ReproPilot — Codex Bug Reproduction Lab

Inspiration

AI-assisted development can produce a patch quickly, but reproducing a bug is often the slowest and least reliable part of the workflow. Developers still spend time rebuilding environments, collecting logs, rerunning commands, and explaining what happened. ReproPilot is designed to make that process isolated, repeatable, and easy to review.

What it does

The current prototype provides a browser-accessible cloud workspace backed by isolated E2B sandboxes. A user can create a temporary development environment, run shell commands, manage files, inspect processes, restore an active workspace, and use an AI-assisted terminal from one interface.

For OpenAI Build Week, I am extending this foundation into ReproPilot: a workflow that accepts a repository or error report, uses GPT-5.6 to create a structured reproduction plan, and uses Codex to run the plan, create a regression test, propose a patch, and produce a replayable evidence report.

How I built it

The application uses TypeScript, React, TanStack Start, Tailwind CSS, and the Vercel AI SDK. E2B provides disposable Linux and Chromium environments. Supabase supports application data and server-side integrations, while the web interface manages sandbox lifecycle, command execution, file transfer, runtime presets, and session recovery.

Build Week work

This project existed before the submission period. During Build Week I am using Codex to add the GPT-5.6 planning and bug-reproduction workflow, automated verification, clearer diagnostics, and an exportable timeline. Dated Git commits and Codex session evidence will distinguish the new work from the original CloudLab prototype.

Challenges

The hardest parts have been safely managing short-lived sandboxes, handling command timeouts and non-interactive login flows, preserving useful logs after failures, and keeping the browser UI synchronized with remote runtime state.

Accomplishments

The prototype already provisions isolated environments, exposes a working browser terminal and file manager, supports multiple runtime presets, restores active sessions, and provides a public demo that can be tested without local setup.

What I learned

Reliable agentic development needs more than a model response. It needs observable execution, explicit state, bounded environments, reproducible commands, and verification artifacts that another developer can inspect.

What's next

Next I will complete the GPT-5.6 planner, Codex execution loop, regression-test generation, patch review, and a downloadable report containing commands, logs, test results, and diffs.

Built With

• codex
• e2b
• gpt-5.6
• lovable
• openai-api
• react
• supabase
• tailwind-css
• tanstack-start
• typescript
• vercel-ai-sdk