Claude Code context limits are a session-wide memory budget, not just prompt length, including history and file reads. In large monorepos, token burn is quick as big files and logs exhaust tokens, so local reads miss dependencies. The article argues architectural context—dependencies and call paths—
AI-assisted coding is only safe if the AI can see the codebase architecture, not just edit files. A code graph MCP server provides a code graph, an MCP interface, and safety via impact analysis before edits. It champions relationship-first workflows—tracing callers, dependencies, routes, tests, and
Code-change blast radius—not file count—drives AI-edit risk, as small tweaks can ripple through callers and deployments. Identify changed symbols, trace callers (direct and transitive), assess failure modes, and rate reversibility to shape tests and rollout. AI often creates plausible code that miss
AI helps speed changes but often misses cross-cutting effects; software impact analysis is essential to assess blast radius, contracts, and dead code before merging. The piece lays out seven questions and a practical workflow to map dependencies, interfaces, tests, owners, and failure modes, plus a
AI edits in multi-service codebases are risky if you judge them by diffs alone; you must map upstream and downstream impact and blast radius across contracts, schemas, and deployment. A useful service dependency graph is artifact-driven, multi-layered (service calls, packages, symbols, schemas, infr
AI coding is risky in real repos because models miss architecture. A codebase knowledge graph records entities (files, modules, functions, endpoints) and relations (calls, imports, reachability) to gauge blast radius before edits. Build it with AST parsing, relationship extraction, incremental updat
Dead code isn’t just what grep misses; it’s code that no longer affects behavior, given reachability, side effects, and who depends on it. The article classifies dead code into unreachable paths, unused definitions, dead variables, and stale comments, and warns that deletion isn’t safe if side effec
AI often writes new helpers that already exist because it only sees a slice of the codebase, especially in multi-module repos. Duplicates can be exact, near, or semantic; semantic duplicates are the riskiest since the behavior is the same but the implementation differs. The article proposes a pre-ed
MCP tools don’t magically smarten Claude Code; they give it enough system context to safely understand dependencies and blast radius when editing code. The article advocates a lean, staged stack: start with codebase-context tools (like Pharaoh) and GitHub, then add databases, UI checks, or incident
In a monorepo, AI edits feel risky because file proximity hides dependencies; a graph of inbound usage and symbol-level links makes blast radius visible before changes. The article favors layered graphs—package builds plus deeper file/symbol maps for safe refactors—static and architecture-focused. P
New posts on codebase intelligence, AI coding, and building Pharaoh.