Documentation Overhaul and Architecture Pruning

The Great Documentation Purge

The documentation had reached a point where it was drifting significantly from the implementation. I spent a large chunk of time today deleting over 2,000 lines of redundant or speculative documentation, including entire files dedicated to the uncertainty engine and various cognitive architectures that have since evolved. I fixed over 60 stale claims—updating details on our use of Moonshine and Kokoro for voice, ONNX for embeddings, and more accurate descriptions of the ACT loop’s iteration limits. The README was also rewritten to prioritize a one-line install and a grounded “Why Chalie?” value proposition over technical jargon.

Skill Consolidation and Loop Simplification

I refactored the skill system to be leaner by merging the notes_skill directly into recall_skill. Transcript searching and memory retrieval are now handled by a single entry point, simplifying the agent’s decision-making process.

Inside the ACT loop, I removed the “scratchpad” mechanism. Previously, large results were gated to a separate scratchpad to save tokens, but this added unnecessary complexity to the orchestration. Now, results stay inline, and we simply drop the oldest history entries when the context exceeds the token budget. This makes the reasoning flow much easier to debug and follow.

World State and Observability

To make the agent’s internal state more transparent, I added a new /system/observability/world-state endpoint. This provides a structured breakdown of the agent’s current understanding, which is now visible in a new “World State” tab in the Brain dashboard. This view includes a prompt injection preview so I can see exactly what the model sees.

I also took the opportunity to prune several dead API endpoints—like the old conversation summary and memory forget routes��that were no longer being used by the frontend. The voice pipeline also received some stability improvements, particularly around the renderer and gateway handling.