Feature Overhaul: Tool Discovery, DMN, and Resilience

Two-axis tool discovery is implemented, scoring tools using both semantic k-NN distance and exact keyword matches, which replaces the previous reliance on full_profile embedding.

Built-in tools (search, weather, news, code_eval, programming_docs_search, browser) now use hardcoded profiles seeded at startup, bypassing the LLM profiler.

The reasoning loop and cognitive drift engine are replaced by a timer-based DMN service, triggering proactive LLM calls after idle time or based on high-salience memories.

Auth sessions are made durable by persisting them to SQLite alongside MemoryStore, ensuring logins survive container restarts.

A major refactor updates the search tool description to enforce plain natural language queries, removing guidance that taught operator syntax.

The package name for duckduckgo-search was updated to ddgs upstream to resolve silent zero-result returns.

Schema validation for skills was tightened, specifically restoring the ‘array’ type for items to prevent JSON string serialization issues from LLMs.

Workflow automation was added to support auto-release via n8n and subsequent syncing of release notes to the website on publish.

• Implements 2-axis tool discovery: semantic k-NN + keyword matching.

• Replaces reasoning loop with a timer-based DMN service.

• Makes auth sessions durable by persisting them to SQLite.

• Hardcodes profiles for built-in tools, bypassing LLM profiler.

• Updates search tool description to enforce natural language queries.

• Adds auto-release workflow via n8n upon tag creation.