* Settings & Storage Screen * * Section order: API Configuration, Generation Settings, Tool Calling, * Storage Overview, Downloaded Models, Storage

* Speech to Text Screen * * Features: * - Batch mode: Record full audio then transcribe * - Live mode: Real-time streaming transcription * - Reco

* Stored Model Row */

SystemTTSRow: card-style row matching ModelCard, "Use" action

* TTS app bar model chip - same style as ChatTopBar: pill Surface, model icon + name + Streaming. */

* Text Input Section */

* Text to Speech Screen * * Features: * - Text input area with character count * - Voice settings (speed, pitch sliders) * - Generate/Speak butto

* Shows streaming LLM tokens in real-time while the model is reasoning. * Styled in amber/yellow to visually distinguish from the completed log panel

* Thinking Progress Indicator - pattern * Shows "Thinking..." with animated dots when message is empty but thinking content exists */

* Tool calling indicator badge - matching iOS ChatInterfaceView toolCallingBadge */

* Tool Calling Settings Section * * Allows users to: * - Enable/disable tool calling * - Register demo tools (weather, time, calculator) * - Clear

* Transcribing badge - iOS style orange processing indicator */

* Transcription display area * iOS Reference: Transcription ScrollView in SpeechToTextView */

* Metrics bar showing transcription statistics * Clean, minimal design that doesn't distract from the transcription */

Typing indicator — text-based shimmer

* VLM Screen — Vision Language Model interface with live camera. * Mirrors iOS VLMCameraView.swift exactly. * * Features: * - Live camera preview

* Vision Hub Screen — Matches iOS VisionHubView exactly. * * Lists vision-related features: * 1. Vision Chat (VLM) — Chat with images using photos

* Voice Assistant screen * * This screen shows: * - VoicePipelineSetupView when not all models are loaded * - Main voice UI with conversation bubb

* Voice Settings Section with Speed and Pitch sliders. * Pitch parameters are currently unused because the pitch slider is commented out, * but reta

* Core initialization logic — only called once, guarded by initialize().

Internal: Transcribe via online recognizer (for streaming models used non-streaming)

* Add a breadcrumb for the log entry

* Add a log destination (non-suspending version). */

* Add a message to a conversation */

* Add a single model to the registry without replacing existing ones. * Used for importing models via file picker or drag-and-drop. * If a model

Outputs double if is_valid is true, otherwise outputs null

Outputs a string if value is non-null, otherwise outputs null

* Build an OPFS key for additional files (e.g., mmproj). * Uses `__` separator to avoid name collisions between * a primary model FILE and a dir

* Analyze a screenshot and return a text description of the screen. * * @param screenshotBase64 Base64-encoded JPEG screenshot * @param

* Append a metrics footer below a message bubble.

* Create an assistant message */

* Current audio level (0.0 - 1.0)

* Authenticate with API key * @param apiKey API key * @returns true if authenticated successfully

Base64 decoding for rgbPixels image format

* Decode base64 string to bytes */

* Simple bilinear resize without stb */

* Bind camera preview + image analysis to the given PreviewView. * Mirrors iOS setupCamera() + startCamera(). * * Uses LifecycleCameraC

* Build the assistant message JSON for a tool_calls response. * Used to construct conversation history for multi-turn tool calling. */

* Build follow-up prompt after tool execution. * Uses C++ when available, falls back to TypeScript.

* Build follow-up prompt using C++ bridge * Falls back to template string if native unavailable

* Build initial prompt with tools (IMPLEMENTED in C++ ToolCallingBridge) * * Combines user prompt with tool definitions and system instructions.

* Build initial prompt using C++ bridge * Falls back to simple concatenation if native unavailable

* Build a RAGConfiguration from selected models, injecting the vocab path. * Returns null if either model is not yet selected or has no local path.

* Build system prompt for structured JSON output

* Build a tool result message for the conversation history. */

* Cache multi-file descriptors for later retrieval during download. * The C++ registry doesn't preserve the file descriptor array. * Implemented via

Get cache size. */

Delegate to C++ for recursive directory size calculation

* Recursive directory size calculation. * This is the core logic duplicated across all SDKs. * * Algorithm (identical in Swift/Kotlin/Flutter/RN):