Skip to content

AGENTS.md

Latest commit

 

History

History
94 lines (66 loc) · 4.79 KB

AGENTS.md

File metadata and controls

94 lines (66 loc) · 4.79 KB
Raw

AGENTS.md

Project

Android FPS game AI aim-assist. Single-module Gradle app (:app) — Kotlin/Java + C++ NDK. Captures screen via MediaProjection, runs YOLOv8/v11 inference via TFLite with Qualcomm QNN HTP delegate (Hexagon NPU), draws overlays, injects touch via Shizuku + uinput.

Package: team.maodie.aimbot | minSdk=31 | targetSdk=35 | compileSdk=36

Build

./gradlew assembleDebug          # debug APK
./gradlew assembleRelease        # release APK (needs app/release.jks)
./gradlew installDebug           # build + install to device
./gradlew clean assembleDebug    # full rebuild

Release APK: app/build/outputs/apk/release/app-release.apk

No CI, no lint/typecheck scripts, no pre-commit hooks. Build errors are the only check.

Architecture

Entry: MainActivity.kt → permissions + model loading → starts FloatService (foreground service). FloatService owns all overlay views, inference loop, and touch injection.

MainActivity.kt          — entry, permissions, Shizuku auth
FloatService.kt          — foreground service, inference loop, PID auto-aim, trigger bot
├── FloatBallView.kt     — draggable FAB toggle
├── OverlayCanvasView.kt — detection boxes overlay
├── GuiPanelView.kt      — MD3 side panel (自瞄/扳机/模型/系统 tabs)
├── AreaSettingsView.kt  — zone config UI
└── JniCallBack.kt       — JNI bridge → libaimbot.so
        └── aimbot.cpp   — TFLite + QNN HTP inference

RemoteInjectorService.java — Shizuku UserService (separate process)
└── uinput_inject.cpp      — uinput touch injection + EVIOCGRAB reader

root_daemon.cpp            — standalone binary for su-based injection (stdin/stdout protocol)

Critical Details

Native Code (CMake at app/src/main/cpp/CMakeLists.txt)

  • Three native targets: aimbot (inference), uinput_inject (Shizuku path), root_daemon (su path)
  • Prebuilt .so files in app/src/main/cpp/lib/ — QNN SDK libs, TFLite JNI, ONNX Runtime, SNPE stubs
  • root_daemon is compiled as libroot_daemon.so (executable disguised as .so for Gradle packaging)
  • aimbot.cpp JNI methods: init(), detect(), setConfidence(), getBackend(), release()
  • Input format auto-detected: NCHW [1,3,H,W] or NHWC [1,H,W,3]
  • Output format: [1, 4+num_classes, num_outputs] — YOLOv8 transpose layout
  • QNN cache dir: /data/data/team.maodie.aimbot/cache/qnn

Touch Injection (Two Paths)

  1. Shizuku path: ShizukuInjectorClient.java → AIDL → RemoteInjectorService.java → native uinput_inject.cpp
  2. Root path: RootInjectorClient.kt → su → root_daemon (stdin/stdout text protocol)

Both use uinput with virtual multitouch device "AimbotTouch". Touch device ABS range auto-detected via getevent -p.

Coordinate System

  • Screen coords (landscape) → device ABS coords: dev_x = (screen_h - y) * max_x / screen_h, dev_y = x * max_y / screen_w
  • Default device: OnePlus Pad Pro — screen 3000x2120, ABS X=[0,21199] Y=[0,29999]

Models

Config: app/src/main/assets/models.json. Models copied to internal storage on first launch. Supported: YOLOv8n, YOLOv11s, YOLOv26n/s. Single-class (head only) and multi-class (body/head/friendly/etc).

AIDL

app/src/main/aidl/team/maodie/aimbot/IRemoteInjector.aidl — touch injection IPC interface.

Conventions

  • No test suite beyond placeholder ExampleUnitTest.kt. Don't expect tests to validate changes.
  • CLAUDE.md exists with detailed architecture docs — read it for deeper context on PID tuning, NMS, area settings.
  • Signing config is in app/build.gradle.kts (passwords in plaintext — release.jks in app/).
  • ViewBinding enabled, no Compose.
  • Mixed Kotlin/Java codebase — Java files: ShizukuInjectorClient.java, RemoteInjectorService.java, UinputInjector.java.
  • Shizuku dependency hardcoded to 13.1.5 (not in version catalog).
  • models.json outputSize field = num_outputs (e.g., 756 for 192 input, 1344 for 256 input).

Gotchas

  • QNN HTP only works on Qualcomm Snapdragon 8 Gen 1+. aimbot.cpp checks /proc/cpuinfo for "Qualcomm" and falls back to NNAPI.
  • libcdsprpc.so / libadsprpc.so are vendor libs — may not be found via dlopen on some devices. QNN delegate handles this internally.
  • extractNativeLibs=true in manifest — required for prebuilt .so files to be accessible at runtime.
  • The root_daemon binary uses stdin/stdout text protocol (not AIDL). Responses are OK, OK:<value>, or ERR:<message>. Debug output goes to stderr only.
  • FloatService inference runs on a single-thread executor at THREAD_PRIORITY_URGENT_DISPLAY.
  • Touch injection is NOT auto-triggered — inference only detects + overlays. Auto-aim/trigger controlled via aimbotOn AtomicBoolean in FloatService.
  • OverlayCanvasView draws on a full-screen TYPE_APPLICATION_OVERLAY window.