continuous-learning-v2

Capture Claude Code session activity and tool usage by enabling deterministic PreToolUse and PostToolUse hooks. These hooks reliably observe every tool call and prompt, providing the raw session data needed to detect repeated workflows and error-resolution patterns within git repositories.

Behavioral instincts are atomic YAML rules generated from session analysis that encode detected corrections, tool preferences, and repeated workflows. Each instinct includes triggers, supporting evidence, and an evolving confidence score, defaulting to project scope with metadata supporting clustering and evolution.

Automatically learn project conventions by running a background observer that analyzes recent session observations. It detects repeated behaviors and error-resolution patterns, then generates project-scoped, confidence-weighted YAML instincts so conventions are operationalized without manual re-encoding.

Yes, project-scoped instincts can be promoted to global scope. When high-confidence behavioral patterns are observed recurring across multiple projects, the governance mechanism supports promoting them, alongside clustering related instincts into reusable skills, commands, or agents.

Session observation requires Claude Code sessions within git repositories to detect project-scoped patterns. The hook-based observation system captures tool calls and prompts reliably, and the generated instincts include scope metadata to distinguish between project-level and global-level conventions.

Export and import behavioral instincts using the built-in import/export pattern support. This allows confidence-weighted YAML instincts, including their triggers and evidence, to be transferred between projects, facilitating the evolution of recurring high-confidence patterns into reusable skills.