---

name: web-ui-runtime description: "Install, configure, launch, and troubleshoot Nesa's encrypted AI web UI and model-management runtime." metadata: disco-role: operating disable-model-invocation: true license: NO_LICENSE

Nesa Web UI Runtime

Use this sub-skill when the user asks about the local browser demo, one-click installation, model downloads, CPU/GPU flags, web UI settings, or how Nesa plugs into the text-generation web UI.

Typical triggers:

• "start the Nesa demo UI"
• "why does the one-click installer fail?"
• "switch the demo from CPU to GPU"
• "download or select encrypted models in the UI"
• "what does equivariant-encrypt mode do?"

Decide the runtime target

Ask or infer these before running commands:

1. Operating system: Linux, macOS, or Windows.
2. Backend target: CPU, NVIDIA CUDA, AMD ROCm, Apple M-series/MPS, Intel Arc, or unknown.
3. Whether the user permits environment mutation and model downloads.
4. Whether the UI should be local-only or reachable from other hosts.
5. Which model path is needed: encrypted DistilBERT local classification or encrypted Llama remote streaming.

If the user only needs the small sentiment demo, route to encrypted-distilbert instead of launching the full web UI.

Safe-first workflow

1. Read references/installation-and-runtime.md for the platform and backend decision matrix.

2. Validate settings/flags before launch:

   python scripts/validate_runtime_config.py --settings /path/to/settings.yaml --cmd-flags /path/to/CMD_FLAGS.txt
   

3. Preview model-download naming before network downloads:

   python scripts/check_hf_model_plan.py nesaorg/distilbert-sentiment-encrypted
   

4. Only then run user-approved installer or launch commands.

5. If anything fails, read references/troubleshooting.md before installing broad extras or switching backend variants.

Important repo-specific behavior

• The checked default command flags include CPU mode.
• The settings template uses mode: equivariant-encrypt.
• The UI model menu dispatches selected model names through Nesa's model registry rather than ordinary generic loader behavior for the supported Nesa models.
• Model download helpers can fetch or checksum Hugging Face model files, but that is a network and disk-writing operation.
• Launching the UI can bind to 0.0.0.0; never expose it without auth unless the user explicitly approves that risk.

References and scripts

• references/installation-and-runtime.md: platform installer flow, dependency variants, backend choices, and launch advice.
• references/configuration.md: settings fields, command flags, and equivariant-encrypt mode details.
• references/webui-integration.md: model menu, registry dispatch, download helper, and prompt-flow integration.
• references/troubleshooting.md: install, runtime, model-download, and public-serving failure modes.
• scripts/validate_runtime_config.py: read-only settings and command-flag validator.
• scripts/check_hf_model_plan.py: safe model id, branch, and output-folder preview without downloading files.

Boundaries

• Do not use this sub-skill for low-level request struct changes; route to backend-protocol.
• Do not use this sub-skill for contest scoring/attack baselines; route to security-contest.
• Do not run one-click installers, large downloads, or web UI launch commands without user approval.