create, fields written in place, AddChild runs Construct (box art, name, tags, rating drawn by the game); price/Old overwritten; bundle title from the string table with the count; new-release body run from @2022; its standee picture is a class neither copy can load from Lua - hidden. Orders: the host calls "Market - Click on Button Product Event" / "...Bundle Movie Event" / "New Release Movie - Product Click on Button Buy Even" with a throwaway card carrying the price. Measured: movie -975 cents, market 19 -> 18, tape 479 -> 480; new release -2500 and 1 tape; bundle -10000 and 10 tapes, 2 -> 1 bundles; a repeat after the window refused. Live with the real mouse on the guest: Market tab, scroll, BUY on The Rocketman Tragedy -> host money -975, tape at the delivery spot on all three screens, guest market refreshed. Bugs found on the way: one card registry for both pages lost the market's cards when the new releases arrived a tick later (Buy did nothing); pages redrawn every 30 s (scroll reset) - now only when their content changed; a guest's clicks on a machine's screen were also forwarded to the host as "primary" - not while at a machine.
  • UE4SS facts measured: all out-parameters land in the first table passed; nested struct wrappers write in place; FDateTime Ticks unreadable; TSoftObjectPtr -> string via Conv_SoftObjectReferenceToString.

Other players' bodies: one look per player, a walk that walks, things in hands (2026-10-15)

  • Reported by the user: "a lot we could be doing better, especially with player pawn animations and their look; some have broken textured shirts". Measured on the host, a guest's body next to a passer-by: SK_CHA_M_01 in the mesh's default material instances (MI_CHA_Cloth_Tshirt_01 ...), bald, staring; the NPC's slots are dynamic materials of the same parents with a texture and tint, hair/beard/glasses on HeadTop_EndSocket. The body's AnimBP_Base_C: Speed 0, Scale Character Play Rate 0 while walking (the NPC beside it: 165, 1.143). Cause: its update casts the pawn to AI_Base_Character_C (@4578) and stops on a player. Each machine also borrowed from whatever NPC it found first, so every screen saw a different person.
  • Fix (bodies.lua, npcsync.lua exports Npc.look = read/dressBody/dressSlot/asset/face/HEAD): host picks a town NPC look per PlayerId (passers-by and customers first, a different NPC each, only a dressed one: its top a dynamic material with a texture), sends "plook" every 10 s; every machine swaps the body by SetSkinnedAssetAndUpdate (also now the first choice in writeMeshAsset), dresses the slots/eyes, sets a neutral face, and adds hair/beard/glasses with AddComponentByClass (StaticMeshComponent, NoCollision) attached like the NPC's own "Attach Child to Skeletal" (@33527). The AnimBP gets Speed, play rate (speed/144) and Is Holding Something (the hand socket has a child) every tick. The held item's socket ("AttachObject Socket", parent "Lag offset") goes on the body's hand_rSocket on machines where the player is remote (the stack's on hand_lSocket); reset to the hand every tick (the host plays a guest's first-person "bring up" timeline on it, @19159). Config.PlayerLooks, Config.CarryInHands (missing = on).
  • Measured live, three copies: 3 looks assigned, each player the same on the other two screens (e.g. guest B a grey-haired man in a striped sweater and jeans on the host and on guest C; the host a moustached man in an orange T-shirt on both guests); B's body Speed 300 / rate 2.08 while walking, 0 / 1.0 standing, a stride in the screenshots; B's tape in its right hand on C's screen; a thrown tape landed in the same spot on all three (worldcheck: 396/396). 0 errors. Smoke 410 PASS (7 new). Cleanup: saves byte-identical, GameUserSettings.ini unchanged, bridge files removed, the released 1.7.3 scripts back in the game folder.

