5 viral repos, distilled

A reversible theme overlay system for the official Codex desktop app. It injects full-window backgrounds and CSS via Chrome DevTools Protocol loopback, without touching the app binary, so themes install, switch, and roll back cleanly on both macOS and Windows.

The wedge is doing this entirely offline: bitwise image metadata parsing and luminance-based safe-area inference replace what most competitors would reach for an ML model or hosted service to do.

Reconstructs 3D objects and humanoid characters from a single reference image as procedural Three.js models. The pipeline stages through image intake, spec authoring with pass/fail quality contracts, code generation, and a visual self-correction loop, output is animation-ready geometry with rigging and PBR materials, not a baked mesh.

What separates it from typical iterative reconstruction is catching small identity-defining details (buttons, scratches, labels) through deterministic zoning before any code gets generated, so tokens aren't spent recovering from malformed specs.

Not a library, but a practice for shaping the environment agents work in: context, tools, permissions, repository structure, feedback loops. Eight core ideas (domain modeling, context routing, fixed-worker qualification, authority separation, maintenance loops, outcome proof, tool legibility, whole-job delegation), documented through theses, playbooks, and real domain examples.

The panel's biggest gap call: no single reference implementation demonstrates all eight ideas running together, so it reads as mature guidance without a working proof.

A terminal UI video downloader wrapping yt-dlp and ffmpeg. Paste a URL, get interactive format and resolution selection, download from 1800+ platforms, all without leaving the terminal.

The wedge is treating the terminal as a first-class UI: React-as-TUI via Ink, a real light/dark/auto theme system, and frame-capture click hit-testing, in a space where most yt-dlp wrappers stop at feature parity with the CLI tool.

An OpenAI-powered, offline-first PWA that extracts individual garments from photos, generates identity-preserving modeled editorial images of each piece, and lets users assemble outfits into lookbook photos. All data stays local in JSON; service workers cache images with trimming to stay usable on storage-constrained mobile devices.

The wedge is refusing synthetic polish: generated garment photos are constrained to match the real item's construction, texture, and color rather than drifting toward an idealized version, which is the trust problem most AI wardrobe apps don't bother solving.