564 skills published by NVIDIA across 46 repositories. Together they weigh 12 776 064 tokens — that is what loading all of them at once would cost you in context. 6 of them have been repackaged into other people's repositories.
564 skills 12 776 064 tokens total 6 copies elsewhere vendor account
Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.
>- Load and run a TensorRT engine (.plan / .engine) from C++ using the TensorRT 11 / 10.x **modern Runtime API**, avoiding the deprecated TRT 8.x binding-index APIs that older guidance still promotes. Use whenever the user asks about loading or running a TensorRT .plan/.engine from C++, even on "minimal example" requests — without this skill the default reply uses deprecated enqueueV2-style code. Also use when the user hits "Engine plan file is generated on an incompatible device", deserializeCudaEngine returns nullptr, gets an enqueueV2 / IStreamReader deprecation warning, or inference, load TensorRT plan C++, run .plan from C++, IRuntime example, deserializeCudaEngine, enqueueV3, enqueueV2 deprecated, setTensorAddress, getBindingIndex, IStreamReaderV2, libnvinfer C++. NOT for building engines (`trt-onnx-quickstart`), Python deploy, plugins, multi-GPU.
> Build and verify a TensorRT engine from a Hugging Face model ID or ONNX file, with numerical parity checked against ONNX Runtime. Use when the user imports a non-LLM model to TensorRT, needs a verified engine from ONNX, hits trtexec "unsupported operator", must verify the engine matches ONNX numerically, debugs a polygraphy parity failure (large max abs diff at FP16), or configures multi-input TensorRT, trtexec onnx, trtexec unsupported operator, optimum-cli export, polygraphy parity check, polygraphy run --trt --onnxrt, parity check failed, max abs diff, verify engine matches ONNX, --minShapes, dynamic shapes trtexec, multi-input shape profile, FP16 engine, INT64 `trt-cpp-runtime-quickstart` (C++ engine load). LLM token generation belongs in TensorRT-LLM, not here.
Validate and analyze TensorRT performance data from paired layer-info JSON and profile/latency JSON files. Use when asked to inspect TensorRT, TRT, torch-tensorrt, or ONNX-TensorRT perf reports, verify that layer/profile JSON files are valid and from the same model, infer basic model information, find likely fusion or latency optimization opportunities, and produce a concise Markdown performance report or structured JSON data.
>- Migrate a TensorRT build from weak typing (deprecated 10.12, removed 11.0) to strong typing — across Python INetworkDefinition builders, the trtexec CLI, and kSTRONGLY_TYPED, weak typing deprecated, kFP16/kINT8 removed, setPrecision rejected, setComputePrecision deprecated, do I still need --stronglyTyped, how to add the kSTRONGLY_TYPED flag, ModelOpt autocast, INT8 on TRT 11. NOT for ONNX import (`trt-onnx-quickstart`), Torch-TRT (`trt-torch-quickstart`), or C++ deploy (`trt-cpp-runtime-quickstart`).
> Compile a PyTorch model to a TensorRT engine via Torch-TensorRT — AOT or JIT — under the new strong-typing default. Use when the user compiles PyTorch to TensorRT without ONNX, hits "enabled_precisions should not be used when use_explicit_typing=True", sees Dynamo graph breaks or PyTorch fallback, debugs ABI errors at import torch_tensorrt, or needs the compatible torch / torch_tensorrt / tensorrt-cu13 version pins for torch.compile backend torch_tensorrt, pytorch to tensorrt, ExportedProgram, Dynamo graph break, use_explicit_typing, enabled_precisions, torch_tensorrt.Input, min_block_size, truncate_double, tensorrt-cu13, `trt-cpp-runtime-quickstart`. LLM token generation belongs in TensorRT-LLM.
Create a GitHub pull request with the NemoClaw template. Then, monitor CI and automated reviews. Use this skill when the user asks to create, open, push, or submit a PR for review. Trigger keywords - create PR, pull request, new PR, submit for review, open PR, push for review.
