Most AI coding tools look at the file you changed. I wanted to know what that change could affect across the repository. So I’ve been building GitHub Autopilot — an open-source AI engineering agent that can review PRs, fix bugs, scan for security issues, manage GitHub workflows, and now build an AST-based dependency graph of the codebase. The graph can identify: - Module dependencies - Fan-in / fan-out - Import cycles - Orphan modules - Dependency hotspots - Runtime vs top-level imports It actually found a 5-module import cycle in my own repository. I fixed it and added CI checks to catch regressions. The project is open source and self-hostable, with local LLM support for privacy-sensitive code. GitHub: https://github.com/Shweta-Mishra-ai/github-autopilot What repository-level context do you think AI coding agents are still missing?

It’s a GitHub App that listens to webhook events and helps automate common repository tasks like PR review, issue triage, and repo health checks. I tried to focus more on architecture reliability — guardrails, async webhook handling, and validation of AI responses. Would love feedback from other developers on the approach. Repo: https://github.com/Shweta-Mishra-ai/github-autopilot