Cost-Aware Python Code Repair via Budget-Threshold LLM Optimization

Repeated test feedback can increase token consumption during large language model (LLM) code repair, particularly for difficult tasks. This study examines whether fixed feedback compression and observable progress signals can control this overhead without changing model parameters. The cost_aware strategy combines feedback filtering, budget-threshold and no-progress stopping, reset prompts, and a final rewrite opportunity. Five strategies were evaluated on 30 QuixBugs-derived Python tasks over three independent API runs. Public tests guided repair, whereas hidden tests were reserved for final evaluation. cost_aware fully passed 27 tasks in every run, averaging 90.0%, and consumed 47.5% fewer tokens than full_loop. Maximum per-task usage dropped substantially from 29,498 to 4,185 tokens, demonstrating effective tail suppression; while conservative multiplicity-adjusted inference reflects the limited sample size (N=30), the observed effect sizes highlight substantial practical utility. A single exploratory DeepSeek run showed a smaller reduction in the same direction. These observations warrant further evaluation of transparent invocation control, with conclusions limited to the tested tasks, models, and sampling conditions.