Context Guardian: helping coding agents understand the impact of a change

A team at Careo Limited built Context Guardian, an open-source tool that pairs a command-line utility with an agent skill to help coding agents understand the downstream impact of a code change. The tool builds a map of the codebase and combines it with project configuration in a `.app-context/` directory to identify affected features, flag matching iOS and Android files, and point agents to relevant documentation and validation commands. The Python analysis engine runs locally with no model calls or telemetry, integrates with Claude Code and Gemini CLI via hooks and with Codex, Cursor and GitHub Copilot through `AGENTS.md`, and is currently in alpha with import and call resolution for TypeScript, JavaScript and Python, developed around React Native and Expo apps.

A coding agent updates a shared module and completes the task. The edit looks reasonable, but another feature depends on that module. Does that feature need checking too? We built Context Guardian to make those connections easier to identify during agent-assisted development. It combines a command-line tool with an agent skill to connect code changes with feature documentation, project rules and validation commands. Consider a checkout module that is also imported by a profile screen. A change to that module can require attention in both places. Context Guardian builds a map of the code and uses it alongside project configuration to select the relevant features. Its report points the agent to their notes and configured checks. For mobile projects, it can also flag corresponding iOS and Android files that may need a matching update. Before a task, Context Guardian selects the context relevant to the work. When the task finishes, it compares the changes against the application’s map and reports areas that need attention. Claude Code and Gemini CLI use hooks for automatic checks. Codex, Cursor and GitHub Copilot can follow the skill and instructions in AGENTS.md . The analysis engine runs locally in Python, with no model calls or telemetry. It produces compact reports for the agent to read. Feature descriptions, project rules and validation commands live in .app-context/ , alongside a generated code map. Teams maintain the curated documentation, while Context Guardian regenerates the map from the code. This gives the analysis both code relationships and application-specific context. Context Guardian is in alpha. We developed it around React Native and Expo applications, with import and call resolution for TypeScript, JavaScript and Python. Other languages receive lighter, pattern-based analysis. Its usefulness depends on the project configuration and the relationships it can resolve. Tests and code review remain part of the workflow. Visit the repository for installation and setup https://github.com/CareoLimited/context-guardian We’d welcome feedback from developers trying it on existing applications, particularly where setup is unclear, a dependency is missed or a report includes too much context.