Prepare a NemoClaw source checkout for compliant contribution through the repository's one-command setup and readiness doctor. Use when a new contributor asks to set up a development machine, prepare a checkout for a first PR, repair local contributor tooling, verify contributor readiness, launch the pinned coding agent, or decide whether optional runtime onboarding is needed. Trigger keywords - contributor setup, developer onboarding, first PR, dev setup, dev doctor, repair checkout, prepare development machine.
Add or review a NemoClaw messaging channel through the current messaging architecture. Use when implementing channel support for one or more agent runtimes, mapping upstream behavior into NemoClaw, or reviewing credentials, packages, reachability, network policy, documentation, and tests for a channel. Trigger keywords - add messaging channel, onboard messaging channel, new channel, messaging integration, channel support.
Audit and implement a dependency upgrade as a semantic migration. Use when changing a library, CLI, service, image, runtime, installer artifact, or transitive dependency, including a Hermes release. Trace upstream changes into current NemoClaw consumers, resolve security and lifecycle concerns, and verify the exact artifacts that NemoClaw uses. Trigger keywords - update dependency, upgrade dependency, bump version, dependency migration, release audit, update Hermes, upgrade Hermes, review Hermes release, publish Hermes base image.
Find user-visible changes in current NemoClaw history and update their owning documentation. Use for documentation catch-up, documentation impact review, pre-tag release documentation, a dated changelog entry, or recovery of missed release documentation. Derive pages, commands, variants, and validation from the current checkout. Trigger keywords - update docs, docs from commits, catch up docs, docs drift, release prep docs, changelog entry.
Scan open issues to find issues that a PR can also fix or conflict with. Report each relationship with file and line evidence. Use this skill during PR review to find related fixes and risks.
Creates deterministic NemoClaw semver release tags on origin/main after verifying the pre-tag dated changelog entry, handles release housekeeping, drafts announcement release notes, and verifies the maintainer-published Announcement. Use when cutting a release, tagging a version, shipping a build, creating vX.Y.Z tags, publishing release announcements, or completing release communication.
Run one NemoClaw daytime maintainer pass. Prioritize items for the release version. Select a merge, salvage, security, test, conflict, or sequencing workflow and report progress. Use during the workday to land PRs and close issues. Designed for /loop, for example /loop 10m /nemoclaw-maintainer-day. Trigger keywords - maintainer day, work on PRs, land PRs, make progress, what's next, keep going, maintainer loop.
Dispatches and verifies trusted GitHub Actions E2E for NemoClaw maintainers. Use for requests such as run the E2E suite, run the Launchable E2E, run the full E2E suite, deploy pre-release full E2E, run pre-tag full E2E, or run release-candidate E2E.
Runs the end-of-day NemoClaw release handoff, including the pre-tag dated changelog PR, version progress, straggler planning, QA summary, tag cut, and announcement draft. Use at the end of the workday. Trigger keywords - evening, end of day, EOD, wrap up, ship it, cut tag, handoff, done for the day, pre-tag release notes.
Find open PRs with the security label and Urgent or High Project Priority. Link each PR to its issue. Identify competing or superseded PRs and report review candidates. Use when looking for the next PR to review. Trigger keywords - find pr, find review, next pr, pr to review, duplicate pr, security pr.
Runs the morning maintainer standup for NemoClaw. Triages the backlog, determines the day's target version, labels selected items, surfaces stragglers from previous versions, and outputs the daily plan. Use at the start of the workday. Trigger keywords - morning, standup, start of day, daily plan, what are we shipping today.
Normalizes GitHub issue and PR titles by removing any bracketed [NemoClaw] tag case-insensitively, even when the tag appears later in the title. Use when cleaning issue tags, bulk-renaming titles, or normalizing repo title hygiene.
Provide read-only NemoClaw maintainer policy. Use for questions about Issue Type, labels, Project fields, release labels, triage, duplicates, blocked items, and maintainer decisions. Trigger keywords - maintainer policy, workflow policy, project workflow, issue type, labels, label taxonomy, needs labels, project status, blocked issue, duplicate issue, daily release label, release train, triage policy.