Memory: per-tick class lookups (2026-10-15)

  • Measured per call (host only, 200,000 calls each, private bytes after two Lua collections): control IsValid(obj) 0-1 bytes; the path lookup UE.findObject made (MakeSoftObjectPath + Conv_SoftObjPathToSoftObjRef + Conv_SoftObjectReferenceToObject) 257; Conv_SoftObjectReference- ToObject on a kept reference 86; K2_GetActorLocation 279 (struct returns and params leak in UE4SS 3.0.1); number/object property reads ~0. Rate in a host+guest session: ~31 path lookups/s on the host, ~34 on the guest (findAll/findFirst resolve each Blueprint class once per tick; sendDrawer's CheckOut_C was 9/s alone).
  • Fix: ue.lua keeps one TSoftObjectPtr per path (softRefOf; bounded at 512) and resolves only that - a soft reference is a value, its weak pointer checked against the object array's serial number, safe across frames where an object pointer is not. Measured: kept references stay independent (A, B, then 50 more: A still A, in a later frame). UE.staticFind uses the same. sendDrawer now runs 5 times a second. rt.status shows the kept references. Saves ~(257-86) x 30/s = ~20 MB/hour per copy. Session windows (4 min, Private Bytes slope) were noise: old lookup +193/+313, then -4/+213 MB/h; new -5/-86, then +83/+130 - the game itself swings that much.
  • Test mistake, user-visible: to measure the game without the mod I paused the mod's tick on both copies (wrapping Log.guard("tick")); the guest's NPCs stopped being restyled and the user saw "all the NPCs are broken again, no hair or animations". Unpaused within a minute; after it, 9/9 guest NPCs with hair and a walk rate. Don't pause the tick while anyone is watching.
  • Live with the new ue.lua: worldcheck matches, NPCs 9/9 paired, player looks sent and worn, 0 errors. Smoke 412 PASS (2 new). Cleanup: saves byte-identical, GameUserSettings.ini unchanged, bridge files removed, released 1.7.3 scripts back in the game folder.

Office computer: Black Market search and Data page for guests (2026-10-15)

  • Black Market = the Search page (switcher value 4): type a SKU, buy that movie at a mark-up. On a guest its Committed event (@10716) runs Create Film Data (works: data tables, deterministic per SKU) and stops at Unlockable_Function "Return Research" (GameMode) -> "no movie found". Guest hooks the event, clears the box and asks (action 13, kind 5, SKU, sequence number). Host: WidgetBlueprintLibrary Create of a second UI_ScreenComputer_Widget (never shown, so the host's own Search page keeps what the host typed), its Committed(SKU), Search_WrapBox child 0 is the UI_Computer_Product-SKU card or NoFoundProduct; price from the card's own Generate Value in Pennies (its Construct never ran); "csku" back with the product line, price, genre-researched, Old. Guest makes an SKU card, writes the fields, sets the bool, overwrites the price. Buy = kind 6: the host searches again and calls its real screen's "Search - Clicked on a Product Event" with that card (whole Product incl. tags; class SKU_C, so achievement 17 can unlock on the host - the user said yes to that, "any buy in the store").
  • Data page: Generate Data Movie Genre Button and Generate Data Page of Genre stop at the GameMode on a guest. "Generate A Data"(ID, genre) does not use it. Host sends "cdata": genres researched (+16 Adult), copies owned, times rented. Guest makes UI_Computer_Data-ButtonSwitcher buttons (Genre set before AddChild), hooks their btn click, runs the screen's own "Button Movie Genre Data - Click Event" (lights the button, clears the lists), then 10 rows a tick: owned (in the count map) into VerticalBox_Data-Owned with copies/rented text, the rest into Non-Owned + Set Disable (as @308/@856).
  • Standee picture: the earlier "class neither copy can load" was a bad path, not a missing asset. Read through UE4SS the host's Standeereference is ".../Standees_Collection.Standees_Collection_ 19759_C"; the market builds ".../Standees_Collection_19759.Standees_Collection_19759_C" (Market_C "Return Standee based on SKU") and that package is in the pak - the FName number of the package name was dropped on the way into Lua. Thumbnails pair 1:1 with movie standees (T_Standees_Collection_, 27/27); video-game banners do not (1 thumbnail for 9), so a game's banner picture stays hidden. Guest loads the thumbnail by path string (UE4SS LoadAsset, then UE.load) and sets Image_Standee's brush.
  • Smoke 417 PASS (6 new: search answered once, not found, search buy once at the card's price, Data payload, standee path). NOT yet run in a live game: released as 1.7.4 for the user's own live test (the user's call).

