GitHub Copilot CLI + GPT-5-mini BYOK: The Code Was Cheap, the Quality Gates Were Expensive

This is part 3 of a series testing GitHub Copilot CLI with different model setups. The author moved from local models to Azure OpenAI's GPT-5-mini through BYOK, building a Windows tray app (ElBruno.NetAgent) for safe, dry-run network selection. While GPT-5-mini generated code efficiently, the main challenge was defining and enforcing quality gates beyond just passing builds and tests—including runtime startup, dependency injection, WPF/XAML loading, tray lifecycle, fast deterministic tests, and manual UX validation. The article emphasizes that code generation is cheap, but ensuring real-world reliability is expensive.