continuous-learning-v2

You can automatically extract reusable patterns by enabling observation hooks that capture PreToolUse and PostToolUse events. An observer agent then analyzes these session activities to generate confidence-scored atomic instincts.

Yes, you can scope learned behaviors to specific repositories. The tool detects project context and stores instincts per-repo under the ~/.claude/homunculus directory, preventing cross-project contamination of conventions.

You can promote project-specific instincts to global configurations using built-in CLI commands. Once patterns are proven and reach sufficient confidence scoring, the promotion mechanism elevates them from project scope to global scope.

Yes, you need Python, Git, and Claude installed to automate session learning. These dependencies are required to run the observer scripts and manage the project-scoped detection of repository contexts.

The best way to manage session instincts is using the provided CLI commands. You can use status to view instincts, evolve to cluster them into skills, and import or export the learned behaviors for governance.

A limitation of continuous learning hooks is that instincts rely on conservative promotion rules and evidence-backed scoring. This means patterns must accumulate sufficient confidence before being applied, preventing immediate automated reuse.