1.7.5: the register lagged for a guest (2026-10-16)

  • Reported in the user's live test (a friend as the guest, over the internet; the guest's own recording): checking customers out was "super laggy". Frame analysis of the clip (ffmpeg mpdecimate, unique frames per second): no freezes over 0.2 s, 14-30 new frames a second; the visible lag was the register's screen - a tape scanned, the guest's screen still on the desktop page, then dark, the bill ~2 s later. Host side: sendBill looked every 0.5 s, sendDrawer every 0.2 s, plus the host's 10 Hz tick for the replay and the guest's for the message. Fix: a guest's action while a guest holds the till opens a 2 s window in which both look every tick (the same tick's tickHost runs after Chan.pump, so the press's own effect goes out at once).
  • rpcchan merged repeats of one action from one guest inside 0.12 s (each press is sent twice). Both copies leave in the same frame; 0.12 s also swallowed real clicks counting change out of the drawer. Now 0.04 s. Test: two clicks 0.1 s apart are two presses.
  • The guest's frame dips (14-20 fps at times in the clip) were not explained: the host is this PC, the guest a friend's; measuring inside the running host would have meant restarting it.
  • Smoke 419 PASS (2 new). Host log for the session: no errors.

1.8.0: characters, every machine, bodies, mornings (2026-10-16)

