Lingtai-AI
Lingtai-AI/lingtai
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Lingtai-AI/lingtai --skill preset-skill-nvidia命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| 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. |
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.
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.
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.
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.
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.