AI Core Updates: DMN, Context, and Performance Boosts

The DMN functionality was stabilized by replacing the broken unified_generate path with DMNMessageProcessor, enabling the full tool loop for DMN interactions.

The system now supports a DB-backed context window with compaction to eliminate Out-of-Memory (OOM) errors during tool loop execution.

Scheduled prompts now use a dedicated ScheduledMessageProcessor, decoupling them from the digest_worker pipeline and matching the UserMessageProcessor pattern.

Memory usage was significantly reduced by replacing the optimum/PyTorch dependency in doc2query_service with raw ONNX Runtime and implementing MemoryStore TTLs.

The persistent_task system was replaced by the goal_pursuit MessageProcessor subclass, simplifying background ACT looping and removing over 3400 lines of code.

Search embeddings are now pre-shipped as a static artifact in the database, eliminating a 3GB+ memory spike during startup.

Multiple fixes addressed memory stability, including disabling CPU memory arenas for the ONNX base model and deferring model loading to first use.

Capability health alerts are now published directly to the output:events WebSocket channel instead of routing through the LLM pipeline.

• Replaced broken DMN unified_generate path with DMNMessageProcessor.

• Implemented DB-backed context window with compaction to eliminate OOM in tool loop.

• Replaced persistent_task with goal_pursuit MessageProcessor.

• Reduced memory by dropping optimum/PyTorch dependency for doc2query_service.

• Pre-shipped search embeddings as a static artifact, removing runtime generation spike.

• Capability failure alerts now publish directly to output:events WebSocket channel.