<![CDATA[Repository-scale software agents need constrained context before reasoning about a defect. This study evaluates retrieval and triage on all 500 human-validated SWE-bench Verified tasks from 12 Python repositories. Unified-diff parsing recovered 373 modified source paths and 1,220 source hunks. In the primary issue-only closed-world experiment, a leave-one-repository-out reranker achieved 0.500 Hit@1, 0.747 Recall@5, and 0.625 MRR; its MRR gain over character TF-IDF was 0.065 (95% CI 0.046-0.087; adjusted p < 0.001) and remained significant with equal repository weighting. Failing-test identifiers increased post-failure RRF MRR from 0.590 to 0.685, while issue-only word TF-IDF reached 0.684 MRR on gold-selected pre-fix hunks. Nested cross-repository calibration yielded 0.139 PR-AUC, 0.632 ROC-AUC, 0.082 Brier score, and 0.028 ECE for high-effort triage. Lexical localization was effective, but pre-execution effort forecasting remained limited; retrieval evidence, escalation, and executable repair should therefore remain separate stages.]]>
Copyrights © 2026