Compare open PRs that address the same issue and recommend one to merge. Apply eligibility, correctness, quality, and tie-break checks. Report the score and evidence. Use when an issue has two or more open PRs.
Plan and execute maintainer-owned refactors of oversized NemoClaw Fern documentation sections into focused one-topic pages with concise prose, journey-based nested navigation, non-clickable group nodes, canonical troubleshooting and reference ownership, deduplicated content, variant-aware route-style links, and complete redirects. Use when a docs page or section has grown too large, when reorganizing documentation information architecture or a table of contents, when splitting pages, shortening dense paragraph blocks, moving content across sections, consolidating duplicate guidance, or migrating URLs in docs/index.yml and fern/docs.yml. Trigger keywords - refactor docs, reorganize docs, split docs, nested TOC, documentation IA, one topic per page, big paragraphs, move troubleshooting, deduplicate docs, docs too long.
Drafts the post-tag NemoClaw Announcement from live GitHub tag and compare data. Produces three lead paragraphs, categorized shipped changes, why-it-matters bullets, and external-only contributor thanks. Use after cutting a release tag or when asked to draft an announcement or summarize v0.0.x. Native `docs/changelog/YYYY-MM-DD.mdx` entries belong to the pre-tag `nemoclaw-contributor-update-docs` workflow.
Review a PR, or a PR linked to an issue, for security risks. Check nine categories and report PASS, WARNING, or FAIL. Use when reviewing code for vulnerabilities, secrets, injection, authorization bypasses, or unsafe configuration. Trigger keywords - security review, code review, appsec, vulnerability assessment, security audit, review PR security.
Triage NemoClaw issues and PRs with Issue Type, Project fields, and allowed labels. Support one item or a batch. Show proposed changes and apply only changes the maintainer accepts. Trigger keywords - triage, label issues, suggest labels, batch triage, triage issue, triage PR, label this, what labels.
Verifies whether stale NVIDIA/NemoClaw bug reports still reproduce on the latest tag. Use when maintainers ask to verify stale issues, reproduce old bugs on latest, or drain the bug backlog. Uses native Issue Type, canonical labels, local/Brev reproduction, by-design detection, confidence scoring, approved Project updates, and redacted verdict comments; never auto-closes.
Start here. Introduces what NemoClaw is, what agent skills are available, and which skill to use for a given task. Use when discovering NemoClaw capabilities, choosing the right skill, or orienting in the project. Trigger keywords - skills, capabilities, what can I do, help, guide, index, overview, start here.
Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.
Bump the NVIDIA PyTorch base image (`nvcr.io/nvidia/pytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker/.ngc_version.dev` and GitLab CI in `.gitlab/stages/01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs #4611 and #4688.
CI/CD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request/BRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.
Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.
Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.
Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.
Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.
How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.
Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.
Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.
Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.
Research and draft a response to a GitHub issue or question from an external contributor.
Refresh golden values from a GitHub Actions workflow run (failing-only or all jobs), score the change with average normalized relative differences, and produce a PR-ready summary. Use when the user asks to update goldens for a CI run, refresh golden values from a workflow ID, or generate a golden-value diff summary for a PR description.
> Debug AutoDeploy accuracy regressions vs a reference score (PyTorch backend or published baseline). Use when an AutoDeploy model's eval score is significantly below the reference and the root cause is unknown.
> transforms under transform/library/ in a TensorRT-LLM checkout. Prefer existing kernels and custom ops; use Triton only when no viable existing-kernel path exists. Use ad-graph-dump for AD_DUMP_GRAPHS_DIR workflows. Covers TRT-LLM paths, registry, default.yaml registration, graph validation, tests, and a review checklist — without prescribing profiling tools or throughput targets.
> Check whether AutoDeploy YAML configs were actually applied by analyzing server logs and optionally graph dumps (AD_DUMP_GRAPHS_DIR). Use when the user wants to verify config application, debug config issues, or check if AutoDeploy transforms (piecewise CUDA graph, multi-stream, sharding, fusion, etc.) were applied or fell back. Triggers "config not working", "check if piecewise is enabled", "check log for config", or any request to compare AD YAML settings against runtime behavior.
