Building AI-Powered GitHub Issue Triage with the Copilot SDK

When managing GitHub open source projects or active repositories, the task of triaging countless GitHub issues causes significant mental fatigue. This is because reading each issue's title and description, prioritizing it, and attaching labels involves constant context switching.

To solve this, we built IssueCrush, an app that leverages the Copilot SDK to process issues quickly via swiping. Users can swipe issue cards left or right to process them, and pressing the 'Get AI Summary' button lets Copilot summarize the issue content and suggest actions.

The key technical challenge was determining the execution environment for the Copilot SDK. Since the SDK depends on the Node.js runtime and the Copilot CLI process, running it directly inside a React Native app—which cannot execute Node.js directly—was not possible.

Therefore, we adopted a Server-side integration architecture as follows:

• Single shared SDK instance: All clients share a single instance on the server, minimizing overhead and authentication handshakes.
• Server-side security: Sensitive information such as API tokens is managed on the server rather than the client, enhancing security.
• Graceful degradation: Even if the AI service goes down, triage work can continue through a basic summary function.
• Logging and monitoring: Since all prompts and responses pass through the server, latency tracking and debugging are easier.