* Detect error messages

* Detect login required

* Detect obstacles that require user intervention

* Detect out of stock items

* Detect the appropriate tool call format based on model name. * LFM2-Tool models use the lfm2 format, others use default JSON format. */

* Detect model format from path. Only Apple CoreML diffusion is supported. * ONNX diffusion is not supported; we only look for CoreML. */

* Detect VLM model type from model name metadata. */

* Release all audio resources.

* Perform a download with custom logic. * * @param url The URL to download from * @param destinationPath The local file pat

* Download an embedding model and its companion files (e.g., vocab.txt) using direct HTTP. * All files are saved into the same directory: {base}/mode

* Download a file using HttpURLConnection with redirect support. */

* Download a model * @param modelId Model identifier * @param url Download URL * @param destPath Destination path * @returns true if downl

* Build a HuggingFace download URL for this chip. * @param modelSlug Model slug (e.g. "qwen3-4b") → produces * "qwen3-4b-genie-w4a16-8elit

* Draws batched points grouped by quantized color to minimize draw calls. * Colors are quantized to ~16 levels per channel to group similar particles

* Draws all particles with minimal allocations. * Small particles are batched into a single drawPoints call. * Only larger/brighter particles get in

* Emit device registered event via C++. */

* Emit device registration failed event via C++. */

* Emit download started event via C++. * * @param modelId The model being downloaded * @param totalBytes Expected total bytes (0 if unk

* Emit SDK init completed event via C++. */

* Emit SDK init started event via C++. */

Helper to emit device registered event

* Registers a service worker that injects COOP/COEP headers for browsers * that don't support `credentialless` COEP (Safari/WebKit). * * - On Chrom

* Ensure the current chat is in the store's list and persisted before showing history. * Syncs latest messages to the store and adds the conversa

* Ensure a directory exists in the sherpa-onnx Emscripten virtual FS.

* Ensure the llama.cpp VLM backend is registered with the service registry. * Must be called before creating the VLM component so it can find a pro

* Ensures espeak-ng voice files from lang/ subdirectories are also * accessible directly under voices/ so that espeak_SetVoiceByName() * can find th

* Ensure trailing slash on directory path. */

* Event callback invoked by C++ core when an analytics event occurs. * * Routes events to the registered [AnalyticsEventListener] if one is

Convert analytics event type to modality

Convert analytics event type to event type string

* Evict old models from OPFS to free space for a new model. * Deletes models sorted by oldest-first until enough space is freed, * skipping the

* Click an element by selector with retry logic

* Extract text content from an element

* Press Enter key on an element (for submitting forms/search)

* Scroll the page

* Execute a tool call by looking up the registered executor.

* Extract alert/notification messages from the page

* Extract Amazon-specific interactive elements with priority

* Extract an archive to the model directory using native C++ extraction (libarchive). * * Supports all formats via auto-detection: ZIP, TAR.GZ, TAR.

* Extract file, line, and function from the first relevant stack frame.

* Extract raw RGB pixels from a canvas 2D context. * * Canvas gives RGBA; this strips the alpha channel to produce the * RGBRGBRGB... byte la

* Extract search query from task without LLM

* Extract plain text from a file identified by a content URI. * * Supports PDF (via pdfbox-android) and JSON (via org.json). * Mirrors

* Extract text from a document using native PDFKit (for PDF) or string parsing. * Mirrors iOS DocumentService.swift - handles PDF, JSON, and plain te

* Extract visible text content from the page

* Extract YouTube-specific interactive elements with priority * YouTube uses web components that standard querySelectorAll may miss

* @brief Extract a JSON array as a raw string (including brackets) */

* @brief Extract a JSON object as a raw string (including braces) */

* @brief Extract tool name and arguments using multiple strategies * * Strategies in order: * 1. Standard format: {"tool": "name", "arguments": {..

Simple URL host extraction (after ://)

Simple URL scheme extraction

* Strip keys that don't exist in the V2 `telemetry_events` table. * Handles both array format (dev flat batch) and single-object format.

* Find the best matching product from search results

* Find button by text content

* Find input element by placeholder, label, or aria-label

* Find link by text content

* @brief Find matching closing bracket for a JSON array * * Tracks string boundaries to ignore brackets inside strings. * * @param str String to s