> Enable and interpret TensorRT-LLM AutoDeploy FX graph text dumps via AD_DUMP_GRAPHS_DIR. Use when you need before/after graphs per transform, to locate subgraphs, or to confirm a rewrite ran. Paths and behavior are grounded in tensorrt_llm/_torch/auto_deploy (GraphWriter, BaseTransform). Complements ad-add-fusion-transformation.
> Visualize a specific transformer decoder layer from an AutoDeploy FX graph text dump as a hierarchical DOT/PNG diagram. Optionally annotate nodes with actual GPU kernel names and durations from an nsys trace. Use when the user wants to visualize, inspect, "show layer", "graph of layer", "layer visualization", "dump graph layer". Assumes graph dumps already exist in a directory (produced by AD_DUMP_GRAPHS_DIR).
Translates a HuggingFace model into a prefill-only AutoDeploy custom model using reference custom ops, validates with hierarchical equivalence tests.
> Translates a HuggingFace model into a prefill-only AutoDeploy custom model using reference custom ops, validates with hierarchical equivalence tests.
> Adds sharding-aware IR hints (op substitutions, sharding kwargs, all_reduce insertions) directly into an existing AutoDeploy custom model (modeling_*.py). Edits the file in place — no separate _ir.py copy. Validates with apply_sharding_hints and end-to-end multi-GPU runs.
Compile TensorRT-LLM on a compute node inside a Docker container. Use this when already on a compute node with GPUs visible.
Compile TensorRT-LLM on a SLURM cluster. Covers submitting a batch job with a container image, monitoring the job, and verifying the build. Use when the user wants to compile TRT-LLM remotely via SLURM rather than on a local compute node.
> Write and implement GPU kernels using NVIDIA CuTe DSL (CUTLASS 4.x Python API) — NOT for Triton, CUDA C++, or conceptual explanations. Trigger only when the user wants to write or implement a kernel, not when asking questions about CuTe DSL concepts or layouts. CuTe DSL uses cute.jit/cute.kernel decorators and cutlass.cute imports. Covers element-wise kernels, GEMM patterns, reductions, memory hierarchy (global/shared/register/TMA), MMA tensor core operations, software pipelining, and framework integration.
> Optimize existing Triton kernels for NVIDIA TileIR backend on Blackwell GPUs (sm_100+). kernel classification (dot-related, norm-like, elementwise, reduction), type-specific "add TileIR configs", "Blackwell optimization", "TMA descriptors", "2CTA mode", "occupancy tuning". Kernels use standard `import triton`; TileIR activates via ENABLE_TILE=1 when nvtriton is installed.
> ONLY for OpenAI Triton (@triton.jit) kernel development. NEVER use for CUDA C++ kernels, TileIR, or profiling tools (ncu, nsys). The user's request must involve Triton explicitly. Covers Triton-specific design, write, verify (with fast-path for explicit requests).
> Performance analysis coordination workflow. Guides profiling delegation, bottleneck classification (compute/memory/launch/communication/sync), and structured report generation. Use when the user asks to analyze performance, profile a workload, check MFU/SOL, or diagnose bottlenecks.
> Analyze host/CPU overhead in TensorRT-LLM inference from nsys traces. Detect whether host overhead is the bottleneck using GPU idle ratio, host prep exposed ratio, and per-phase evidence. For regressions, isolate forward steps via allreduce/NVTX patterns, compare host operation breakdowns across versions, and identify scheduling or request-management overhead. Supports optional inter-kernel gap, eager-vs-graph, pattern mapping, and multi-rank straggler inter-step gap, scheduling overhead, forward step isolation, nsys iteration analysis, NVTX breakdown, request management overhead, GPU idle, host bottleneck, host prep exposed, inter-kernel gap, bubble analysis, graph coverage, eager kernel, rank imbalance, straggler detection.
Profiles and optimizes TensorRT-LLM host/CPU overhead using line_profiler (with nsys support planned). Runs iterative profile-analyze-optimize-validate rounds. Use when GPU utilization is low or optimizing PyExecutor throughput.
