# Botty Architecture ## Overview Botty is an autonomous bot for Diablo II Resurrected (D2R) that uses computer vision and native Windows API input to run endgame content without kernel drivers. It's built in Python 3.10 and distributed as a PyInstaller onefile executable. ## Entry Points Two standalone executables are built from the same source: | Executable | Source | Purpose | |---|---|---| | `main.exe` | `src/main.py` | Main bot controller | | `shop.exe` | `src/shopper.py` | Standalone vendor shopping | ## Threading Model main.py spawns three concurrent threads: ``` main.py (main thread) ├── game_controller.py (bot loop thread) │ └── bot.py (state machine thread) ├── health_manager.py (background thread) └── death_manager.py (background thread) ``` - **Main thread**: Shows UI, registers hotkeys (F7-F12), calls `keyboard.wait()` to block - **Game Controller**: Orchestrates the bot loop - starts/stops the bot thread, handles game recovery, tracks stats - **Health Manager**: Polls health/mana/merc every ~1s, auto-potions, chickens when low - **Death Manager**: Polls every ~1s for "You Have Died" screen, triggers recovery ## Core Modules ### Input Layer (`src/input_layer/`) Native Windows API input replacement designed to evade kernel-level anti-cheat (Warden). It bypasses common Python libraries like `pyautogui` or `pynput` which can be easily detected. | File | Purpose | |---|---| | `win_input.py` | Low-level `ctypes` wrappers for `SendInput`, `GetAsyncKeyState`, `GetCursorPos`. Uses standard Windows user-mode APIs. | | `mouse_impl.py` | Humanized mouse controller. Features include: Bezier curve trajectories, Gaussian noise (hand tremor), endpoint wobble, and randomized arrival-to-click delays. | | `hotkey.py` | Polling-based hotkey manager that avoids global hooks. All polling intervals include micro-jitter. | | `__init__.py` | Drop-in API that shims standard input calls with stealth timing and variable duration automatically. | All key presses include stealth micro-pauses and variable press duration automatically. ### Screen Capture (`src/screen.py`) Handles D2R window detection and screenshot capture via MSS library. Converts between coordinate systems: Monitor (top-left first monitor), Screen (per-monitor), Absolute (character center), Relative (template-matched). ### Image Recognition (`src/template_finder.py`) Template matching via OpenCV (`cv2.matchTemplate`). Searches against pre-captured asset templates in `assets/templates/`. Returns match position and validity. **Asset conventions — read before adding or recapturing any template:** - Every `.png` under the `TEMPLATE_PATHS` roots is loaded recursively and keyed by its **uppercased filename** (`a5_red_portal.png` → `A5_RED_PORTAL`). Keys are one flat namespace across all roots, so filenames must be globally unique — and **never leave a backup or scratch `.png` anywhere under `assets/`**, or it silently becomes a live template. - **Keep templates fully opaque (3-channel, or 4-channel with no zero-alpha pixel).** `alpha_to_mask` only produces a mask when the image has 4 channels *and* contains a fully transparent pixel; that mask is then passed to `cv2.matchTemplate(..., TM_CCOEFF_NORMED, mask=...)`. **OpenCV only properly supports masks for `TM_SQDIFF` and `TM_CCORR_NORMED`** — masked `TM_CCOEFF_NORMED` returns unreliable scores and wandering match positions. A masked template will appear to "work" in isolation and then fail at random. See CLAUDE.md Bug 23. - Crop something **structurally stable and unoccluded**. Animated or partly hidden features (a swirling portal's interior, a ring occluded by scenery) make poor anchors. - **Validate on held-out frames**: build the crop from one capture, score it against *other* captures, and — critically — against frames where the subject is **absent**. A template that scores high on both is matching background, not the subject. Aim for a clear gap straddling the 0.68 default threshold. ### UI Detection (`src/ui/`) | File | Detects | |---|---| | `main_menu.py` | Main menu buttons (play, save & exit) | | `character_select.py` | Character portraits and selection | | `skills.py` | Skill bar state, skill availability | | `meters.py` | Health/mana globes, potion charges | | `player_bar.py` | Player status bar, XP bar | | `view.py` | Minimap, inventory screen detection | | `waypoint.py` | Waypoint portal selection | | `error_screens.py` | Error dialogs, death screen | | `loading.py` | Loading screen detection | ### Pathing (`src/pather.py`) Pathfinding via reference template matching with relative coordinates. Each location has numbered nodes with template references. The `Location` class defines named destinations (towns, dungeons). Path definitions map (start, end) pairs to node sequences. ### Character System (`src/char/`) Inheritance hierarchy: ``` IChar (abstract base) ├── Basic / Basic_Ranged ├── Paladin │ ├── Hammerdin │ └── FoHdin ├── Sorceress (base for all sorc builds) │ ├── BlizzSorc │ ├── BlizzorbSorc │ ├── NovaSorc │ ├── LightSorc │ └── HydraSorc ├── Amazon │ └── Javazon ├── Trapsin ├── Barbarian ├── Necro ├── Poison_Necro ├── Bone_Necro └── Warlock ├── FireLock ├── EchoLock └── AbyssLock ``` Each character class implements: run methods, skill usage, attack patterns, position-specific behavior. ### Run System (`src/run/`) | File | Run | |---|---| | `pindle.py` | Pindle of Purity | | `shenk_eld.py` | Shenk + Eldritch | | `trav.py` | Travincal | | `nihlathak.py` | Nihlathak's Temple | | `arcane.py` | Arcane Sanctuary | | `diablo.py` | Chaos Sanctuary + Diablo | | `vizier.py` | Vizier (Seal Boss) | | `baal.py` | Baal | | `mephisto.py` | Mephisto | | `andariel.py` | Andariel | | `countess.py` | Countess | | `level.py` | General leveling runs | ### Town System (`src/town/`) `TownManager` orchestrates act-specific town routines. Each act (A1-A5) has its own module with NPC interactions, waypoint usage, and stash handling. ### Inventory (`src/inventory/`) | File | Purpose | |---|---| | `belt.py` | Belt slot management, potion detection | | `personal.py` | Personal inventory grid, item positions | | `vendor.py` | Vendor trade window | | `stash.py` | Personal stash, shared stash | | `cube.py` | Horadric Cube | | `common.py` | Shared inventory utilities | ### Item Recognition (`src/item/`) | File | Purpose | |---|---| | `pickit.py` | Pickup decisions, item filtering | | `consumables.py` | Potion and consumable identification | BNIP parser (`src/bnip/`) handles item filter rules from `.bnip` files. ### Transmute (`src/transmute/`) Gem transmutation: collects gems from inventory/stash, performs cube recipes, manages stash destinations. ### Messages (`src/messages/`) | File | Purpose | |---|---| | `messenger.py` | Message dispatcher | | `discord_embeds.py` | Discord webhook formatting | | `generic_api.py` | Generic HTTP webhook support | ### Configuration (`src/config.py`) Singleton that merges config files in priority order: `custom.ini` > `params.ini` > `game.ini` > `shop.ini` > `transmute.ini` Supports variable substitution via `[variables]` sections. ### Stealth System (`src/utils/stealth.py`) A multi-tiered approach to mimicking human behavior and evading detection: - **Tier 1: Input Stealth**: Automatic micro-pauses (20-120ms), variable key press durations (20-200ms), and non-linear mouse paths via `input_layer`. - **Tier 2: Behavioral Stealth**: Probabilistic "mistakes" such as clicking the wrong waypoint (2.5% chance) or skill hesitation (80-300ms) before casting. - **Tier 3: Session Stealth**: Randomized run durations (+/-15%), AFK breaks (2-12 mins), and shuffling of farming routes between rotations. All timing across the bot is routed through `utils.misc.wait()`, which applies Gaussian jitter to every sleep call. ### Utilities (`src/utils/`) | File | Purpose | |---|---| | `misc.py` | Window management, DPI awareness, timing utilities | | `restart.py` | Game launch/kill, D2R always-on-top | | `auto_settings.py` | D2R game settings adjustment | | `key_detector.py` | Auto-detect skill bindings from `.key`/`.keyo` files | | `graphic_debugger.py` | Visual debugging overlay | | `node_recorder.py` | Record new path nodes | | `stealth.py` | Stealth behavior randomization | ### Auxiliary Tools Standalone tools located in the root directory for project maintenance and development: - `asset_manager.py`: Unified interface for auditing, searching, and optimizing template assets. - `asset_extractor.py`: Screenshot capture and AI-assisted entity cropping workflow. - `build.py`: PyInstaller wrapper for building production executables. - `desktop_snap.py`: Lightweight tool for capturing full desktop screenshots. - `quest_debug.py`: Debugging interface for the questing system. - `screenshot_tool.py`: Simple utility for taking D2R client area screenshots. ## Data Flow ``` D2R Game Window │ ▼ screen.grab() ──→ Image (numpy array) │ ▼ template_finder.search() ──→ Position + Validity │ ▼ ui_manager.detect_screen_object() ──→ Current UI state │ ▼ bot.py (state machine) ──→ Decides next action │ ▼ char.run() / pather.go_to() ──→ Movement + Interaction │ ▼ input_layer.keyboard.send() / mouse.click() ──→ SendInput API │ ▼ D2R Game Window (response captured on next grab) ``` ## Build System - **Build script**: `build.py` (PyInstaller with `--onefile --noconsole`) - **Dependencies**: `pyproject.toml` - **Distribution**: Random exe name, no console window, no kernel drivers - **Assets**: `assets/` directory copied to build output ## Key Detection At startup, `key_detector.py` reads the character's `.key`/`.keyo` file from `Saved Games/Diablo II Resurrected/` to auto-detect skill bindings and non-skill keys (inventory, show items, etc.). This eliminates manual key configuration. Slot-to-key mapping (`CHAR_BINDING_SLOTS`): | Slot | Config key | |---|---| | 41 | `show_belt` | | 36 | `stand_still` | | 44 | `weapon_switch` | | 43 | `force_move` | If params.ini and the `.keyo` file disagree, params.ini wins and a `"Keeping configured key binding"` line is logged. No disagreement = no log line. ## Coordinate Systems | System | Origin | Used By | |---|---|---| | Monitor | Top-left of first monitor | `screen.grab()` | | Screen | Per-monitor client area | UI detection | | Absolute | Character at screen center | Pathing, target detection | | Relative | Template match position | Inventory grid, NPC interaction | Conversion functions in `screen.py`: `convert_monitor_to_screen()`, `convert_screen_to_abs()`, etc. ## State Machine `bot.py` uses the `transitions` library. States and transitions are defined at `Bot.