Adaptive Efficient Token Hybrid Evaluation Router

AETHER is a high-performance, token-efficient hybrid AI agent designed to process natural language tasks across 8 capability categories, intelligently routing simple/context-contained tasks to a local LLM running on CPU, and falling back to Fireworks cloud models for complex reasoning. Context-Contained Tasks (Local Engine): Tasks like Sentiment Analysis, Summarization, and Named Entity Recognition (NER) are routed locally to a GGUF quantized Qwen2.5-1.5B-Instruct model running directly on CPU (costing 0 API tokens). Complex Synthesized Reasoning (Cloud Engine): Tasks requiring math, logic, code generation, or bug fixing are routed to the most capable permitted Fireworks models in the ALLOWED_MODELS list (leveraging minimal system prompts for maximum token efficiency). Hybrid Factual Checking: Factual queries are first checked by the local model; if the model is unsure of the facts, it transparently routes to the cloud engine to prevent hallucination. Fail-Safe Fallbacks: If the cloud API fails or the local model fails to load, the router falls back gracefully to secondary models, ensuring a valid exit code 0 and complete outputs.