> classification, roofline analysis, occupancy diagnosis, memory hierarchy analysis, warp stall analysis, metric interpretation, and programmatic .ncu-rep report analysis. NOT for kernel writing or code generation, Nsight Systems (nsys), host-side profiling, or system-level profiling.
>- Nsight Systems (nsys) CLI for system-level timeline profiling. Use when the user wants to run nsys profile, analyze .nsys-rep reports, use nsys stats/analyze/recipe commands, diagnose GPU idle time from timeline traces, or profile distributed training with NCCL overlap analysis. NOT for kernel-level metrics like SOL%, occupancy, or roofline (use perf-nsight-compute-analysis for ncu). NOT for writing or generating kernels. NOT for applying optimizations like CUDA Graphs.
> Performance optimization coordination playbook. Contains specialist routing table, TileIR two-step pipeline, kernel generation specialist selection, prioritization criteria, and safe modification workflow. Use when the user asks to apply optimizations, write kernels, or improve performance. Covers both user-specified optimization and autopilot-driven iterative optimization.
>- Apply CUDA Graphs to PyTorch workloads — API selection (torch.compile, PyTorch make_graphed_callables, TE make_graphed_callables, MCore CudaGraphManager, FullCudaGraphWrapper, manual torch.cuda.graph), code compatibility, capture workflows, dynamic pattern handling, and troubleshooting. graph capture, graph replay, kernel launch overhead, CudaGraphManager, FullCudaGraphWrapper, full-iteration graph, stream capture.
>- Identify and eliminate host-device synchronizations in PyTorch code. Detects sync points (.item(), .cpu(), boolean indexing, torch.tensor on CUDA), classifies false vs true dependencies, provides sync-free alternatives. eliminate syncs, CPU stall, non_blocking, set_sync_debug_mode, cudaStreamSynchronize, cudaEventSynchronize, remove syncs, async GPU.
> (1) Training loop — inject manual timing to report per-iteration latency, throughput (samples/sec), and data load time. (2) Standalone kernel/op — write CUDA event timing code with warmup, per-iteration statistics, and anti-pattern avoidance. Also covers NVTX annotation for labeling profiler timelines. Nsight Compute), writing kernels (Triton, CuTe, CUDA), applying optimizations (CUDA Graphs, gradient checkpointing, fusion), or interpreting roofline/SOL% metrics. training loop", "samples per second", "NVTX annotate", "instrument my dataloader", "data load time", "kernel timing", "how do I time".
> Systematic approach to exploring the TensorRT-LLM codebase before implementing new features or optimizations. Teaches how to discover existing infrastructure, trace code paths, and avoid reimplementing what already exists. Derived from real mistakes where ~250 lines of code were written and deleted because existing forward methods weren't discovered upfront. Use when starting any new feature, optimization, or code modification in TRT-LLM.
> Best practices for contributing code to TensorRT-LLM. Covers the official contribution process (issue tracking, fork workflow, DCO signing), coding guidelines, implementation workflow, common mistakes, testing strategy, commit hygiene, and review readiness. Incorporates rules from CONTRIBUTING.md and CODING_GUIDELINES.md plus lessons distilled from real PR retrospectives. Use when implementing new features, optimizations, or bug fixes in the TensorRT-LLM codebase.
>- Upgrade flashinfer-python version in TensorRT-LLM. Fetches the latest releases from GitHub (stable and nightly), compares with the current pinned version, lets the user pick a target version, and updates all version references across the repo. Use when the user wants to bump or upgrade flashinfer.
> Onboard a HuggingFace multimodal model (vision/audio/video + text) to the TensorRT-LLM PyTorch backend. Use when writing a new `tensorrt_llm/_torch/models/modeling_<vlm>.py` plus its input processor and weight mapper, or extending an existing VLM. Not for AutoDeploy — use `ad-model-onboard` for that path.
>- Review, design, and refactor TensorRT-LLM PyTorch MoE code for architecture fit, clean code, maintainability, and testability. Always use for any modification, review, refactor, or design planning that touches MoE modules, including tensorrt_llm/_torch/modules/fused_moe, ConfigurableMoE, MoE backends, MoEScheduler/moe_scheduler.py, forward execution/chunking, communication strategies, EPLB, quantization/weight handling, routing, factories, MoE docs, or MoE tests. Also use when the user asks whether a MoE design follows the current architecture or whether a MoE refactor is reasonable.
