---

name: app-and-ui description: "Run and troubleshoot the H2O LLM Studio Wave GUI, app runtime, Docker entry points, settings storage, and UI/server lifecycle." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

App and UI

Use this sub-skill when the task is about starting, operating, or debugging the H2O LLM Studio graphical app: H2O Wave launch commands, Docker runtime, browser/service lifecycle, app settings, user database paths, credential persistence, download links, and UI-server connectivity.

Route here for

• Starting the GUI with make llmstudio or a direct wave run llm_studio.app command.
• Choosing H2O Wave environment variables for uploads, downloads, reverse proxies, remote access, and timeouts.
• Understanding how the Wave app initializes q.app, q.client, settings, datasets, and the local SQLite database.
• Locating the persistent data/ and output/ directories controlled by H2O_LLM_STUDIO_WORKDIR.
• Explaining or troubleshooting keyring, .env credential storage, YAML user settings, and settings restore/save behavior.
• Running safe preflight checks before starting a long-lived Wave server.
• Planning optional browser/UI verification with Playwright when a server and browser stack are intentionally available.

Route elsewhere

• Experiment YAML schema, dataset columns, problem types, and import-format validation: ../configuration-and-data/.
• Training commands, train.py, GPU/DeepSpeed training failures, and experiment artifacts after a run starts: ../training-and-experiments/.
• Interactive prompting, model download/export, and Hugging Face publishing: ../export-and-prompt/.
• Model wrapper, loss, metric, inference, and evaluation internals: ../modeling-and-evaluation/.

First actions

1. Read runtime setup for launch commands, runtime-root assumptions, Docker commands, workdir layout, and Wave environment variables.

2. Use the safe checker before starting a service:

   python sub-skills/app-and-ui/scripts/check_app_runtime.py --runtime-root .
   

   Add --prepare-dirs only when you intentionally want the checker to create missing data/ and output/ directories.

3. Read app architecture when diagnosing initialization order, settings persistence, SQLite DB state, default dataset import, or download-link behavior.

4. Read troubleshooting for common service, browser, remote-origin, keyring, Docker, and filesystem failures.

Operating reminders

• make llmstudio is GPU-oriented and runs nvidia-smi before launching Wave. For app-import or UI-layout debugging without proving GPU readiness, use the direct wave run llm_studio.app form from the runtime root.
• Run the app from a runtime root containing the package assets expected by the app, not from an arbitrary directory. The app uses relative paths for static assets, Python config modules, prompts, and model-card templates.
• The app stores persistent user data under the resolved H2O LLM Studio workdir, not under Wave's process directory unless those are the same.
• For remote or proxied access, treat browser origin, allowed origins, public base URL, and Wave app timeouts as part of the app configuration, not as training issues.