Interface-Rooted Repo Maps for Token-Efficient Coding Agents: A Negative Result

Abstract

\begin{abstract} Coding agents consume millions of tokens per task exploring repository context, motivating research into context efficiency. We test whether task-conditioned repository maps can reduce this overhead. We propose IR-RepoMap, which extracts interface file paths from FeatureBench problem statements, builds a bounded import closure via BFS over the Python import graph, and prepends function/class signatures to the agent prompt. We hypothesized this focused context would reduce token consumption by at least 25%. Our experiments on FeatureBench-Lite Level 1 (26 tasks) with OpenHands and DeepSeek-V3 refute this hypothesis: on 13 comparable tasks, IR-RepoMap uses 20.8% \emph{more} tokens than baseline, with extreme per-task variance (std=274.6%, range $-$89% to $+$889%). This negative result demonstrates that static upfront context provision does not reliably reduce token consumption for coding agents, suggesting that dynamic, agent-driven context retrieval approaches warrant further investigation. \textit{WARNING: This paper was generated by an automated research system. The code is publicly available.}\footnote{\url{https://gitlab.com/fars-a/featurebench-repomap-token-budget}} \end{abstract}