Generate a source-backed starting `trtllm-serve --config` YAML for basic aggregate single-node PyTorch serving, aligned with checked-in TensorRT-LLM configs and deployment docs. Preserves explicit latency / balanced / throughput objectives. Excludes disaggregated, multi-node, and non-MTP speculative configs.
Given a GitHub issue number, plan and implement the work described in the issue. Supports direct user requests and unattended queue processing through the `agent:*` workflow labels. Includes tests, documentation updates, and PR creation. Trigger keywords - build from issue, implement issue, work on issue, build issue, start issue.
Create GitHub issues using the gh CLI. Use when the user wants to create a new issue, report a bug, request a feature, or create a task in GitHub. Trigger keywords - create issue, new issue, file bug, report bug, feature request, github issue.
Create GitHub pull requests using the gh CLI. Use when the user wants to create a new PR, submit code for review, or open a pull request. Trigger keywords - create PR, pull request, new PR, submit for review, code review.
Create OpenShell RFC proposals in rfc/ from a design request. Use when the user asks to write, draft, start, create, or update an RFC, Request for Comments, architecture proposal, API proposal, process proposal, or cross-cutting design proposal that should follow the OpenShell RFC process and template.
Investigate a plain-language problem description by deeply exploring the codebase, then create a structured GitHub issue with technical findings. Prequel to build-from-issue — maps vague ideas to concrete, buildable issues. Trigger keywords - spike, investigate, explore, research issue, technical investigation, create spike, new spike, feasibility, codebase exploration.
Debug why inference.local, direct external inference, or supervisor-only system inference is failing. Use when the user cannot reach a local model server, has provider base URL issues, sees inference verification failures, hits protocol mismatches, or needs to diagnose inference on local vs remote gateways. Trigger keywords - debug inference, inference.local, system inference, sandbox-system, local inference, ollama, vllm, sglang, trtllm, NIM, inference failing, model server unreachable, failed to verify inference endpoint, host.openshell.internal.
Debug why an OpenShell gateway deployment is unhealthy, unreachable, or unable to create sandboxes. Use for gateway health failures, Docker/Podman runtime issues, Helm failures, Kubernetes scheduling, TLS or auth, gateway interceptors, supervisor middleware startup or runtime failures, external compute-driver sockets, VM drivers, or sandbox startup. Trigger keywords - debug gateway, gateway failing, deployment failing, helm install failing, cluster health, gateway health, gateway not starting, health check failed, sandbox pending, docker driver, podman driver, kubernetes driver, external driver, compute driver socket, gateway interceptor, supervisor middleware, middleware failed, vm driver.
Implement a fix for a reviewed security issue. Takes a directly requested issue number or scans for issues labeled `topic:security` and `agent:implementation-requested`. Reads the security review from the issue comments and implements the remediation plan. Trigger keywords - fix security issue, remediate security, implement security fix, patch vulnerability.
Validate and monitor OpenShell GitHub issues and PRs using the gator:* state machine. Use when asked to triage issues/PRs for project validity, gate PRs, run gator, validate submissions, or monitor PRs toward merge readiness.
Generate sandbox security policies from plain-language requirements and optional REST API documentation. Produces L4 or fine-grained L7 network policies and ordered network middleware configuration. Use for API access rules, middleware host selection, failure behavior, or built-in and operator-run middleware attachment. Trigger keywords - generate policy, create policy, update policy, change policy, sandbox policy, network policy, API policy, security policy, allow API, restrict API, network middleware, supervisor middleware.
Start up, tear down, and configure the local Kubernetes development environment for OpenShell. Uses k3d (Docker-backed k3s) + Skaffold + Helm. Covers cluster lifecycle, optional add-ons (Keycloak OIDC, Envoy Gateway), HA testing, and port mappings. Trigger keywords - local k8s, local cluster, k3d, skaffold, helm dev, start cluster, stop cluster, tear down cluster, delete cluster, create cluster, helm:k3s, helm:skaffold, local dev environment, dev cluster, k8s dev, envoy gateway local, keycloak local, high availability, HA.
