ContextStrategies

Challenge: Prepare massive codebases for LLM consumption—raw repos overwhelm context windows with noise (node_modules, build artifacts, redundant code). Approach: Built comprehensive analysis toolkit with Clean Architecture—Domain layer handles scanning/flattening with gitignore intelligence, Application layer orchestrates processing pipelines, Infrastructure abstracts AI clients and storage. Content cleaning pipeline strips comments, normalizes whitespace, deduplicates code blocks. AI-powered scope selection uses pydantic_ai to intelligently focus on relevant files for specific queries. AST-based patching workflow extracts target scopes, generates modifications via LLM, applies changes safely. Multi-faceted analysis includes architecture validation (layer violations), code smell detection, cyclomatic complexity metrics, duplicate code identification, security pattern scanning, and dependency mapping. Golden testing framework captures AI responses for regression validation. Interfaces: Unified Typer CLI with server fallback, Textual TUI with split panes, Streamlit dashboard for history visualization, FastAPI HTTP server for remote access. DuckDB analytics backend. Innovation: BaseAnalyzer with caching prevents redundant processing; unified interface abstracts deployment modes (local/server/TUI).