Preprint measures coding repair failures caused by context compaction

Replacing coding agents' full histories with summaries damaged 5-64% of 22 successful repairs across 14 summarizer backbones in the Compact-Bench preprint. Context compaction lets agents continue beyond their context windows, but can erase information needed to finish a task. The benchmark compares full-history and summary-only continuations from the same repository state. It covers 35 SWE-bench repairs, 146 agentic sessions, and 106 production chains. Language-model summaries retained at most 16.0% of recurring details, compared with 79.1% for mechanical extraction. However, retaining details did not ensure successful continuation: with Kimi-K3 summarizing, the Kimi Code prompt preserved the most task continuations, while Claude Code retained the most recurring details.