Launch and supervise OpenShell gator agents. Use when starting gator on issues or PRs, checking gator sandboxes, building the gator sandbox image, restarting stuck gators, inspecting gator logs, or experimenting with gator harness/model overrides. Trigger keywords - launch gator, start gator, run gator, gator sandbox, supervised gator, gator logs, restart gator.
Guide agents through using the OpenShell CLI (openshell) for sandbox management, gateway registration, provider configuration and refresh, policy iteration, settings, service exposure, BYOC workflows, and inference routing. Covers basic through advanced multi-step workflows. Trigger keywords - openshell, sandbox create, sandbox exec, sandbox connect, logs, provider create, provider profile, provider refresh, policy set, policy get, settings, service expose, forward, port forward, BYOC, bring your own container, inference, use openshell, run openshell, CLI usage, manage sandbox, manage provider, gateway add, gateway select.
Use when an OpenShell sandbox returns policy_denied, mentions policy.local, or needs a narrow network policy proposal.
Review a GitHub pull request by summarizing its diff and key design decisions. Use when the user wants to review a PR, understand changes in a branch, or get a code review summary. Trigger keywords - review PR, review pull request, summarize PR, summarize diff, code review, review branch, PR summary, diff summary.
Given a GitHub issue, review the issue for security implications. You'll make a determination if the claim in the issue is legitimate and should be addressed or will be a "won't fix." Trigger keywords - security issue, review security ticket, review security issue.
Generate and manage Software Bill of Materials (SBOMs) for the OpenShell project. Covers SBOM generation with Syft, license resolution via public registries, and CSV export for compliance review. Trigger keywords - SBOM, sbom, bill of materials, license audit, license resolution, generate sbom, sbom csv, dependency license, supply chain, license scan.
Detect and fix drift across agent-first infrastructure files. Ensures skill inventories, workflow chains, architecture tables, issue/PR templates, and cross-references stay consistent when skills, crates, or workflows change. Run after adding, removing, or renaming skills or components. Trigger keywords - sync agent infra, sync skills, update agent docs, check agent consistency, agent infra drift, sync contributing, sync agents.
Manually dispatch and iterate on the Release Canary workflow that smoke-tests published OpenShell artifacts (install.sh on macOS/Ubuntu/Fedora, Helm chart on kind) after each Release Dev publish. Use when changing `.github/workflows/release-canary.yml`, validating a release before tagging, debugging a canary failure, or reproducing a canary job locally. Trigger keywords - release canary, release-canary, canary failed, canary dispatch, test release canary, post-release smoke, install.sh canary, helm chart canary, kind canary, dispatch canary.
Assess, validate, and route community-filed issues for human disposition and roadmap placement. Takes a specific issue number or processes a confirmed batch of issues labeled state:triage-needed. Investigates reported behavior, separates objective findings from product decisions, and prepares validated issues for a human yes/no decision. Trigger keywords - triage issue, triage, assess issue, review incoming issue, triage issues.
Guide for developing the OpenShell TUI — a ratatui-based terminal UI for the OpenShell platform. Covers architecture, navigation, data fetching, theming, UX conventions, and development workflow. Trigger keywords - term, TUI, terminal UI, ratatui, openshell-tui, tui development, tui feature, tui bug.
Scan recent git commits for changes that affect user-facing behavior, then draft or update the corresponding documentation pages. Use when docs have fallen behind code changes, after a batch of features lands, or when preparing a release. Trigger keywords - update docs, draft docs, docs from commits, sync docs, catch up docs, doc debt, docs behind, docs drift.
Watch and monitor GitHub Actions workflow runs using the gh CLI. Use when the user wants to check workflow status, watch a running workflow, view CI/CD jobs, or monitor build progress. Trigger keywords - watch pipeline, pipeline status, CI status, check build, monitor CI, view pipeline, pipeline progress, workflow status, actions status.
