Rocket Agent Lab

Source is publicly available for Build Week judging under LICENSE.md; redistribution and commercial reuse are not granted.

Interactive 3D rocket education prototype with nine Three.js rocket models, structured lessons, part picking, focused/all label modes, live challenge checklists, before/after slider impact readouts, physics equations, hardware/cost/failure breakdowns, WhatsApp sharing, and Astra, an OpenAI Realtime voice tutor with a prompt bar for scene-aware questions and replies.

The Build Week judge path is deliberately short: open Flight, press Run engine-out, inspect the deterministic before/after proof, then ask Realtime to coach from that exact receipt. The proof remains available without an API key; live voice requires a server-side key.

The viewport uses orbit-camera inspection with horizontal rotation, vertical pitch, shift-drag stack panning, ray-picked rocket parts, inspection shortcuts for engines/tanks/payload/thermal loads, split/cutaway modes, fuel and payload callouts, launch-pad environment detail, ignition plumes, shock diamonds, Max-Q pressure waves, staging flashes, heat glow, smoke, and ascent camera following. The Flight Lab adds beginner-readable orbit, escape, heavy-payload, and Max-Q stress tests with wind shear, guidance error, launch/fail/pass states, animated breakup, staged delta-v, mass ratio, live mass, peak Max-Q, risk, slider threshold cues, and mission margin from the same classroom model.

The selected-part panel includes educational estimates for mass, relative cost driver, materials, failure watchpoints, and optimization levers. Failure states now show a causal chain such as mass ratio loss, Max-Q overload, or wind/guidance coupling. These are learning approximations, not procurement-grade cost or mass data.

The 3D renderer is local/offline: rocket-three-scene.js imports vendored Three.js files from vendor/.

Supported platforms

• macOS, Windows, or Linux
• Node.js 20+
• A current Chromium, Firefox, or Safari browser
• Microphone permission and a server-side OpenAI API key only for the optional live Realtime voice path

Judge smoke test

node --test test/causal-engine-lesson.test.mjs test/realtime-context.test.mjs test/realtime-session.test.mjs test/realtime-client-wiring.test.mjs
node --check app.js

Then open the local app and run the engine-out lesson. Representative browser proof is stored at outputs/rocket-agent-lab-engine-out-proof-card-2026-07-19.png.

Run

cd /Users/arnav/Desktop/testing/rocket-agent-lab
node server.mjs

Open:

http://127.0.0.1:5186

OpenAI Realtime Voice

Set the API key only on the server:

OPENAI_API_KEY=... node server.mjs

The browser posts a WebRTC SDP offer to /api/realtime-call; the server forwards it to OpenAI and returns the SDP answer without exposing the standard API key to the browser. The app intentionally does not use Web Speech or a Mac voice fallback. Realtime is treated as the tutor backbone: voice, the prompt bar, launch-event narration, and quiz turns all route through GPT Realtime 2 with live scene context.

The prototype follows OpenAI's WebRTC Realtime pattern: the browser creates an SDP offer, the server creates the Realtime session, then the browser completes the WebRTC connection with the returned SDP answer.

For Vercel, set OPENAI_API_KEY as a project environment variable. The matching serverless endpoint is api/realtime-call.js. The default public model id is gpt-realtime-2 with the marin voice; set OPENAI_REALTIME_MODEL to override it.

Authorized human Realtime check

The local causal-context and data-channel tests are not a substitute for a live voice result. Before a demo or submission claim, run exactly one authorized microphone/WebRTC session using outputs/realtime-human-test-protocol-2026-07-18.md. It records only the selected scene, changed value, visible connection state, and pass/fail outcome—never an API key or raw SDP.

Build Week collaboration and provenance

Every capability presented for judging was built after July 13, 2026. The primary Codex build task is 019f5e81-8bde-79d1-b2d6-416776726a63. Codex/GPT-5.6 helped isolate the causal context contract, identify stale-state and opening-turn failure modes, build tests, and implement the deterministic proof surface. Human decisions kept simulation outcomes deterministic, restricted GPT to explanation/coaching, separated Rocket Agent Lab from Light Years From Home, and preserved the live-microphone test as an explicit human gate.

Physics Sheet

The equation source is in assets/rocket-equations.tex, compiled to assets/rocket-equations.pdf with the bundled LaTeX Tectonic runtime.