Chorusborn

Raised by chat. Proven in the arena. Twitch communities teach rival champions, test their tactics in branching replays, and earn credit when their lessons are used.

Comment

Inspiration

Twitch chat has ideas worth more than a fleeting message. I wanted two communities to raise recognizable rivals: champions whose habits come from the people watching, with a way to inspect whether a lesson actually helps.

What it does

Chorusborn turns streamer rivalry into a game of community coaching. Each streamer connects their own Twitch account and fields a champion named after their channel. Viewers submit natural-language tactics with !teach in their streamer's chat.

The loop is teach → prove → adopt → compete. The coach reviews an interpreted lesson, compares the current and proposed playbooks in branching replays from the same game state, then decides whether to adopt it for a future round. When an adopted rule actually executes, its viewer receives visible credit. More lessons do not automatically mean a stronger champion: timing, conditions, priority and trade-offs matter.

In Core Run, champions contest a core and carry it home while attacking, shielding and navigating hazards. The first to three captures wins, with a 60-second round limit. Both communities train independently, accept a challenge and ready their champions for one shared match. An OBS-friendly read-only overlay shows the arena, while a sign-in-free viewer guide explains the rules and gives concrete teaching examples.

How I built it

I built Chorusborn as a solo project in a TypeScript workspace with React, Vite, Three.js/React Three Fiber, Fastify and a deterministic simulation shared by live matches and replay workers. Twitch OAuth verifies each broadcaster; EventSub brings chat lessons into a channel-specific review queue. A bounded Groq-hosted OpenAI GPT-OSS 120B compiler translates supported language into validated behavior data. It cannot execute viewer code, publish a playbook or control the match clock. The server owns outcomes, and each active match pins its playbooks.

OpenAI Codex helped me develop the product scope and architecture, implement the engine and Twitch integrations, constrain and validate the compiler, build the coach/replay/overlay UI, animate the original procedural stick fighters, and debug failures with automated checks. The repository and build log record those changes and verification.

Challenges and what I learned

A fair-looking animation is not enough: equal core claims originally exposed ordering bias and later a zero-possession loop. I added explicit tie initiative, alternating starts for rematches and paired neutral-opening comparisons. I also separated receiving, testing, adopting and actually using a lesson so the UI does not mistake publication for execution—or execution for causing a win.

A targeted replay can expose a trade-off, but it does not establish universal tactic superiority. Coaches need inspectable evidence and clear limits, not a fabricated intelligence score.

Accomplishments and next steps

The prototype is live at chorusborn.vercel.app, with a viewer guide and a sample teaching/replay flow that needs no Twitch login. The Vercel runtime separates web gateways from one authoritative game daemon, with encrypted Redis checkpoints, fenced ownership and durable command receipts. Recovery pauses matches and revalidates saved Twitch sessions.

I tested a real two-account local match. The latest complete GitHub Actions run passed 102 unit tests, 18 browser tests, formatting, types and the production build. I also ran isolated real-Redis checks for independent gateways, duplicate adoption, spectator denial, two mocked Twitch communities, WebSocket delivery and paused recovery. Hosted sample proof and adoption have also been observed in the browser. Mock provider checks are labeled as such; they are not a substitute for a real hosted community demonstration.

The production Twitch callback is registered, and the hosted studio has successfully authenticated developersimiyu and established live EventSub chat intake. A fresh hosted community teaching demonstration remains to be recorded. Independent audience/broadcast validation, broader scenario evaluation and the final demo video remain release work. This is a bounded hackathon prototype, and a targeted replay result is not a claim of universal tactic superiority.

Built With

Share this project:

Updates

Submission history