Skip to content

PHASE1.md

Latest commit

 

History

History
86 lines (71 loc) · 5.63 KB

PHASE1.md

File metadata and controls

86 lines (71 loc) · 5.63 KB
Raw

ArkWeb Phase 1: link

Goal: both games handshake over shared memory. The coordinate mapping is unit-tested. Each side can be developed and tested with only its own game running.

What exists

Path What
protocol/arkweb_protocol.h Shared-memory layout Local\ArkWeb_v1: header and heartbeats, HostState, PadState (XInput-shaped), GuestState, SPSC rings
src/common/link.h Mapping, seqlock read/write, SPSC byte ring
src/common/coords.h AK (UE3, Z-up, LH, cm) <-> Spider-Man (Y-up, m)
src/common/util.h Logging (ArkWeb/logs), safe reads, inline / IAT / vtable-capture hooks
src/ak_host/ dinput8.dll for Arkham Knight. Hooks UObject::ProcessEvent and, on the player controller's PlayerTick, publishes Batman and drives him from the guest's hero (SetPhysics(None), RPawn.SetLocationIgnoringCollision, SetRotation). Lets go with PHYS_Falling when the guest disappears.
src/sm_guest/ winmm.dll for Spider-Man. Captures Hero::HeroLocal through its vtable and publishes the hero transform every ~4 ms. Plays the link's virtual pad as XInput controller 0 (through GetProcAddress). Fakes focus while a host is alive (GetForegroundWindow/GetFocus, swallowed deactivation messages, ClipCursor/SetCursorPos ignored while really in the background).
tests/link_test.cpp Coordinate round trips and handedness, seqlock under contention, ring integrity, --layout JSON
tools/arkweb_proto.py Python mirror; --check verifies it against the C++ layout
tools/fake_spidey.py Guest stand-in: replays the recorded swing, anchored at Batman
tools/fake_arkham.py Host stand-in: heartbeat plus scripted pad (monitor, handedness, swing), records the guest
build.bat, install.bat, uninstall.bat Build (low priority, one core), copy into both games, remove

arkweb.ini next to either DLL can override [ArkWeb] LogDir=, DrivePuppet= (AK, default 1), TickEvent= (AK, default PlayerTick) and ForceFocus= (SM, default 0).

Offline results (2026-10-01)

  • link_test: all passed. Seqlock: 0 torn reads under a hot writer. Ring: 200k random records through a 4 KB ring, 0 bad.
  • arkweb_proto.py --check: layout OK.
  • fake_arkham <-> fake_spidey end to end: linked, 59 Hz guest publish, the host recorded 698 samples in 12 s.

In-game tests (one game at a time)

  1. AK + fake_spidey. Batman should be carried along the recorded swing path through Gotham, then dropped when the replay ends. Check logs/ak_host.log (resolve line, ProcessEvent census, "driving Batman") and logs/fake_spidey.log (Batman's distance from the hero should stay small).
  2. SM + fake_arkham.
    • monitor: the hero is published and focus is faked, so the game keeps running when you alt-tab.
    • handedness: settles coords::kGuestRightHanded.
    • swing: the virtual pad makes Spider-Man run and swing.

In-game results (2026-10-01)

Test 1: Arkham Knight + fake_spidey (pass)

  • The ProcessEvent hook resolved everything. PlayerTick ticks the host on the game thread.
  • Batman was carried along the full 235 s replay, 2.2 m from the hero on average (3.8 m max), and the camera followed.
  • Expected artifacts:
    • falling pose (physics is off; Batman gets hidden in Phase 4 and replaced in Phase 5)
    • clipping through buildings (the path is a New York path; Phase 2 makes Spider-Man collide with Gotham)
  • Fixed afterwards: a torn read of the guest state counted as "guest gone", which caused 6 brief drop-and-regrab blips. The host now keeps the last good state until the heartbeat times out.

Test 2: Spider-Man + fake_arkham (pass, after four fixes)

  1. Windows.Gaming.Input. The game prefers WGI, and once WGI found the user's Xbox controller it logged "XInput disabled". The guest now IAT-hooks the registry reads (RegQueryValueExW/A, RegGetValueA) and reports EnableWindowsGamingInput = 0 for that run. The user's registry is never written (VirtualPad=1, the default).
  2. Keyboard path. XInput is only polled for slots with a real Xbox device, so the virtual pad also drives the keyboard bindings, read from the game's own settings.
    • The guest posts WM_INPUT with tagged handles and answers its own GetRawInputData (IAT hook).
  3. Stale window. The game destroys its first window and creates a new one. The guest now finds the window by class GameNxApp and re-finds it whenever the handle dies.
  4. App-active flag. The window procedure only reads raw input while NxGameWindowImpl::IsActive() holds: flags +0x18 and +0x19 set, +0x1A and +0x1B clear, on the object at [[exe+7b11770]+0xa0]. +0x19 drops when focus is lost.
    • While focus is faked, the guest holds +0x18 and +0x19 at 1 and also posts activation messages.
    • This was validated by writing the flag from outside first; it is now built into the DLL but not yet re-tested from the DLL.

Results

  • Handedness: three forward/right stick runs gave (forward × up) · right = +1.00, +1.00, +0.99. Spider-Man is right-handed, so kGuestRightHanded = true is confirmed.
  • Swing by virtual input: run, jump and swing from a rooftop covered 428 m in 20 s, between street level and 106 m up.
  • Once active, both input paths work (XInput virtual pad served, injected keys read).
  • Velocity spikes fixed: the guest samples every 4 ms but the hero moves once per frame. Velocity is now computed only across real position changes.

Phase 1 is complete. Still open from it: a quick re-run of fake_arkham.py handedness once the rebuilt guest (built-in active-flag hold) is installed, to confirm that no outside helper is needed.