FlowProof turns operational runbooks into executable browser assurance. A persistent Playwright worker replays seeded offboarding, refund, and access-policy workflows; captures step screenshots and traces; evaluates deterministic business invariants; and creates JSON, Markdown, and GitHub issue evidence.
A browser action can appear successful while the intended business operation still fails. The hosted demo injects repository permission drift during contractor offboarding: the UI action completes, but authoritative membership remains active. FlowProof fails the deterministic membership-removed invariant, preserves evidence, and requires a successful repair rerun.
GPT-5.6 has two bounded jobs: compiling natural-language SOPs into typed runbooks and diagnosing failures from structured evidence. Deterministic code remains the sole PASS/FAIL authority. The public judge path uses seeded GPT-5.6 output for reproducibility.
Codex accelerated product design, TypeScript implementation, test construction, debugging, deployment hardening, and hosted verification. Human decisions retained deterministic verdict authority, a zero-credential judge path, SQLite persistence, local evidence files, and strict separation between the Vercel dashboard and Railway execution plane.
Judge path: open https://flowproof-ten.vercel.app with no login, click Run failure demo, inspect membership-removed, screenshots, trace, and diagnosis, then click Repair & rerun and confirm PASS. Seeded applications affect no real accounts.
This is a factual submission draft. The project owner should personalize the final narrative before submitting.
Built With
- chromium
- codex
- gpt-5.6
- next.js
- node.js
- openai
- playwright
- prisma
- railway
- sqlite
- typescript
- vercel
- zod
Log in or sign up for Devpost to join the conversation.