AI Assistant Updates: Watch Folders, Classification & Stability

Watched Folder functionality is now available with a dedicated service handling CRUD, server-side browsing, and a scan algorithm using mtime fast-skip.

A background worker thread manages folder scanning on a 30-second cycle, offering rate limiting and semantic environment tagging based on folder structure.

A new Vision LLM OCR service cascades through Gemini, Anthropic, and OpenAI before falling back to Tesseract for image-only PDFs and scanned documents.

The system now supports LLM-driven automatic document classification into categories, projects, and dates, with configurations managed via a JSON file.

The Moments to Documents pipeline migration is complete, storing pinned messages as documents while preserving API contracts and salience boosting logic.

A major overhaul to document processing concurrency fixed an infinite retry loop and replaced serial processing with a thread-safe queue using three concurrent workers.

Infrastructure saw significant stability improvements, including consistent UTC handling via a new time_utils module and fixes for database schema mismatches.

Ambient tool intelligence now proactively invokes external tools to bridge user drift and curiosity threads, with tools being ambient-eligible by default.

• Watched FolderService introduced with mtime fast-skip and rate limiting.

• LLM now drives automatic classification into category, project, and date pillars.

• Vision LLM OCR service implemented as a cascade fallback for documents.

• Concurrency fixed in watched folder processing by replacing serial queue with 3 concurrent workers.

• Moments are now migrated to be stored as documents in the DocumentService.

• New time_utils module ensures consistent UTC handling across core services.