version: "1.0.1" name: rtdt-core description: Shared foundation for Real-Time Digital Twins (RTDT) work. Use when starting or health-checking the standard or lite streaming stack, deciding whether a task needs a built Kit application, or reasoning about the current control and inference contracts. Read this before using rtdt-customize or rtdt-aero.

RTDT Core

Use this skill as the operational and contract foundation for RTDT work. The repository combines a workflow/controller layer, a specialization, USD stages, a browser UI, Kit-CAE 3.0 visualization, and either live Triton inference (standard) or an offline result cache (lite). Some controller interfaces are workflow-driven, but the shipped frontend and a few runtime paths still contain auto-aero conventions; do not assume the application is fully specialization-neutral.

The supported build, container, and runtime host is Linux (Ubuntu 22.04 or 24.04). Do not infer Windows or WSL support from cross-platform tooling in dependencies.

Choose the execution mode

• Static authoring or review: inspect and edit TOML, USD text, Python, or documentation without

starting Compose. A built checkout is required only for checks that import Kit/USD schemas.

• Standard profile: live Kit rendering plus AeroNIM/Triton inference. Use either two GPUs with

at least 40 GB each or one GPU with more than 80 GB, and follow the NGC authentication requirements. On a one-GPU host, set CUDA_DEVICE_KIT and CUDA_DEVICE_AERONIM to the same host index (normally 0); the .env_template defaults assume a two-GPU host.

• Lite profile: one-GPU rendering with a fixed pre-baked cache. It cannot synthesize a missing

inference result.

Before a runtime launch, initialize the Kit-CAE submodule and Git LFS assets, copy .env_template to .env, and select exactly one profile; standard and lite cannot run together because they publish the same Kit streaming ports. Compose builds and packages the Kit application inside Dockerfile; a separate host ./repo.sh build is required only for host-side tests or launches. Standard additionally requires a valid NGC API key and registry login.

Treat .env as secret-bearing. Do not print the fully resolved Compose model with docker compose config, because it expands values such as NGC_API_KEY. Use docker compose --profile <name> config --quiet for validation, or extract only a known-safe field without emitting the full command output.

Start, stop, and check the application

Run these commands from the repository root:

docker compose --profile standard up -d
docker compose --profile lite up -d

Stop the profile explicitly. When changing profiles, remove every service from the Compose project before starting the other profile:

docker compose --profile standard down
docker compose --profile lite down
docker compose --profile '*' down

The first image build and Kit startup can take several minutes. Use the profile-aware preflight with a bounded retry window:

skills/core/rtdt-core/scripts/check_readiness.sh --profile standard --wait-seconds 600
skills/core/rtdt-core/scripts/check_readiness.sh --profile lite --wait-seconds 600

The preflight does not parse .env. When .env changes the published web or proxy port, pass the same value explicitly with --web-port or --proxy-port (or export the matching environment variable before running it). The shipped standard Compose topology publishes Triton on host port 8080; use --triton-port only when a Compose override changes that host mapping. This preflight validates the bundled local topology, not an AWS split-host or other remote-Triton deployment.

The preflight reports Compose services, the direct web endpoint, the launch proxy, and Triton/model readiness where applicable. It does not prove that a browser established the WebRTC stream, that the scene rendered correctly, or that a control round-trip succeeded. Verify those separately when the task requires an end-to-end runtime claim.

For a blank scene or missing controls during startup:

1. Read docker compose logs kit-lite (or kit for standard). app ready can precede

RTX ready by several minutes while shaders compile; a reachable web page is not renderer readiness.

1. Keep the browser open through initialization. Confirm the scene appears and the loading message

gives way to controls, including velocity options, without a reload.

1. Change velocity and visualization mode; verify the rendered result changes, not just the selected

button. If it fails, capture Kit logs, browser console errors, a screenshot, and OS/GPU/driver details before diagnosing. Compare a restart with retained caches to distinguish cold-start delay.

Execution rules

• Run repository commands with the documented ./repo.sh entry points. Code importing pxr,

omni.*, or carb must execute in the built Kit application, commonly with ./repo.sh launch -n <app> -- --exec <script> --no-window.

• Use python for pure-standard-library orchestration. Do not use a system interpreter for

Kit-dependent code.

• A one-shot Kit script must flush its output, request application shutdown, and return a reliable

exit status. A long-lived stream handler must never terminate the host application.

• The current request/response control API is transported through the browser's WebRTC messaging

channel. Do not claim a direct external HTTP or Python control endpoint exists.

Current contracts

Read contracts.md before changing controller behavior, output fields, cache semantics, or inference integration. In particular:

• get_control_manifest is useful metadata, but the shipped frontend does not render arbitrary

controls from it.

• Stage-authored result-selection APIs determine which inference outputs are requested.
• Kit-CAE 3.0 results declare native OmniSci field metadata, while RTDT stores immutable NanoVDB

buffers in omni.cae.core.cache and reconstructs a warp_simdata.Field through GetField on demand. It does not author native USD array values.

• The AeroNIM base models retain their validated Warp version from aeronim/requirements.txt.

RTWT's Warp 1.13 and Warp SimData dependencies are installed from requirements-rtwt.txt into an isolated model path. Do not merge these environments unless the upstream NIM supports the newer Warp runtime.

• Lite cache availability is enforced per control axis; it is not a list of guaranteed-valid tuples.

For workflow/schema/geometry authoring, continue with rtdt-customize. For the shipped wind tunnel, continue with rtdt-aero.

Resources

• contracts.md — control, output, cache, and inference boundaries.
• check_readiness.sh — bounded profile-aware runtime preflight.