Boundary Reliability and Interface Blocks

Topic Boundary Refactor

We’ve replaced the complex 3-layer AdaptiveBoundaryDetector—which utilized NEWMA, Transient Surprise, and Leaky Accumulators—with a more streamlined TwoSignalBoundaryService. Despite its theoretical alignment with Event Segmentation Theory, the previous model consistently underperformed in real-world benchmarks. The new approach combines consecutive and window similarity checks with self-calibrating thresholds and a fast path for specific discourse markers. This shift improved our F1 score from 0.50 to 0.67, providing a much more reliable foundation for thread segmentation and focus switching.

P5 Block Protocol

The interface daemons now support the P5 block protocol, introducing 11 new renderers for layout and interaction, including columns, tabs, forms, and toggles. These components are styled with the Radiant CSS framework and wired to handle complex actions like form collection and asynchronous polling. As part of this UI stabilization, we’ve also cleaned up the Brain dashboard’s entry point to properly redirect unauthenticated sessions back to the main login flow with a return path.

Execution and Routing Fixes

A critical bug in the digest_worker was identified where responses were being generated but never delivered to the frontend. The worker was calling the generation service directly, bypassing the orchestrator responsible for WebSocket dispatch. We’ve restored the proper dispatch pattern, ensuring that asynchronous background processing correctly routes back to the user.

On the frontend, we fixed a race condition in the ACT loop. Previously, if a narration event removed a pending form from the DOM before a response arrived, the response was rendered into a detached element and lost to the user. The renderer now checks for connection status and falls back to a fresh form append if the original container is gone.

Cognitive Cleanup

To keep the backend lean, we’ve removed three deprecated cognitive jobs: mode-reflection, tool-prefilter, and fact-store. These jobs were either uncalled or referenced services that had been previously refactored or deleted, such as the old ONNX skill selector.