diff --git a/CLAUDE.md b/CLAUDE.md new file mode 100644 index 0000000..d32b00b --- /dev/null +++ b/CLAUDE.md @@ -0,0 +1,65 @@ +# CLAUDE.md + +This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository. + +## Project Overview + +Android Action Kernel — an AI agent that controls Android devices through the Accessibility API. It runs a Perception → Reasoning → Action loop: captures the screen state via `uiautomator dump`, sends it to an LLM for decision-making, and executes the chosen action via ADB. + +**Runtime:** Bun (TypeScript, ES2022 modules). Bun natively loads `.env` files — no dotenv needed. + +## Commands + +All commands run from `android-action-kernel/`: + +```bash +bun install # Install dependencies +bun run src/kernel.ts # Start the agent (interactive, prompts for goal) +bun run build # Compile to dist/ (bun build --target bun) +bun run typecheck # Type-check only (tsc --noEmit) +``` + +There are no tests currently. + +## Architecture + +Six source files in `android-action-kernel/src/`, no subdirectories: + +- **kernel.ts** — Entry point and main agent loop. Reads goal from stdin, runs up to MAX_STEPS iterations of: capture screen → diff with previous → call LLM → execute action → track history. Handles stuck-loop detection and vision fallback when the accessibility tree is empty. +- **actions.ts** — 15 action implementations (tap, type, enter, swipe, home, back, wait, done, longpress, screenshot, launch, clear, clipboard_get, clipboard_set, shell). Each wraps ADB commands via `Bun.spawnSync()`. `runAdbCommand()` provides exponential backoff retry. +- **llm-providers.ts** — LLM abstraction with `LLMProvider` interface and factory (`getLlmProvider()`). Four providers: OpenAI, Groq (OpenAI-compatible endpoint), AWS Bedrock (Anthropic + Meta model formats), OpenRouter (Vercel AI SDK). Contains the full SYSTEM_PROMPT with all 15 action definitions and rules. +- **sanitizer.ts** — Parses Android Accessibility XML (via `fast-xml-parser`) into `UIElement[]`. Depth-first walk extracting bounds, center coordinates, state flags (enabled, checked, focused, etc.), and parent context. `computeScreenHash()` used for stuck-loop detection. +- **config.ts** — Singleton `Config` object reading from `process.env` with defaults from constants. `Config.validate()` checks required API keys at startup. +- **constants.ts** — All magic values: ADB keycodes, swipe coordinates (hardcoded for 1080px-wide screens), default models, file paths, agent defaults. + +## Key Patterns + +- **Provider factory:** `getLlmProvider()` returns the appropriate `LLMProvider` based on `Config.LLM_PROVIDER`. Groq reuses the `OpenAIProvider` class with a different base URL. +- **Screen state diffing:** Hash-based comparison (id + text + center + state). After STUCK_THRESHOLD unchanged steps, recovery hints are injected into the LLM prompt. +- **Vision fallback:** When `getInteractiveElements()` returns empty (custom UI, WebView, Flutter), a screenshot is captured and the LLM gets a fallback context suggesting coordinate-based taps. +- **LLM response parsing:** `parseJsonResponse()` handles both clean JSON and markdown-wrapped code blocks. Falls back to "wait" action on parse failure. +- **Long press via swipe:** Implemented as `input swipe x y x y 1000` (swipe from point to same point with long duration). +- **Text escaping for ADB:** Spaces become `%s`, shell metacharacters are backslash-escaped in `executeType()`. + +## Adding a New LLM Provider + +1. Implement `LLMProvider` interface in `llm-providers.ts` +2. Add case to `getLlmProvider()` factory +3. Add config fields to `config.ts` and env vars to `.env.example` + +## Adding a New Action + +1. Add fields to `ActionDecision` interface in `actions.ts` +2. Implement `executeNewAction()` function +3. Add case to `executeAction()` switch +4. Document the action JSON format in `SYSTEM_PROMPT` in `llm-providers.ts` + +## Environment Setup + +Requires: