TUI audit finds only 12% of tests touch the interface

A survey of 197 terminal-interface applications found only 12% of test code exercises the interface, and 45% of those tests never send input. The accompanying headless benchmark compares four frontier models with random exploration and finds no model dominates: random testing discovers more crashes through throughput, while LLM guidance is more efficient per interaction and uniquely reaches input-gated faults. For teams building developer agents, the result complements the finding that browser-agent latency is often a queue problem: invest first in launch automation, deterministic instrumentation, and honest baselines, then choose the model.