what is codex-speedup-proof fr?

Codex Speedup Proof is a skill add-on for OpenAI's Codex coding assistant that tries to make performance optimization claims trustworthy instead of taking them on faith. Rather than looking at code and declaring it faster because the structure looks better, the skill finds a likely performance hotspot, proposes exactly one focused change to fix it, and then measures the same realistic workload both before and after the change to see whether it actually helped. The result is always one of three honest verdicts: proven, inconclusive, or regression, never a made up percentage improvement. Before touching any files, the skill has to show the user a full proposal: which files it plans to change, the exact transformation it intends to make, what behavior needs to stay identical, the benchmark and test plan, and the main risk involved. It then stops and waits for explicit approval, and that approval only covers the one proposal shown, so a different optimization needs a fresh approval. Once approved, it applies the single change, preserving things like outputs, ordering, error behavior, APIs, and permissions, then runs the project's existing tests, typechecks, and build process. It rejects changes that turn out to be noisy in their measurements, actually slower, or that break existing behavior. A result only counts as proven when correctness checks pass, the same representative workload is used before and after, the measurements are stable, and the improvement clearly exceeds normal benchmarking noise. The skill also supports an audit mode that ranks performance opportunities without changing any code, and a regression mode for comparing a branch against main to find out what got slower. Installation is a single command using npx, which places the skill into a Codex skills folder, or it can be installed manually by cloning the repository and copying the folder in directly. After installing, Codex needs to be restarted. Output is written as a Markdown report that renders on GitHub plus a machine readable JSON evidence file, so the result can be reproduced. The project is released under the MIT license.