Helix-agi sensory-actualizor

Hi everyone, I've been working on this agent harness for a while and the most recent version uses dual RWKV models so one can speak and one can handle sensory intake and action output. The idea was to bypass the standard natural-language JSON tool calling in favor of a Bicameral Dual-RWKV non-language action loop. Rather than forcing a stateless model to break conversation, write out stringified JSON arguments, and parse regex, Helix-AGI splits cognition into a conversational Speaking Head and a Sensory-Actualizer Head coupled in latent space via a hidden-state gain of 0.15. The sensory head operates directly over a non-language vocabulary of primitive action and perception tokens (sense, move, read, write, run), routing operations through a shared Habitus hourglass bicone graph where conserved-fluid edge weights (\sum w = 1.0) and Y-axis travel-time ciphers turn raw operating-system file inspections and script executions into reflexive motor pathways without needing to continue to take up token space every time as with conventional markdown skills and pure RAG based behaviors. Essentially, the agent continually develops habitual behaviors and preferences, and uses that function to continually sync its dual LLM heads to ensure correct and accurate intentional actions based on actually learned desired outcomes and seamless sensory returns that imply self action in the context window without the need for system prompt schema. The UI lets you watch the graph function in real time and see what directory the agent's sensory intakes are pointed at, basically where it is, in it's environment. You can watch the agent move around and interact in its environment. It can switch between internal monologue and message output based on hidden state intent. Focus trajectory mapping and rolling JIT context window bidding lets the agent go on 'tangents' and sub tasks and 'snap back' to the main or previous topic or focus. The repo is pretty messy and some of the docs are still outdated but the main readme and important technical documents are updated and the setup wizard has a chatbot assistant. Would really love for some testers to give some feedback and even better if people want to collaborate! Thanks in advance! Link to the GitHub repo in the comments!