mountain-dew-esports

Hybrid Multi-Backend Compute Routing: Omen abstracts inference behind a pluggable orchestration layer. Developers can assign a different runtime to each agent and hot-swap it mid-session. Supported runtimes include local daemons (Ollama, llama.cpp), serverless cloud GPUs (Modal T4/A100/H100), and external APIs (Meta Muse Spark/Wonder, DeepSeek-V3/R1, OpenAI GPT-4o). Distributed LAN Hardware and VRAM Pooling: Omen clusters multiple machines on the same local network (e.g., an Apple M4 workstation and a Windows RTX 3050 laptop). A zero-install, 4-digit PIN web pairing portal (omen-worker.exe) connects them. Omen then splits model layers across the combined VRAM and distributes CPU threads, enabling fully offline token ingestion, AST parsing, and large-context inference. Multi-File Agentic Diff Engine and Workspace Scoping: Through the Agent Hub and execution manager, models can reason over any project subtree and stream edits across multiple files simultaneously. Changes appear as atomic unified or side-by-side diffs with transactional Accept and Reject controls. Local-First RAG and Real-Time Context Telemetry: An embedded vector database (supporting nomic-embed-text and bge-m3) provides privacy-preserving semantic codebase indexing. A live Context Token and KV-Cache Meter shows prompt capacity and context window usage, up to 1,000,000 tokens.