Context-aware code generation needs more than the current file: four layers of context—local, repository, type/dependency, and architecture—guide changes and reduce blast radius. In real repos, file-level completions often duplicate logic or misuse abstractions, causing downstream breaks. A practica
Architecture analysis for AI tools should reveal blast radius, dead code, and how modules connect, not just pretty diagrams. The piece compares Pharaoh (codebase intelligence via a knowledge graph for active editing), ArchAgent (architecture recovery for legacy drift), and SMART (high-level system r
AI coding tools often make risky edits because they operate on a thin slice of code and miss cross-cutting dependencies. MCP developer tools should provide architecture context via a code/knowledge graph that shows dependencies, blast radius, reused patterns, and dead code. The current landscape mix
The article argues that in large codebases a software architecture graph should reveal dependencies, blast radius, test coverage, and dead code before changes. It urges moving beyond import graphs to include build, test, runtime, and ownership edges with provenance. It advocates layered views (file,
AI-assisted refactors fail when the blast radius isn’t visible in the diff, especially in multi-module systems. The article advocates pre-edit impact analysis: define the seed change, enumerate direct and transitive dependents, contracts, and runtime paths, and decide sequencing. It argues that diff
Refactoring with Claude Code in monorepos is risky because changes can affect unseen callers and distant dependencies. Safer edits come from richer repo context, not smarter prompts—layered in-session and persistent rules, per-directory CLAUDE.md, live discovery, and a repo map to reveal blast radiu
Safer AI-assisted coding comes from a code knowledge graph that encodes real code relationships—files, modules, calls, tests, and contracts—over prompts. It answers who calls what, what depends on it, and which tests are affected, reducing the blast radius of edits. Static analysis is the floor; dyn
AI coding in multi-repo systems is fast but fragile without architecture intelligence—context about dependencies and blast radius must precede edits. A graph-based architecture map shows ownership, connections, and contracts, enabling quick impact questions like what breaks and who’s affected. Use a
Eight AI coding assistants are evaluated as alternatives to Sourcegraph Cody for small teams, focusing on real workflows over flashy demos. They fall into AI-native editors (Cursor, Windsurf), broad IDE assistants (Copilot), terminal-first (Aider), and self-hosted options (Tabnine, Tabby), plus JetB
Large monorepos make Cursor edits risky: a rename can pass local tests yet break a downstream package that shared a hidden type. The article says map context before editing—check shared libs, indirect consumers, and blast radius to ship cleanly. Cursor excels at exact and semantic search with backgr
New posts on codebase intelligence, AI coding, and building Pharaoh.