A Salesforce research team built a framework that improves how AI agents find code by rewriting the repository's own navigation files, not the agent itself.
The system, called Libra, plants hierarchical Markdown files called catalogs into a code repository. An LLM then generates synthetic search queries, a second model tries to locate the right code by following those catalogs, and a third model rewrites the catalogs wherever the search fails. The loop repeats until the catalogs get accurate enough to stop misleading the agent. Tested across 12 repositories from the SWE-bench Lite benchmark, the researchers report continual accuracy gains that follow a logarithmic curve - meaning early rounds improve things fast, with diminishing but still positive returns over time. The team also showed that a minimal coding agent using Libra-optimized catalogs beat existing state-of-the-art baselines.
Most AI coding research focuses on making the model smarter; Libra bets that the environment the model operates in is just as important. The more striking result is transferability: catalogs optimized for one LLM improved localization accuracy for different models without any retraining, suggesting the improvements are structural rather than model-specific.
Salesforce is not the only lab chasing better code retrieval - retrieval-augmented generation over large codebases is a known weak point for every major coding assistant - but packaging the fix as a self-healing index layer rather than a bigger model is a notably cheaper angle to pursue.