Survey of 197 Real TUI Apps: Only 12% of Test Code Touches the Interface, and 45% of Those Tests Never Send Input

arXiv 2608.03743·medium signalXBlueskyLinkedInCopy link

Researchers packaged 197 terminal UI applications spanning ratatui/Rust, bubbletea/Go, textual/Python and ink/TypeScript as instrumented Docker images, then pitted four frontier LLMs against random exploration under equal wall-clock budgets. No model dominates — random is a strong time-budgeted baseline whose crash advantage comes from raw throughput, while per-interaction LLM guidance is more efficient and uniquely reaches input-gated faults. Line coverage poorly predicts crash discovery, undercutting it as a proxy for test effectiveness; the coverage tool tuicov and framework tuibot are released on GitHub.