Developers building on-device AI experiences using Google's AI Edge library faced critical debugging blind spots: there was no way to view real-time logs from WebView-based Agent Skills, custom imported models crashed with native C++ Status Code 13 due to hardcoded multimedia requirements, and packaging release APKs required manual execution of local Gradle scripts.

Gallery-Extended is a developer-focused fork of the official Google AI Edge Gallery application. Designed to run local on-device machine learning models (like Gemma and Qwen) on Android, this project bridges critical debugging gaps for creators. It introduces a unified Developer Console that captures console logs from WebView Agent Skills and routes native C++ LiteRT engine errors. Additionally, it implements dynamic capability checks to allow text-only custom-imported models to execute without crashes, and integrates a robust CircleCI pipeline for automated release builds.

1. / 01Built a Unified Developer Console & Logging System: Created a thread-safe singleton (DevConsoleManager) to collect, wipe, and display both real-time JavaScript WebView console logs from Agent Skills and native C++ LiteRT LLM engine errors directly in a bottom sheet UI.
2. / 02Enabled Dynamic Model Capability Support: Decoupled hardcoded multimedia requirements in the Agent Chatroom, allowing text-only custom-imported models (like custom Gemma variants) to initialize safely by dynamically referencing their actual properties

• Zero-Blind Debugging: Reduced Agent Skill debug cycles by allowing developers to view WebView console outputs and native C++ error traces directly inside the chat UI without launching Android Studio.
• Robust Custom Model Support: Restored full compatibility for custom text-only imported model executions, eliminating native Status Code 13 (DYNAMIC_UPDATE_SLICE) prep crashes when loading text-only models in Agent Chat.

• eveloper Console
• WebView Agent Skills Logging
• Dynamic Model Loader