  • Audit: docs/MULTIPLAYER-AUDIT.md (status by system, measured on the two-copy rig).
  • Return desk: machines.lua seats a guest at the HOST's desk (Toggle Machine as the guest, glide stopped, finishing step run as the guest) and replays clicks and drags from the guest's camera; a drag cursor spawned outside a guest window is the host's and is handed over (Toggle Input off, on again as the guest). mirror.lua hooks the desk's cosmetic functions on the host and plays them on the guest's copy (screen frame, case open/close, tape in the rewinder) and watches the rewinder door/base and the case's tape. Verified: a guest rewinds a tape end to end.
  • Carried items: each socket offset by minus its first child's relative location -> the item on the hand on every remote screen.
  • Day/night on guests: the time-of-day listeners bind through the GameMode; each is told the host's DayTime once per change (94 told). TVs: the guest's movie players never start; started and each TV wired to a player on its channel (8/8 pictures).
  • Bought machines: on a guest the popcorn machine had Unlocked/Bought false (save state, loaded only by the host's GameMode), so Interaction stopped @49. worldsync sends which machines the host has bought; the guest runs "Make Machine Usable For Player"(true) (popcorn, slushie, sidewalk sign differed; also covers cotton candy, console rental, custom movie station, repair station).
  • Concessions: replaying them on the host goes half way (the kettle pour works) but the shovel follows player 0's mouse/camera every tick and the full cup goes to player 0 a tick later, so the guest plays its own copy and the host makes results real (concessions.lua): popcorn level both ways (action 17), a product the guest's game put in its hands is put away and made for real by the host in the guest's hands (action 18: kind x 60 + fill x 50 in yaw, flavour in pitch - a rotator crosses as three 0..360 angles), fill and flavour shown on the guest's copies (they do not replicate: the real cup read 0 on the guest). Verified live: popcorn scoop/drop/full cup, slushie hand-off with flavour. Cotton candy not bought in the test save.
  • "The pawns move in circles" (the user's recording): a remote body faced the view 1:1, and standing still its hips traced a 30 cm loop every 1.6 s. AnimBP_Base adds Blend1D_Rotate (the town's turn-on-the-spot steps) at full weight when standing (ApplyMeshSpaceAdditive, alpha 1 - Speed/15), picked by "Rotation Delta Z Normalize" - an NPC's update makes it (still to turn
    • rate/2)/rate = 0.5 standing still; on a player nothing writes it, 0 = turning for ever. Measured hips 21x16 cm at 0, 17x12 at 1, 0x1 at 0.5. Now 0.5, and the NPC formula while the body turns on the spot. Bodies also turn like people (head look-at node from "Look At Pos"/"Look at Player Alpha", body turns past 60 degrees or walking), per frame from the pawns' ReceiveTick - on the host the guest pawn's tick is stopped (guestinput.lua), so the host's own pawn's tick turns the guests' bodies, found through the controllers. Engine rotation smoothing set to 0 on remote pawns. Verified on video: idle body still (hips 4x4 cm over 8 s).
  • Mornings: "Create a new day" teleports only player 0. dayend.lua waits until the host is at the Actor Gatherer's "PlayerStart Game Zone" and puts every guest beside them, facing the same way (ClientSetRotation). Verified: guest moved from the store to the start, Mon 8:00.
  • Characters (requested: "show all possible characters and let people pick one", "make it like the developers made it themselves"): the character screen (charscreen.lua) is a second UI_ESC-Menu_C (its Construct sets the input mode, the cursor and the menu's own mapping; the host's pause is undone at once) with its blur hidden and its column replaced by a dark panel in the menu's own font and UX_Button_ESC_C buttons; a SkeletalMeshActor figure (not replicated) dressed by bodies.lua, playing Anim_CHA_Idle_1. creator.lua: the 14 named regulars (AI_Data_Membership-Unique), every body's own clothes plus the regulars' textures, 26/17 hair meshes by gender, 9 hair colours, beards, glasses, 13 tints, heights 0.85-1.10; wardrobe.lua is generated from the dumps (dev/make_wardrobe.py). Opens by itself the first time (10 s after a pawn, not at a machine/till/talk/summary), from the pause menu's "Character" and F4. A guest's look crosses as FNames over ServerNotifyLoadedWorld ("RTLK:i/n:piece", 800 chars each).
    • A shipping build forgets a SkeletalMeshComponent's single-node animation when its mesh is swapped (AnimationData is editor-only): the figure went to its bind pose on every body change. PlayAnimation again after each new body.
    • The figure's place: traces (LineTraceSingle; its FHitResult lands in its own out table: Distance, ImpactPoint...) at knee, chest and eye height toward the spot, 12 directions, plus other players' pawns (they do not stop a trace), then a floor trace under it.
    • Height (cs) now goes on every player's body (mesh RelativeScale3D, feet stay on the floor) and into the look's signature; the walk's play rate divides by it, as the town's does.
  • THE HUD STAND-IN (critical): asGuest nilled the GameMode's "UI Game Interface" so a guest's panels stayed off the host's screen - and Change Money, with no widget, writes the total and returns WITHOUT setting "Have enough money" (@430 -> @810). The snack shelf charged, spawned the snack at the box, failed the check and destroyed it (pending-kill actors at (0,0,0) in FindAllOf). A manual replay with the widget left in place stocked 3/3. Now asGuest swaps in a second UI_Player_Interface_C made with WidgetBlueprintLibrary:Create and never constructed or shown (its money/XP pop-ups cleared after each call). Snack adds 3/3, the till sale (scan x3, cash, $4.50 change, Space) and catalogue orders all verified after it.
  • Screen machines (uimachines.lua, from a full read of all ten in the dumps): catalogue cards uncovered by the mirrored level (class CDO "Unlock Level" / "Product Unlock Level"), station cards via their own Generate Text with the machine's own TSoftObjectPtr property value handed straight back (works; a Lua string there is a native crash), purchases = action 21 run by the host widget's own handler with a nil card (never read); array out-parameters come back as a Lua list of RemoteUnrealParams (:get()). Expansion map linked per room from the host's sectors' refs; sidewalk sign ad; level board cross-mark/post-its; store name via an "RTSN:" FName (hex: an FName keeps its first casing), plus the sign's font now mirrored. Clicks on screen machines are no longer replayed on the host (the repair desk, sidewalk sign and store name sign never ask "Can Player use the machine?"; their glide would run on the host).
  • Phone: its ring sound, light, shake, interact icon and handset mirrored (mirror.lua LIST).
  • Bridge probes, two ways to crash the game outright (pcall cannot catch either): GetFName() on an invalid wrapper (a property read that came back empty: reads 0x18, UObject's name), and member names typed only into the bridge (UE4SS keeps a pointer into the Lua string; use UE.TryGet/TryCall, which pin the name).
  • Smoke 487 PASS.

A character per store (2026-10-16)

  • Requested: "i want it to save the character i make per save". 1.8.0 kept one look for every game (top of RewindTogether.character.txt). Now each store has its own section in that file, [store <key>], key = FNV-1a of "<host's Steam ID or name>/" as 8 hex digits, so the same store has the same key on every player's machine and nothing about the host is written in the file. The host reads its store from the GameMode's "Save Slot Name" (Core_Gamemode @13660, set from GameInstance "Save Slot", e.g. Player_Save1) every 3 s and broadcasts it to guests as "chr" { k, l }; a guest enters the store the host names.
  • First time in a store: the last look made is worn while the screen opens by itself; DONE keeps a new one for this store, CANCEL keeps the one worn. A 1.8.0 file's look is the "last" one, so the first store after the update asks once, already on it. A choice made before the store is known goes to the first store without its own.
  • The pause menu's Character and F4 were already there (1.8.0); they change this store's.

