Cost-efficient pinpointing of target code contexts in complex repositories

Develop methods that cost‑efficiently pinpoint the target code contexts required to answer diverse natural language queries in complex software repositories, achieving reliable localization of relevant files, classes, and functions while minimizing token and computational overhead during repository‑level code reasoning.

Background

Repository-scale code reasoning tasks require locating small, relevant regions of large codebases to support multi-hop reasoning. Existing retrieval and graph-augmented methods improve over flat text chunking but still incur significant cost due to context size and structural complexity.

The paper highlights that, despite progress, efficiently identifying only the necessary code elements across varied query types remains unresolved, motivating approaches like FastCode that decouple exploration from full-content consumption.

References

Despite these advancements, it remains an open challenge to cost-efficiently pinpoint target contexts in complex repositories across diverse query scenarios.

— FastCode: Fast and Cost-Efficient Code Understanding and Reasoning  (2603.01012 - Li et al., 1 Mar 2026) in Related Work, Retrieval and Graph-Augmented Code Reasoning

Agents that can re-open files may recover some of the losses we measure, because a compressed view tells them which file to read in full. Each re-read, however, costs a turn and the tokens that compression was meant to save, and a model may not notice that the text it sees is altered: in Figure~\ref{fig:main}, it rewrote a line that does not exist in the file. Measuring this trade-off needs a full agent loop, which we leave to future work.

— Skip the Compressor: Coding Agents Fix More Issues When They Pick Whole Files  (2610.21073 - Yang et al., 15 Oct 2026) in Discussion, paragraph “Re-read on demand or compress once?”