Exhibit

AI coding tools make it easier to produce code. Maintainers still have to answer a harder question: what actually works? Exhibit turns AI-authored pull requests into evidence a human can inspect.

An opt-in GitHub Actions workflow runs browser and backend API tests, captures frontend recordings, and collects screenshots, traces, results, and the exact tested commit. Codex reviews the structured results and recording metadata to highlight coverage gaps and suggest further checks. Exhibit posts the evidence back to the pull request, with animated previews and links to original artifacts.

Our working demonstration covers an express-checkout change: four browser checks and nine backend API checks passed, with recordings and an advisory Codex report published directly on the PR. The review workspace provides video playback, slow motion, frame stepping, API/cURL inspection, artifact downloads, and source provenance.

Built with Codex, React, TypeScript, Go, PostgreSQL, Playwright, and GitHub Actions, Exhibit fits the Developer Tools track. It brings test outcomes, observed behavior, and AI-assisted insights into one review workflow—while keeping the final decision with the developer.

The prototype demonstrates the complete GitHub Actions loop on our demonstration repository. The dashboard currently displays archived captures; automatic live ingestion is the next step.