The repair desk and the console rental for guests (2026-10-16)

  • Repair desk (repairs.lua): a tape's Error is not replicated (Cartridge_Base: only "Product Structure" is), so a guest's desk always took the "not a broken cartridge" path (RepairStation Interaction @695/@1340). The minigame is the player's own (zones hit-tested from the local cursor, tools following the local mouse, a client-spawned videotape_Repair-Detail), so the guest's desk plays it: the host sends each guest the Error of the tape in its hand ("rph" { e, p = product name }), the guest writes it on its copy and puts the tape in its pawn's "Object Hold" (the desk reads it; hud.lua only does that while the native HUD stand-in is armed). The guest hooks RepairStation_C:Repair cartridge End and sends action 24 (Pitch = seq); the host mends the real tape, still in that guest's hand there: Set Error(desk, 1), Toggle Broken Sticker(false), Change XP(2) (Pickup The Container @434), inside asGuest. Steam stat 16 is gated by a GameMode: not counted for guests. Leaving mid-repair: the guest's desk keeps the tape in its own slot (in single player it waits there), so the guest's game runs the desk's own Destroy The Video Tape Repair Detail + Pickup The Container to put it back in the hand, as the host has it. A fixed tape ignores a stale "broken" from the host for 10 s.
  • Console rental: the mat and the station are worked by world clicks the host replays. The guest's copies get the host's state ("crs" { s, f, p, b, l, c1..cn }): Status For Monitor(step, true, frame) (Force: @5/@207), the monitor widget's Toggle Progression Bar / Set Progression, Set A Container Light per slot, and each console's State (0 if the host's is 0, else 1, which the guest's own slot and mat refuse) and "is Broken".
  • mirror.lua plays the broken stickers (tapes and consoles), the console's scan sticker and door, and the mat's drawers and test cable on the guest's copies.
  • Live (two copies, 16 Oct 2026): a broken tape's sticker reached the guest (mirror.lua); the guest held it, its copy was marked broken and its desk took it; the whole minigame with real clicks on the guest's window (open, screwdriver, three screws, scissors, two cuts); the host mended the real tape in the guest's hand (Error 1, XP 7112 -> 7114); leaving mid-repair handed it back. Console rental at level 21's machines made usable by bridge: a returned VHS player (State 2) on the mat from the guest, the 40 s scan on the guest's monitor (the test picture and its progress bar, then CHECKED IN), taken back and shelved, with the shelf light on the guest.
  • Found live, fixed: (1) a guest carrying something who clicked a screen machine lost it on the host - with something in hand the host's look-at is not the machine, so the replayed click put the tape down; the guest now holds back a click whose own look-at is a screen machine. (2) The guest's console copy runs its own Verify 20 s after it appears, lifts itself and turns physics on, and its mesh stays behind; repairs.lua turns it off and puts the copy where the host's is. (3) With the host's State 0, the guest's own shelf took the console at the same click as the host's; the guest's copies now always read State 1.
  • Rig notes: the user's display settings are borderless full screen, and each copy snaps back to it on load and on joining; coop_test.py now runs r.SetRes 1254x705w and the layout after both. A one-call "aim, then click" sends the old aim: the camera manager's rotation lags SetControlRotation by a frame, so aim in one bridge call and click in the next. And never call a method on an unchecked wrapper in a probe: GetAttachParent():GetFName() crashed both copies (EXCEPTION_ACCESS_VIOLATION reading 0x18 in UE4SS).

A customer's console rental with a guest (2026-10-16, 1.9.1)

Working tree only: native.lua, talksync.lua, repairs.lua, replication.lua, test/smoke.lua. 540 checks pass. The live test was stopped before the end (the user's call).

