continuous-learning-v2

The skill captures atomic learned behaviors by hooking into PreToolUse and PostToolUse events to detect user corrections and repeated tool sequences. It uses background Haiku analysis for pattern detection and stores results as confidence-weighted instincts.

To prevent cross-project pattern contamination, store learned coding instincts using project-scoped isolation via git remote URL hashing. This ensures React patterns stay in React projects and Python conventions stay in Python projects without bleeding across contexts.

Yes, this approach uses PreToolUse and PostToolUse hooks to capture 100% of tool use events for pattern detection. Background Haiku analysis then processes these events to generate confidence-scored atomic instincts tagged by domain.

Yes, you can cluster learned instincts into reusable skills, slash commands, or background agents using the /evolve command. This evolution pipeline groups related atomic behaviors into full skills for broader application across sessions.

High-confidence coding patterns are automatically promoted from project scope to global scope when they appear in 2 or more projects. This cross-project promotion ensures universally applicable patterns are available across all development environments.

Pattern detection supports instinct tagging across domains including code-style, testing, git, and workflow. Each atomic instinct receives a confidence score between 0.3 and 0.9 with full evidence tracking for reliable pattern application.