A development team recently conducted an experiment: feeding a 180,000-line Spring Boot monorepo into Claude Code for a full architecture analysis. Initially, the results were impressive—Claude produced a dependency graph more detailed than their tech lead's and even uncovered three circular dependencies they had missed. However, before reaching that point, the team spent six weeks troubleshooting a silent failure mode. They discovered that when the number of tools in the codebase exceeded 50, Claude Code would silently drop some tools without any error or warning. This meant that the AI was working with an incomplete set of capabilities, potentially missing critical parts of the codebase. The team's experience highlights a significant, undocumented limitation in current AI coding assistants. For engineering teams adopting tools like Claude Code for large-scale refactoring or analysis, this finding is crucial: always verify that the AI has full context, and be aware of hidden constraints that could compromise results. The post serves as a practical warning for the industry.
A team feeding a 180K-line Spring Boot codebase to Claude Code discovered that when the number of tools exceeds 50, the AI silently drops some without warning. This undocumented limitation can lead to incomplete analysis and false confidence. The finding is critical for any team relying on AI coding assistants for large-scale code understanding.