Self-Refining Skill Library and Cooperative Cancellation

Extracted a SearchableAbility base class to replace procedural logic in ability modules. Built out a self-refining skill library with a Layer 1 search ability using vector and FTS5 RRF fusion, alongside a Layer 2 association service that maps behavioral patterns to skills after each processing pass. Fixed tier classification to ensure the skill finder remains permanently available and hardened its parsing logic.

Unlocked behavioral patterns from the memory recall graph so they consistently surface during relevant turns. Added rendering utilities to convert stored JSON blobs into compact prompts for the LLM. Also fixed a reinforcement bug where rediscovered patterns failed to update evidence count, storage strength, and retrieval weight columns in the database.

Replaced hard-coded iteration caps on user-facing processors with cooperative cancellation via threading.Event. Exposed REST endpoints for stopping active sessions and wired frontend controls plus WebSocket disconnect resets to prevent stuck states. Fixed a tool-call dispatch issue where sub_id was being lost by switching to in-place dictionary mutation and routing it through the _params_update pipeline.

Added a self-reference tool that maps queries directly to live documentation URLs without hardcoding knowledge. Updated the installer to reliably fetch fresh sources during upgrades and bumped the sqlite-vec dependency floor to prevent version regressions. Extended the Brain UI to support model listing for OpenAI-compatible providers and cleaned up phase 4 invariants to match the centralized tool tier spec.

• Extracted SearchableAbility base class and rebuilt skill playbooks as direct tool-calling sequences

• Implemented Layer 1 vector search and Layer 2 association service for automatic pattern-to-skill mapping

• Unlocked behavioral patterns in memory recall and fixed reinforcement columns for evidence-based storage strength boosts

• Replaced hard iteration caps with cooperative cancellation via threading.Event and new stop endpoints

• Fixed sub_id persistence through the tool dispatch pipeline by enforcing in-place parameter mutation

• Bumped sqlite-vec dependency floor and corrected installer upgrade flow to always fetch fresh sources