Skip the Compressor: Coding Agents Fix More Issues When They Pick Whole Files

Skip the Compressor. Coding Agents Fix More Issues When They Pick Whole Files. By Xiaocong Yang and four others, from the University of Illinois Urbana-Champaign. Published on the fifteenth of October. Its code repository is worked out: about forty source files, with a readme and a licence.

A coding agent often works on a repository far bigger than its context window can hold, so something has to decide what the model gets to see. Prompt compressors and retrievers are the usual answers. The authors argue that patching an issue needs two things at the same time: the files that must change, and their code exactly as written, because the model has to reproduce the lines it edits. Retrievers rank code by how similar it is to the issue, not by whether it needs to change. Compressors delete text to fit the budget.

The claim follows from that. At an equal token budget, an agent resolves more issues when the budget is spent choosing which whole files to show verbatim than when it is spent pruning tokens or retrieving chunks. The authors' simple selector for doing this is called CQ2-CARE.

The test uses three hundred recent SWE-rebench issues. Every context is fitted to the same per-issue budget, and each patch is graded with the benchmark's own tests. With GPT-5.6, CQ2-CARE resolves forty-three point seven percent of the issues. With Claude Opus 5, it resolves forty-six point seven percent. With either model, that is six point three points more than the best of three compressors.

Against the best of three retrievers, the margins are wider: nine point three points with GPT-5.6, and twelve point seven points with Claude Opus 5. CQ2-CARE also resolves more issues than two methods trained specifically for issue resolution. It beats SweRank by zero point seven points with the first model and one point seven with the second, and SWE-Pruner by two points with the first and seven point three with the second.

The analysis locates where the baselines go wrong. Retrievers often miss the files that need editing in the first place. Compressors do keep those files, but they alter the text, so fewer of the resulting patches apply.

The setting is specific. The evidence comes from three hundred issues on one benchmark, two models, and an equal token budget per issue, with the comparison set made of three compressors, three retrievers, and two trained methods. Within that setting, the result is consistent across both models. The abstract does not report other budgets or other benchmarks, so how far the gap holds beyond this setup is left open here.

Fidelity of the text turns out to matter as much as relevance. A patch has to quote the code it changes, and exact-file selection shows every line as the repository has it, with no compression model involved at all. Under a fixed budget, spending tokens on whole files kept intact beat spending them on text that had been pruned, or on chunks ranked by similarity.