  • VERIFIED LIVE, fixed: a guest's right click on a customer asking to rent a console opened the talk on the HOST's screen. BTTask_ConsoleRent-Asking notes player 0 (the host) when the customer reaches the desk (Move to Help Desk Completed @7529-@7634) and gives that pawn to the talk ('Player', Enter in Talk System @9626), whoever clicks. Native.asPlayer now records whose window is open (Native.playingAs()), and talksync adopts a talk that begins inside a guest's window: the talk's "Player" and the task's "As Player Character" become that guest's pawn. Live: "Can I rent a VHS Player here?" opened on the guest, nothing on the host; the guest's "Sure!" reached the host and the request zone appeared with ref Player Character = the guest's pawn (the answer runs in the guest's window). The answer was pressed with the button's own Button Clicked Event, not a mouse click (the user was at the PC). BTTask_Checkout-AskingForHelp (movie and other requests) works the same way and is covered by the same code - not tried live.
  • VERIFIED LIVE, already worked: paying. The guest's left click on the customer's money (Money_NPC-Stash_C, replicated by npcsync) was replayed on the host: Player click on Money → Change Money, Make Client to pickup case; the customer took the case and left; the console became State 2, rented on day 28.
  • Hand-over, read off the dumps, done by hand on the host only: Request-QuestZone_ConsoleRental Do I accept(true) → Give The Console To Client (@3042) → Get the Object from Player (@1697: takes ref Player Character's Object Hold, Throw, a tick later Object Accepted). Called on the host with the console in the host's hand, the console went into the case and the customer held out money. The base Interaction refuses once the zone's "ref Object Hold" is set - an earlier probe had set it, which is why the first real clicks did nothing. A left click with something in hand is Trace from Camera for Physic → Trace from Camera for Interaction (box trace; the first hit's Interaction; the case's own answers true and does nothing). NOT TESTED: a guest's replayed click handing it over (zone lending from 1.9.0 is in place for it).
  • The counter's case: the host now has the engine send the help desk's own case (NeedHelp "Case Slot", filled by Create a case) and sends its lid angle with "crs". Seen live: the case appeared on the guest at the host's place. NOT VERIFIED: the lid on the guest.
  • REGRESSION made and fixed during the test: the first version sent every ConsoleRent-Case_C, but every customer carries a hidden one of its own (AI_Client_Character "ConsoleRent-Case", a child actor built on every machine). The guest got a second case floating at every customer's side (52 cases for 26 customers) - the user saw them. A GetParentActor check still let 56 through: a child actor is not linked to its parent when it first appears. Now only the "Case Slot" is sent, and Replication.patchActor refuses child actors. On the running copies the 26 already sent were switched off on the host and hidden on the guest (a guest keeps what it was sent until it rejoins).
  • NOT TESTED: the guest's click hand-over; the lid; the customer's own case and bag shown on the guest when they leave (Toggle Visibility of Case, Toggle Bag: not mirrored); a console coming back (BTTask_Bring-ConsoleRent: a case on the counter, the player takes the console out of it. To trigger: put a rented console in RentSystem "Console Base out Ready to Return" and raise "Weight Chance of Returning a Console").
  • Rig notes. A request zone gives up after 120 s ("Time before Timeout Quest"; set it on the class default for a test). The save's clock sat at 20:00 with its minute timer stopped and customers stopped coming; Hour 15 plus Weather Actor "Start Timer - Add one Minute" restarted it (an in-game hour is 30 s). A console shelved with "Store Object From Game Code And No Animation" also needs its "Container where Pickup Actor is Stored" set: without it, it ended up under the floor at the delivery spot, and customers had no console to ask for and gave up at the desk. Two bridge calls to one slot at the same time can lose one. Most of the hour went on this setup, not the mod.
  • Second run (same day, scene set up first: zone timeout 900 s on the class default, the clock slowed with "Time for one minute" 30 and restarted, the console shelved with its container set). VERIFIED LIVE: the talk on the guest (Nathan); "Sure!" (button event); the console in the guest's hands on the host (put there by the host's PickUp in the guest's window - see below); THE GUEST'S LEFT CLICK at the case handed it over (the guest's hand empty on the host, the zone took it, the console in the case, the customer held out money); the case's lid on the guest followed the host's (open, then shut on the console: 3 lid updates); the guest's click on the money paid ($35 and 5 XP; the console State 2, rented). A console coming back: made due by hand (its rent day set back, RentSystem "Get Console Rent ready for return", "Weight Chance of Returning a Console" 1.0); a customer brought it, and the open case with the console in it was on the counter on both screens. NOT DONE: the guest taking it out of that case (the user's own play in the windows took over).
  • Found, not fixed: the guest's own aim at a shelved console hits the ConsoleRental_Station_C, not the console, so its right click took nothing (the console had been shelved by game code, which may not be the same as a player shelving it).