Bun 1.0+, ADB (Android SDK Platform Tools) in PATH, an Android device connected via USB/WiFi with accessibility enabled, and an API key for at least one LLM provider (Groq, OpenAI, Bedrock, or OpenRouter). + +Copy `.env.example` to `.env` and configure `LLM_PROVIDER` + the corresponding API key. + +## Device Assumptions + +Swipe coordinates in `constants.ts` are hardcoded for 1080px-wide screens (center X=540, center Y=1200). Adjust `SWIPE_COORDS` and `SCREEN_CENTER_*` for different resolutions. diff --git a/android-action-kernel/README.md b/android-action-kernel/README.md new file mode 100644 index 0000000..83dacd4 --- /dev/null +++ b/android-action-kernel/README.md @@ -0,0 +1,82 @@ +# Android Action Kernel + +AI agent that controls Android devices through the Accessibility API. Give it a goal in plain English and it autonomously navigates the device using a Perception → Reasoning → Action loop. + +## How It Works + +1. **Perceive** — Captures the screen's accessibility tree via `adb shell uiautomator dump`, parses it into interactive UI elements with coordinates and state +2. **Reason** — Sends the screen context, action history, and goal to an LLM which decides the next action as a JSON object +3. **Act** — Executes the action (tap, type, swipe, launch app, etc.) via ADB +4. **Repeat** — Diffs the screen state, detects stuck loops, and continues until the goal is done or max steps reached + +Falls back to screenshot-based vision when the accessibility tree is empty (games, WebViews, Flutter). + +## Prerequisites + +- [Bun](https://bun.sh) 1.0+ +- [Android SDK Platform Tools](https://developer.android.com/tools/releases/platform-tools) (ADB in PATH) +- Android device connected via USB or WiFi ADB +- API key for one of: Groq, OpenAI, AWS Bedrock, or OpenRouter + +## Quick Start + +```bash +cd android-action-kernel +bun install +cp .env.example .env +# Edit .env — set LLM_PROVIDER and the corresponding API key +bun run src/kernel.ts +``` + +The agent will prompt you for a goal, then start controlling the device. + +## Configuration + +Copy `.env.example` to `.env`. Key settings: + +| Variable | Default | Description | +|---|---|---| +| `LLM_PROVIDER` | `groq` | `groq`, `openai`, `bedrock`, or `openrouter` | +| `MAX_STEPS` | `30` | Maximum actions before stopping | +| `STEP_DELAY` | `2` | Seconds between actions (lets UI settle) | +| `STUCK_THRESHOLD` | `3` | Unchanged screens before recovery kicks in | +| `VISION_ENABLED` | `true` | Screenshot fallback when accessibility tree is empty | + +### LLM Providers + +| Provider | Key Variable | Default Model | +|---|---|---| +| Groq (free tier) | `GROQ_API_KEY` | `llama-3.3-70b-versatile` | +| OpenAI | `OPENAI_API_KEY` | `gpt-4o` | +| AWS Bedrock | AWS credential chain | `us.meta.llama3-3-70b-instruct-v1:0` | +| OpenRouter | `OPENROUTER_API_KEY` | `anthropic/claude-3.5-sonnet` | + +## Available Actions + +The agent can perform 15 actions: + +| Category | Actions | +|---|---| +| Navigation | `tap`, `longpress`, `swipe`, `enter`, `back`, `home` | +| Text | `type`, `clear` | +| App Control | `launch` (by package, activity, or URI with extras) | +| Data | `screenshot`, `clipboard_get`, `clipboard_set` | +| System | `shell`, `wait`, `done` | + +## Project Structure + +``` +src/ + kernel.ts # Main agent loop (entry point) + actions.ts # ADB action implementations with retry + llm-providers.ts # LLM abstraction (OpenAI, Groq, Bedrock, OpenRouter) + sanitizer.ts # Accessibility XML parser + config.ts # Environment config loader + constants.ts # ADB keycodes, coordinates, defaults +``` + +## Notes + +- Swipe coordinates in `constants.ts` are calibrated for 1080px-wide screens. Adjust `SWIPE_COORDS` for different resolutions. +- The agent automatically detects stuck loops and injects recovery hints after `STUCK_THRESHOLD` steps without screen changes. +- ADB commands retry with exponential backoff (up to `MAX_RETRIES` attempts).