Case study

rauchg.com: rebuilding an RSC blog from captured Flight output

Documents a website-rebuild-skill case that decodes browser-delivered React Server Component Flight streams, reconstructs a buildable Next project and compares its output with a captured site. The reported rauchg.com campaign uses route-tree comparison and module-reference mapping; the new server source is inferred from observable output.

Requirements and limitations

The documented rauchg.com snapshot used Next 16.1.1, Turbopack and React 19 canary. The tooling requires Node.js 22+ and a local Chrome/Chromium for comparison; some stages invoke version-pinned external tools. Repeating the method needs an authorized target snapshot and a separately built reconstruction, not just the skill repository.
Flight decoding exposes delivered element trees, route state and client-module references. The reconstructed server-component source is explicitly an inference from that output; flight-to-mdx also needs site-specific shape/link adaptations. This is not extraction of undistributed server source or automatic recovery of an arbitrary Next.js application.
The author reports 18/18 routes passing a Flight semantic gate and a runtime sweep for this case. The gate compares normalized trees and enforces a global module-ID bijection. It normalizes build-hash namespaces and selected encoding differences; configured prop/class exclusions can omit registered fields or subtrees. A pass is not byte identity or full behavioral equivalence.
The gate does not check CSS, so layout, fonts and rendered pixels need separate comparisons. Public synthetic self-tests include changed-text rejection and rejection of a one-to-many module mapping. This review inspected those checks as source and did not run them or reproduce the reported website campaign.
The method explicitly leaves unobserved error responses, database/write paths, server caches, dynamic-image generator source and unlinked routes unresolved. Historical public-source comparisons and reported reconstruction percentages do not prove access to hidden backend logic; the author also notes possible model-training familiarity with a nominally blind target.
MIT applies to the skill and tooling, not to reconstructed websites, their scripts, fonts or other assets. The source documents separate authorization/publication decisions and private, noindex output defaults. No original-site assets or model payloads were downloaded, and no browser session or upstream script was run in this review.

About this entry

Type: Case study
Category: Independent reconstruction cases
Relationship: Independent reconstruction project
Targets: Website

Sources reviewed 2026-10-17

The description was reviewed against primary sources. This does not establish runtime testing, security auditing or verified compatibility.

Requirements and limitations

About this entry

Primary sources