AI Autopilot: A Closed-loop C++ Development Pipeline from Ticket to GUI Verification

The author created an 'AI Autopilot' for a desktop geophysical application (about 3 million lines of C/C++, Qt widgets, QML scenes, and Qt3D). The system ingests a ticket from YouTrack, where an 'Issue Understand' component reads the description, attachments, SVN history, and video frames using ffprobe, extracting context. Then 'Solver' uses Claude Code to plan and write C++ code changes, followed by a build. The 'Test Engine' (nicknamed 'Uprovlala') allows the agent to control the running application via a named pipe from a Python client, using methods like button clicks, tree navigation, and 3D view capture, even in a minimized RDP session. The agent can also record and replay macros. After GUI testing, a review component screenshots are checked. Finally, the ticket is committed and closed with a reply. The system grew from 2,899 lines and 37 C++ methods in May to 13,836 lines and 120 methods by July, with each method added to overcome real task blockers. If the agent cannot perform a step, it honestly reports 'can't do' rather than inventing false results.