Other players' bodies: crouching, jumping, a stack of tapes (2026-10-16, 1.9.1)

Reported by the user while watching the two copies.

  • "No crouch animation, I just sink in the floor": a crouch shortens the capsule (bIsCrouched replicates; OnStartCrouch moves the mesh and BaseTranslationOffset by the change), and holdPlacement put the standing offset back on its next 2 s poll - feet under the floor, and after standing up the body in the air until the poll. The offset now follows the capsule every tick (wantedOffset: dress()'s Z plus how much shorter the capsule is). MEASURED: crouched, half height 20, base Z -20; standing again, -90.
  • No crouch or jump pose: AnimBP_Base has neither. ANIM_CHA_CrouchToStand_01 (2.07 s, S_Mixamo_Base) kneels at 0 s and has bent knees at 0.9 s (looked at on a body): held single-node while crouched / in the air (MovementMode 3), then back to the blueprint. Live: kneeling on the counter the guest was put on; the blueprint back after standing. SetAnimationMode wants both parameters; SetAnimInstanceClass fails where SetAnimClass works. The jump pose first came only as the body landed (the host hears the movement mode late, plus a 0.15 s wait): the wait is gone - not re-checked live after that change.
  • "Carrying more than one movie z-fights": the stack socket's tapes (LA_VHS_Box_01, 3.5 cm thick along its Y) sat 2.6 cm apart along that axis. Each extra tape's Mesh (a child of its root, not replicated) is pushed out along the tape's thin axis to 3.6 cm from the one before, on other players' bodies only, and put back when it leaves the stack. Not checked live yet.

1.9.2: the mod's own stutter (2026-10-16)

  • Reported from a real two-PC session on 1.9.1 (host here, a friend over the internet): "lots of desync", "can't carry the movies", "really choppy". The host's log had no errors.
  • Measured on two copies, the guest through dev/netsim.py (150 ms, +/-25 ms, 2% loss):
    • Not bandwidth: the guest's connection runs at CurrentNetSpeed 100000 on both sides (driver MaxClientRate/MaxInternetClientRate 100000; UPlayer ConfiguredInternetSpeed 100000), about 30 packets/s to the guest; 747 open actor channels.
    • The host's walk on the guest: steady 2-9 cm per frame with a light recorder. A heavy one (FindAllOf every frame) made both copies 20 fps and produced the jerks it was measuring.
    • Frame times (LoopInGameThreadWithDelay(1) logging os.clock deltas): every 6th frame +10 ms (the 10 Hz tick), and 50-150 ms frames every couple of seconds.
    • Per-phase timing by swapping DevBridge.phase for a timer (time between phase calls): ~15 ms of mod work per host tick; the worst single runs were the replication sweep (87 ms, once 3-5 s), the calendar (28), the screen machines (23), the skin sync (23-58).
  • Fixed:
    • Replication.sweep: Shelve_Container and BoxStorage_Container are StaticMeshComponents (header dump), not actors - listing them cost a walk of every object (155 ms) and nothing applies to them: skipped (UE.classIsNotActor). Classes a store lacks: ue.lua's learning walk retried 6 times per name, and the bare-name fallback walked for names the game does not have ("AI_Character", "Customer"); now a known path is the lookup (walked once as a backstop) and a missing bare name is asked once. One class per call every 0.5 s, at most 120 actors per call (the list is fetched again each call); valid() for the reads, alive() only before a write; the child-actor question asked of a whole class once.
    • Calendar, staff book, screen machines: sendState rebuilt its state on every tick once SEND_EVERY had passed whenever nothing had changed or no guest was there to send to (the broadcast never "succeeded", so lastAt never moved). A lastBuild timer; SEND_EVERY 10/11/13.
    • Skin sync: AC_MeshSkin (and _Shelve) Load Material / Load Asset / Manual Load Asset are hooked; the scan runs after one fires, and in full every 30 s and for a new guest.
    • repairs.lua's lid check skips the case search with no case on the counter.
  • Result (30 s of frames, host / guest): no multi-second stalls; the worst job about 25-35 ms (was 87-155); worst frames ~90-120 ms (two copies on one PC, ~30 fps each).
  • Not done: the guest applying a full skin list (to 62 ms), NpcSync (32), Bodies (30), the guest's movement check (46), Chan.pump with big messages (40); Lua GC ~17 ms per full cycle.
  • "Can't carry the movies": not reproduced. Under lag the guest's pickup from a shelf landed in its hand on both machines. (A second right click on the same, now empty slot puts the tape back: the game's own rule, not a bug.)

