Discover and explore top open-source AI tools and projects—updated daily.
GongShichenOn-device Agent Runtime for controllable, observable, and recoverable AI agents
Top 93.1% on SourcePulse
Lumina provides a runtime for edge-native agents, enabling controllable, observable, and recoverable execution on devices. It targets developers building agent applications for iOS, Android, and HarmonyOS, offering a stable execution contract that integrates model outputs, tool schemas, permissions, and session state.
How It Works
Lumina features a cross-platform C/C++ core runtime that orchestrates agent execution. It implements a canonical ReAct loop, parsing model callbacks, validating steps, and dispatching tools. Key components include a Tool Registry for schema management and validation, Deferred Tool Loading to optimize prompts, Progressive Context Loading managed by the host, and robust Session/Checkpoint/Replay capabilities. Hooks and Guardrails allow host-level control, while optional Observability sinks provide tracing and auditing.
Quick Start & Requirements
Integration targets iOS (Swift Package), Android/HarmonyOS (C ABI/JNI). Model management and preparation are handled via the ./model/lumina_model.sh script, supporting download, training, and installation of models like MiniCPM-V 4.6. Development requires Xcode and Android NDK. Building involves cmake for Android and xcodebuild for Apple platforms. Official quick-start guides or demo links are not explicitly provided within the README.
Highlighted Details
Designed for edge deployment with a cross-platform C/C++ core. Offers advanced agent execution features like deferred tool loading and progressive context management. Demonstrates significant performance gains with its trained MiniCPM-V 4.6 model, improving pass@1 from 3/200 to 101/200 and tool recall from 5% to 82%.
Maintenance & Community
The README does not detail specific contributors, community channels (e.g., Discord, Slack), roadmap, or sponsorship information.
Licensing & Compatibility
The repository's license is not specified in the README. This omission prevents evaluation of commercial use compatibility or any copyleft implications.
Limitations & Caveats
The primary adoption blocker is the unspecified license. Detailed setup time, resource footprints, and specific bug reports are absent. While testing scripts are present, the overall stability and maturity (e.g., alpha/beta status) are not explicitly stated.
1 month ago
Inactive