Claude Skills by nebius

Prepare, verify, and resume human-bound Hugging Face or NVIDIA NGC access approvals for exact NPA catalog and workflow dependencies, including agent prompts that may open official pages only after affirmative consent.

Use when building, enhancing, or testing the NPA chat agent backend — grounded-first routing, cost-aware Token Factory model selection, the embedded-backend mechanism, and cheap-token test tiers.

Use when collecting NPA agent work for fine-tuning or evaluation so every goal-level episode emits a sanitized, outcome-linked trajectory to an append-only S3 dataset in an operator-specified Nebius tenant and bucket.

Use when the NPA agent should describe or critique the current viewer (Rerun, video, image, or data) via the Describe this control or multimodal chat.

Use for Claude Code architectural review of the Nebius Physical AI platform, workbench layer, orchestrator choices, and partner model.

Reconcile NPA container-image catalogs and related documentation whenever a workbench image or solution is added, removed, renamed, retagged, republished, reclassified, or materially changed. Use for docs/workbench/container-image-catalog.md drift, public-release inventory checks, and image or solution onboarding.

Use when building, tagging, validating, or publishing NPA workbench container images through immutable full-SHA public development tags and digest-identical supported GHCR releases.

Write or review NPA contributions, including workbench agent changes, using the repository's readability, documentation, and anti-pattern rules.

Use when navigating the Cosmos3 integration in NPA or locating upstream Cosmos3 framework files, defaults, scripts, configs, recipes, and docs.

Use when Cosmos3 setup, fetch, inference, CUDA, uv, Docker, Hugging Face, GitHub, NGC, or checkpoint staging fails in NPA or in an upstream Cosmos framework checkout.

Use when setting up Cosmos3 access through NPA, checking source or Hugging Face reachability, staging the public Cosmos3 framework and checkpoint cache, or deciding which NPA workflow to use before inference.

Use when a workflow run failed, hung, or produced no artifacts — an ordered triage from run id to root cause across status, stage logs, S3 evidence, pod-level reasons, and the resume-vs-cancel decision.

Use when discovering or loading run artifacts in npa agent without workflow/type/path allowlists.

Track typed GPU placement failures in the NPA agent and offer a consent-gated on-demand-to-preemptible fallback after repeated failures or deterministic preflight. Use when quota, capacity, Unschedulable GPU, or compatible-product placement blocks an allocation.

Use when recommending how to run a workbench workload, writing user-facing docs, or answering "what should I run first?" — lead with the Nebius GPU path and keep hosted-inference options framed as what they are, rather than as ways to avoid GPUs.

Use when choosing or reviewing GPU targets for NPA workbench tools, training, rendering, inference, or workflow YAML resources.

Use when a guardrail test or CI gate in npa fails and you need to map the failure to its cause and fix rather than reverse-engineering the assertion.

Use before any deploy, image build, provisioning, or GPU submit to prove service credentials with `npa workbench health preflight`, request `--checks nebius` before provisioning, and verify gated-model access with `npa workbench health access`.

Use when writing or modifying npa CLI commands or SDK functions — registration, option naming, path contract, output format, error handling, exit codes, the decorators that are easy to forget, and config/credential access.

Use for Claude Code reviews that need robotics, simulation, GPU-routing, sim-to-real, or BDD100K pipeline domain context.

Use before pushing an npa change to pick which gates apply and run them locally in cheapest-first order — the map from each CI job to its exact local command.

Use when preparing commits, documentation, reports, examples, tests, pull-request bodies, issues, or live-validation handoffs involving Nebius, to prevent concrete live infrastructure details or credentials from reaching Git or public collaboration surfaces.

Use when authoring or reviewing an NPA workbench pipeline/blueprint that advertises specific components (Cosmos Transfer, Cosmos Evaluator, Cosmos Curator, FiftyOne, VLM eval, etc.) — ensure every advertised stage invokes the REAL component, not an echo/manifest stub masquerading as real work.

Use during Claude Code reviews to classify API, IAM, cleanup, exception, concurrency, config, temp-file, and version-pin risks.

Use whenever an NPA container image is built, tagged, pushed, copied, or promoted, including public full-SHA development builds, release publication, build scripts, and image-producing GitHub Actions workflows. Enforce mandatory safety gates and refuse insecure or restricted publication.

Use when adding, onboarding, or repackaging any solution, tool, container image, model, weights, dataset, or runtime cache; classify each artifact separately, keep restricted weights out of images, design safe runtime caching, and record redistribution decisions where guards enforce them.

Use when submitting, validating, or debugging NPA SkyPilot workflow YAMLs and workflow runner paths.