The staff book and the calendar's rewards for guests (2026-10-16)

  • Staff book (staffbook.lua): on a guest both pages were empty (Generate Container Page stops at its GameMode guards, @145 and @826). The host sends who is in the book ("stf"); the guest's game draws its own cards. The FStaff's plain ints (pay, the two skills, job) and its photo are written into the card's own Staff before AddChild, so the card's Construct writes "$40.00 / DAY", the skill words and their colours itself. The name, about-me, favourite movie and the trait lines (UI_StaffBook_Txt_GameplayTraits_C with its "Gameplay traits" byte) go on after. Measured: the FStaff's TSet of traits reads with set:ForEach(function(elem) ... elem:get() end) in this UE4SS build, and struct ints written in place in a widget's struct property hold.
  • Hire, fire and job are action 22: what, the card's place in its list, and its pay to check it by (if the list moved, the only card with that pay). The host makes an unseen UI_StaffBook_Widget_C (WidgetBlueprintLibrary Create; given "As Core Gamemode" and "Staff Book Ref", the two things its Construct and the book's BeginPlay would give it), has it Generate Container Page from the GameMode, and calls that card's own button event. So the vanilla handlers run: Hire Staff- Click button Event (staff places from the research done, the employee spawned, the applicant gone, Steam stat 8 and achievements 8/9 on the host), Event Fired Staff (AI Director Remove Staff), and the card's Set Staff Job. No FStaff is ever built from Lua. Run inside asGuest, so a refusal's Show Error goes to the stand-in HUD. The guest shows the game's own words, from its own string table (ErrorPlayer_StaffBook_*, "{level}" filled in), chosen by the unseen screen's "Maximum Staff Hired".
  • Calendar (calendar.lua): the guest's machine never initialises (Set init Calendar waits for a GameMode, @30). Giving the guest's machine its own WeatherSystem ("Weather Actor ref", the widget's "Calendar Ref" / "ref Weather Actor") and "Save Load" = true, then calling Achievement - Check which is unlock and the widget's own Init Calendar, draws the task list from the guest's own Steam, the weather and the check marks from the mirrored day. Init Calendar's GameMode reads just fail (Accessed None) and carry on. The month is the host's: each grid child's events and slot row/column ("cal" ev "r,c,e.e;..."), drawn with UI_Calendar_Sector_C / _EmptySlot_C. Assigning a Lua table to the sector's "Calender Event" TArray before AddChildToUniformGrid works: the icons and the bonus tooltips are right. Each row's Reclame is the store's, set again whenever the guest's game rebuilds the list. A Claim is action 23: the host clicks its own row's Button_Claim event (Map_Add Achievement Reclame, Change Money +10000, Change XP 50).
  • Live (two copies, 16 Oct 2026): the guest's applicants and employees pages matched the host's (Gabriel, Layla; Adam at the register); the RETURN STATION radio changed the host's job (2 -> 1) and the guest's radios followed; a hire with the one place taken showed "You must be level 20 or higher to hire your second employee."; FIRE let Adam go (he walked out), HIRE took Gabriel (he walked in on both screens, the guest's list redrawn without him). The guest's calendar page was identical to the host's (28 days, events, claimed rows), and a claim paid the store $100 and 50 XP once (money and XP on the host, CLAIMED on both lists). Not yet seen live: a new day's redraw on a guest.
  • Measuring tip: gw.py shot takes about three seconds, longer than a three-second notice, so read the HUD's UI_Error (txt_error, Current Time / Animation Time) through the bridge instead of screenshotting it.