* Handle unexpected stream end

* Handles streaming response from the API * @param response The fetch response object * @param onStreamChunk Optional callback for each chunk * @re

* Streaming handler for tool-calling requests. A single turn may interleave prose and * tool-call fragments; prose is forwarded via onStreamChunk whi

Process the request asynchronously. Subclasses MUST override this method. Args: request_data: The parsed JSON re

* Check if a service is active for a stream type.

Initializes the screen capture plugin

* Initialize the llama.cpp backend engine explicitly. * Idempotent — safe to call multiple times. On first call this triggers * Metal shader com

Initialize the ring buffer (creates/opens shared file)

Initialize the memory-mapped file

Initialize the memory-mapped file

Install our log hook into llama.cpp and ggml so all C-level logs flow through `_emit_log` and reach the JS Logger.

* An iteration is "settled" once the micro-agent finished a full pass — the model has * responded (or the pass was skipped). Until then its screensho

* Check if the loaded model supports multimodal (vision) input * Returns true if model has an associated mmproj file loaded

* Check if a stream is available

Check if this is a first run (no default channel configured)

Check if shorter is a subsequence of longer (all chars appear in same order).

Allow localhost, tauri, or observer-ai.com origins.

Check if video streaming is active

Build the full multimodal ToolResult (summary + prompt + sensor list + images) for one iteration.

List files in the App Group container (iOS only, for debugging)

* Load a local model by its ID and server type. * Routes to the appropriate underlying manager based on the server sentinel.

* Load a model using saved settings (or defaults if none saved). * This is the primary API - settings are automatically fetched from storage.

Load a multimodal projector (mmproj) file for vision support