When a repo exceeds the context window, spending the budget on choosing which files to show verbatim resolves more issues than prompt compression or chunk retrieval. That shapes how you build context for coding agents.
Key takeaways · AI-distilled
The authors argue a patch needs both the files that must change and their exact text, because the model has to reproduce the lines it edits. Retrievers rank code by similarity to the issue, not by whether it must change.
The whole-file selector, CQ2-CARE, also beat two methods trained for issue resolution: SweRank by 0.7 and 1.7 points and SWE-Pruner by 2.0 and 7.3 points, with GPT-5.6 and Claude Opus 5 respectively.
In the authors' analysis, retrievers often miss the files that need editing, while compressors keep those files but alter their text, so fewer of the resulting patches apply.
Every method's context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → was fit to the same per-issue token budgetA cap on how many tokens a task, session, or agent run may consume — the practical control on both cost and how long an agent will grind.Full definition → and each patch was graded with the benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition →'s own tests. Exact-file selection also needs no separate compression model.
Terms in this piece · Glossary
token budget — A cap on how many tokens a task, session, or agent run may consume — the practical control on both cost and how long an agent will grind.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
articleNot All Agents Are Equal: Code Quality and Post-Merge Maintenance Across Five Autonomous Coding Agents in the WildObada Kraishan
articleRefine After Generation: Toward Correct and Concise Patches in LLM-based Program RepairWenqiang Luo, Jacky Keung, Xiaoyu Shi, Yicheng Sun, Boyang Yang, Zhou Yang, Haoye Tian
articleGroundability, Not Scale Alone: When Weak Reviewers Can Audit Strong Coding AgentsJunyu Guo, Shangding Gu, Ming Jin, Javad Lavaei