what is codescalebench fr?

CodeScaleBench is a benchmark that measures how AI coding assistants perform when they can search large, enterprise-scale codebases through external retrieval tools rather than browsing files locally. The core idea is to isolate the impact of code search: the same AI agent runs each task twice, once with the code available locally and once with it accessible only via a retrieval server, and the difference between the two runs becomes the key result. The project includes 370 realistic tasks drawn from more than 40 repositories across nine languages, covering work like cross-repo navigation, refactoring, security review, debugging, and feature implementation. Tasks are organized by developer work type, and indexes let you slice the data by complexity, language, repository size, or whether a task spans multiple repositories. Each task is scored by deterministic verifiers, and the results are published as auditable snapshots with per-task traces, timing, cost, and reward scores. Who would use this? Researchers and engineers evaluating AI coding tools, especially those building or comparing retrieval-augmented agents. If you're deciding whether your assistant benefits from a Sourcegraph-style search server versus a local-filesystem baseline, or you want a rigorous, reproducible way to benchmark different models and configurations, this gives you a ready-made evaluation kit. The included snapshots already show, for example, how a Haiku 4.5 model performed under baseline versus Sourcegraph MCP conditions across the full task set. Notable design choices include the calibration triad, a quality gate that ensures each task's verifier can distinguish a correct answer from an empty or keyword-stuffed one before it's published. The repo also provides a side-by-side browser so you can visually compare, for every task where retrieval tooling helped, exactly what the agent did differently on both runs: search queries, file reads, edits, and bash commands laid out in two columns. Full agent trajectories are available as compressed archives in GitHub Releases for deeper analysis.