preset-skill-nvidia

name	preset-skill-nvidia
description	Use when revising the built-in nvidia TUI preset.
version	3.0.0
last_changed_at	2026-09-07T00:00:00Z
related_files	["tui/internal/preset/preset.go","tui/internal/tui/preset_editor.go","tui/internal/tui/SKILL.md","tui/CONTRACT.md","tui/internal/preset/revision.go","tui/internal/headless/preset_revision.go"]
maintenance	If you find stale or incorrect information here, use the lingtai-issue-report skill to assemble evidence and obtain per-issue human consent before filing an issue. Never include secrets, credentials, tokens, or private paths.

nvidia preset revision

Use this child for the named built-in nvidia preset. nvidiaPreset in tui/internal/preset/preset.go:1397 ships the bounded NVIDIA gateway snapshot nvidia/nemotron-3-ultra-550b-a55b, https://integrate.api.nvidia.com/v1, NVIDIA_API_KEY, OpenAI compatibility, web_search, and skills; it has no default vision capability. This is a gateway catalog, not a single NVIDIA model family.

Template-specific settings

Read the official NVIDIA API Catalog, model catalog, and NIM LLM API reference. Use the catalog's exact model slug and, for the selected account, query the authenticated OpenAI-compatible GET https://integrate.api.nvidia.com/v1/models when available. Check the model page for modality and tool support; a VLM in the gateway catalog does not automatically add this preset's vision key. Keep catalog slugs distinct from provider names and never copy a volatile catalog wholesale into the picker.

The curated picker is exactly the reviewed stable snapshot: Nemotron Ultra, Nemotron Super/Nano Omni, DeepSeek v4 pro/flash gateway slugs, Kimi K3, MiniMax M3, Mistral Nemotron, GPT-OSS 20B, and the Nemotron Ultra 253B entry. This is not an all-model universal generation ladder. The kernel's nvidia registration disables prompt_cache_key because NIM rejects that OpenAI-only field.

TUI surfaces to revise

Start at nvidiaPreset in tui/internal/preset/preset.go. Revise providerModels["nvidia"] in tui/internal/tui/preset_editor.go for the exact route-served snapshot. This target is unsupported by the universal two-generation JSON revision engine; maintain it as manual code curation. Do not add modelHasVision entries unless the TUI constructor also wires a verified vision route; recheck the constructor's absent vision capability and explicit base_url/env fields. Follow the gateway exemption and curation rules in tui/internal/tui/SKILL.md and tui/CONTRACT.md.

Reviewed deterministic revision

Prepare an evidence-bound manifest and explicit input, then run lingtai-tui presets revise --manifest PATH --input PATH --mode dry-run|check|apply [--output-dir PATH]. Review the JSON plan, use dry-run/check before apply, and apply only to a new explicit output directory. revision.go validates hashes, route bindings, and evidence and preserves unowned bytes. This amendment is unsupported for universal JSON generation; the constructor and exact picker are manual route-served curation with stock no vision.

Maintenance: If the relevant TUI preset/page is revised, check whether this sub-skill also needs revision and, if so, include it in the same PR.

Operations

For save, endpoint/capability, availability, activation/refresh, or troubleshooting, use the five shared operation children under reference/operations/; this child owns NVIDIA gateway and capability facts.