flutter_mind_local 0.2.1

On-device LLM inference via llama.cpp. Drop LocalEngine into flutter_mind's FlutterMindClient — same API as every other engine.

0.2.1

• Fixed the README logo not rendering on pub.dev — pinned to a raw.githubusercontent.com URL instead of a relative path.

0.2.0

• LocalEngine is now a drop-in AiEngine for flutter_mind's FlutterMindClient — same send/stream/countTokens API as every other engine.
• Real token-by-token streaming via stream() — tokens arrive as they're generated, not all at once at the end.
• Fixed an Isolate.run bug where the whole LocalConfig (including the unsendable onEvent closure) was captured instead of just the plain init args, causing a runtime crash.
• Fixed android/build.gradle.kts namespace colliding with the consuming app's own namespace.
• Native build now forces Release optimizations regardless of Flutter build variant — debug builds were previously 10-20x slower at inference.
• Android support verified on a real device. iOS/macOS build files exist (Swift Package Manager) but remain untested — not yet enabled in pubspec.yaml.

0.1.0

• Initial release.
• LocalEngine — on-device LLM inference via llama.cpp, no API key or internet required.
• LocalConfig — full configuration: model path, system prompt, temperature, context size, sampling parameters.
• LocalModelType — built-in chat templates for Qwen, Llama 3, Gemma, Phi, Mistral, DeepSeek, with auto-detection from .gguf metadata.
• Lifecycle events via onEvent: ModelLoadStarted, ModelReady, ModelFailed, InferenceStarted, InferenceCompleted, InferenceFailed, ContextCleared, ModelDisposed.
• Android support via CMake + FetchContent (llama.cpp built from source).
• iOS and macOS support via Swift Package Manager (llama.cpp built from source).
• Model loading and inference run on background isolates — UI thread never blocked.