Large code repositories often exceed the context window of a coding agent, forcing reliance on prompt compressors or retrievers to fit relevant information into the model. A new arXiv preprint argues that these workarounds may be unnecessary.

Instead, the paper reports that coding agents fix more issues when they pick whole files rather than compressed or retrieved snippets. The abstract notes that patching an issue requires two things at once, though the text cuts off before listing them. The title's conclusion is clear: skipping the compressor and selecting entire files leads to better outcomes.

This result suggests that preserving full file context can be more valuable than reducing input size. For practitioners, it points to a simpler design choice: give the agent the whole file and let it decide what matters.