Use when drafting, executing, or reviewing Codex super-prompts for this repository.

Use to stop spend safely and audit orphaned resources — the mandatory cancel-before-destroy ordering, which npa commands touch cloud versus local state, and how to find leaked clusters, agent VMs, controllers, and buckets.

Use before running or interpreting NPA tests, lint checks, or validation reports.

Use before provisioning, building, downloading, or submitting a workload whose image, software, model, or data requires third-party EULAs or gated terms; apply the scoped Isaac ACCEPT_EULA default and explicit opt-out correctly without conflating credentials, privacy, telemetry, or redistribution rights.

Use when adding or changing an npa.workflow toolRef argv template, or when changing a CLI option that a toolRef already passes — the contract that stops a stage from rendering cleanly and then crashing in the pod.

Start, recover, or verify human Nebius CLI authentication on a remote operator/dev VM with no browser, a safe loopback callback tunnel, and secret-free identity/IAM verification. Use for remote CLI profile setup; do not replace an npa-agent VM's attached-service-account metadata profile.

Package, preflight, run, validate, or troubleshoot NVIDIA Alpamayo 2 Super 34B VLA inference in NPA, including its OpenMDW model terms, separately gated PhysicalAI-AV dataset, public runtime-fetch image, B200/RTX PRO 6000 routing, workflow artifacts, and real-GPU evidence.

Use to turn run outputs into something a human can look at and to hand it to someone else — `npa adapter convert` (sim demos to LeRobotDataset), `npa convert lerobot-to-rrd|lerobot-to-mp4`, and `npa rerun host|share|list-shares|revoke` for time-boxed presigned links.

Use to run one gang-scheduled multi-node GPU job through SkyPilot without authoring a workflow — `npa burst submit` wires torchrun rendezvous across nodes, while `submit-yaml` supports one task, `${VAR}` substitution, and optional generic registry credentials.

Use when working on Cosmos world model serving, inference, serverless training smoke validation, backend selection, or rendering limitations.

Operate, validate, or troubleshoot persistent Cosmos3-Nano generation through NVIDIA Cosmos Framework's native Ray Serve implementation, including authenticated readiness, dynamic request batching, B200 versus RTX PRO 6000 placement, runtime-fetched weights/guardrails, and durable S3 batch outputs.

Reproduce, operate, validate, or troubleshoot the fixed Cosmos3-Super eight-GPU benchmark or isolated single-H200 TP-1 validation through the immutable public vLLM-Omni image.

Use when running, validating or reviewing NVIDIA cuRobo V2 Franka pose planning and complete MotionBenchMaker/MPiNets benchmarks on CUDA, with factual joint/FK Rerun artifacts.

Use when ingesting, validating, curating, or querying production sensor data as a versioned dataset-of-record, or wiring the dataset-ingest-curate workflow.

Use to train and evaluate Faster R-CNN detectors from LanceDB materialized views (BDD100K failure-mode slices) — direct versus deployed-service execution, the mandatory label map for string categories, and checkpoint discovery at eval time.

Use when deploying, launching, loading data into, or reviewing the FiftyOne workbench dataset curation and visualization tool.

Use to deploy or operate a fleet of Nebius Managed Kubernetes (k8s-training) clusters across one or many projects in a tenant from an npa.fleet/v0.0.1 spec — including strict capacity-block-backed GPU pools, identical and/or custom clusters, create-on-demand projects, and a k8s-training recipe source that can consume the latest upstream changes.

Use when embedding, operating, or debugging the Foxglove viewer in the NPA agent (the @foxglove/embed TypeScript SDK, MCAP recordings, the npa-foxglove-embed container, or /api/foxglove/* endpoints).

Use when working on Genesis simulation, RL teacher training, visual demo generation, or related serverless/EGL behavior.

Use to prove a workbench container image actually works — the per-container hello-world manifest, dry-run/local/serverless execution tiers, batch runs across every image, and the offline manifest validation that gates CI.

Use when creating or repairing a GPU Kubernetes cluster for npa — the managed-image vs GPU-Operator driver decision, why operator mode is unsafe on NVSwitch, the post-apply health gates (fabric, CUDA vectorAdd, stability window), and triage for nodes that come up without working GPUs.

Use when working on NVIDIA GR00T deployment, model download, finetuning, evaluation, serving, inference, conversion, status checks, validation, routing, or CUDA alignment.

Use when turning the reports/manifests other workbench tools emit into a queryable lineage graph + common metrics store, or when querying/comparing metrics, tracing artifact lineage, or building a run dashboard.