__init__` time. **States**: `initialization`, `hero_selection`, `town`, `level`, `cold_plains`, `baal_xp`, `pindle`, `shenk`, `trav`, `nihlathak`, `arcane`, `diablo`, `vizier`, `baal`, `mephisto`, `andariel`, `countess` **Baal XP leech** (`baal_xp` state, `on_run_baal_xp`): a self-contained cycle that does NOT use the run_wrapper/approach/battle pattern. It leaves the bot's own game, joins a public Baal game via the online lobby (`src/ui/game_browser.py`), hides and stands still to collect shared area XP, then leaves and re-enters its own game. The hide-and-wait phase is a pure wait — no maintenance, no runs, no pathing; the health manager thread keeps auto-potioning and the death manager handles the death screen. Config lives in the `[baal_xp]` section of `params.ini`; the route is `run_baal_xp` and re-arms itself between cycles via `_baal_xp_rearm`. **Key transitions**: | Trigger | Source | Dest | Notes | |---|---|---|---| | `init` | initialization | initialization | Screen detection, routes to create_game or start_from_town | | `select_character` | initialization | hero_selection | | | `start_from_town` | initialization/hero_selection | town | | | `maintenance` | town | town | Heal, buy pots, stash, repair, resurrect merc | | `run_pindle` | town | pindle | | | `run_arcane` | town | arcane | | | `end_run` | any run state | town | TP back; calls `on_end_run` which TPs to town | | `end_game` | town/any run state | initialization | Save & exit; use when no TP scrolls or unrecoverable | `end_run` requires working TP scrolls. If charges = 0, trigger `end_game` instead — otherwise the bot loops trying to TP back indefinitely. ## HealthManager — Internal Timing The background health monitor (`src/health_manager.py`) polls at: ``` interval = max(0, (3/25 - fn_elapsed) * jitter(0.8–1.2)) ``` That's approximately every 3 game frames (96–144ms at 25 FPS). The jitter prevents perfectly regular polling patterns from being detectable. **Rejuv logic**: 1. Minimum 0.60s between rejuv drinks (hit recovery guard). 2. Drinks rejuv if `health ≤ take_rejuv_potion_health` OR `mana ≤ take_rejuv_potion_mana`. 3. "Double rejuv" chicken fires only if `last_drink < 8s` **AND** `health ≤ take_rejuv_potion_health`. - The HP check is critical: mana-triggered rejuvs can legitimately fire back-to-back at full HP (Hammerdin spending mana fast). Without the HP check, false chickens occur at 99.9% HP. ## PickIt — Item Identity `GroundItem` has two identity fields: | Field | Formula | Purpose | |---|---|---| | `ID` | `slugify(Name + all as_dict() values including Amount)` | Pickit cache key; different gold amounts = different IDs | | `UID` | `ID + screen center position` | Deduplication within a single items list; same pile at same coords = same UID | The fail-detection in `_pick_up_item` uses `item.ID == prev.ID` for gold and `item.UID == prev.UID` for everything else. Two nearby gold piles with different amounts bypass the ID check since they produce different IDs. `_yoink_item()` always returns `PickedUpResult.PickedUp` regardless of actual success. True pickup failures surface only through `_pick_up_item`'s same-ID/UID repeat detection. On confirmed failure, the item's `ID` is blacklisted in `_cached_pickit_items` so it isn't retried in the same session. ## Configuration Priority Merge order (highest priority first): ``` custom.ini > params.ini > game.ini > shop.ini > transmute.ini ``` `Config` is a singleton (`__new__` + `data_loaded` class variable). First instantiation loads all files; subsequent calls return the same instance. Key detection runs during `__init__` after `self.char` is populated. ## Known Architectural Issues See `IMPROVEMENTS.md` for the full list. Highest-priority unresolved items: - **C10**: `kill_thread()` uses `PyThreadState_SetAsyncExc` (CPython private API). Can corrupt locks/GIL. Replace with `threading.Event` cooperative shutdown. - **H12/H14**: Health/death managers use module-level globals for state. HealthManager now has a `_state_lock`; death manager does not yet. - **M14**: `PickedUpResult` enum has a gap (values 0,1,3,4,5 — missing 2). - **H10/H11**: `pather.py` (750 lines) and `config.py` are oversized and should be split.