20 tools with consistent naming (adb_* prefix, verb-noun pattern) and descriptions in Portuguese. All tools have input schemas with type definitions and parameter descriptions. However, output schemas are not formally documented, responses are inferred from code (imageResult, textResult helpers). Error handling is basic (generic error wrapping). No tool annotations (readOnlyHint/destructiveHint). Parameter descriptions are adequate but lack format constraints (e.g., no regex for device_id, no bounds for numeric grid_step). Missing idempotency guidance for write operations. Average tool description ~80 chars, within baseline but terse.

• Document output schemas for all tools. Define return types explicitly: e.g., adb_list_devices returns {devices: [{serial, state, model?, product?}]}, adb_screenshot returns {image_base64, width, height, scale}. This enables LLM planning.
• Add tool annotations: mark adb_screenshot, adb_get_screen_info, adb_ui_dump, adb_find_elements, adb_get_current_app, adb_pull with readOnlyHint=true. Mark adb_tap, adb_swipe, adb_input_text, adb_keyevent, adb_start_activity, adb_stop_app, adb_install_apk, adb_push, adb_shell with destructiveHint=true.
• Enhance error handling: categorize failures (e.g., 'Device not found (retryable)' vs 'Invalid coordinates (user-fixable)'). Include recovery hints: 'No device connected. Call adb_list_devices to check available devices.'
• Add parameter constraints: device_id should note 'Required if multiple devices connected'; grid_step should have min=10, max=500; x/y in adb_tap should validate against screen bounds.
• Sanitize adb_shell input: reject commands containing pipes, redirects, or shell metacharacters unless explicitly whitelisted. Return 'Command contains unsafe characters' rather than executing. Or restrict to a whitelist of safe commands (getprop, dumpsys, etc.).