Audit Warp CHANGELOG.md before a release: recover lost entries, sort by user impact, refine entry language, line-wrap, and (release-branch mode) bump compare refs.
Generate a Warp release audit report (pre-release or RC, auto-detected from version string and head ref).
Generate a starting-point GitHub release notes draft for an upcoming Warp release (feature or bugfix, auto-detected from the version). Run from a release branch, or pass a branch / git ref as the argument.
Use when the user provides Warp commit SHA(s) and GitHub issue number(s) to assess, draft issue comments, post progress updates, or recommend whether issue threads should stay open or close.
>- a request to improve, optimize, or cut compile times; an app that is slow to start or stalls at the first wp.launch; seconds of compiling before real work begins; JIT modules recompiling on every run or every CI job. Only applies when the code being optimized uses Warp kernels. Not for steady-state kernel runtime, memory, correctness, building Warp itself from source, or nvcc/C++ build times.
>- Use to diagnose and fix incorrect gradients in differentiable Warp programs. Anything trained, optimized, calibrated, or fit through Warp kernels depends on wp.Tape gradients, so treat any misbehavior of such a workflow as a gradient problem until proven otherwise — use this when training diverges or NaNs, won't train at all, stalls or plateaus above the expected loss, converges to a wrong or biased answer, is worse than a reference implementation, works at small scale but fails at production scale, or fails a QA/validation recheck. Also for explicit symptoms — exploding, NaN/inf, zero, or subtly wrong gradients, suspected wp.Tape/backward issues, gradcheck failures — but users usually describe only the surface symptom ("the sim explodes", "the fit gets dragged toward outliers") without mentioning problems, or autograd issues in other frameworks without Warp.
> Evaluate whether an existing hot path is a credible NVIDIA Warp candidate. Use for irregular or spatial queries, particle or geometry simulation, branch-heavy loops, many small launches, host fallbacks, or large intermediates. CPU-only code and absent GPU dependencies are normal unless NVIDIA is prohibited. Exclude required cross-vendor or CPU-only deployment, vendor-lowered dense or NN layers, general Warp API questions, and already-selected Warp kernels. Contribution policy alone is not exclusion.
DALI imperative dynamic mode (`nvidia.dali.experimental.dynamic`, ndd): use when working on ndd code or migrating pipelines; skip pipeline-only tasks.
Parse, modify, validate, and patch simulator input files. Use when working with reservoir simulation input files, testing scenarios, or validating simulation configurations. This implementation supports reference format (.DATA); other simulators use different extensions (e.g., .afi, .DAT). Supports natural language modifications, keyword patching, and syntax validation.
Plot and compare simulation summary metrics. Use when visualizing time-series results, comparing multiple cases, or analyzing production performance. Supports single and multi-metric plots, case comparisons, and automatic metric keyword resolution.
Retrieve information from simulator manual and example DATA files. Use when answering keyword format questions, syntax queries, or when looking up official documentation and working examples. Essential for understanding keyword definitions, parameter tables, and concrete usage patterns.
Read and analyze simulation binary output files. Use when extracting summary data, grid properties, or running flow diagnostics (time-of-flight, tracer, allocation, F-Phi, Lorenz) from completed simulations.
Run, monitor, and control simulations. Use when executing simulations, checking progress, or stopping running simulations. Supports foreground and background execution, progress monitoring, and process management.
Query and browse evaluation results stored in MLflow. Use when the user wants to look up runs by invocation ID, compare metrics across models, fetch artifacts (configs, logs, results), or set up the MLflow MCP server. ALWAYS triggers on mentions of MLflow, experiment results, run comparison, invocation IDs in the context of results, or MLflow MCP setup.
>- Inspect Hugging Face decoder layers on meta tensors and plan or run per-rank BF16, FP8, and NVFP4 GEMM or fused-MoE microbenchmarks with the bundled scripts and a local FlashInfer checkout. Use when choosing a model, GPU, TP, EP, or M/token-concurrency sweep; deriving common fused QKV and gate/up shapes without loading checkpoint weights; or using FlashInfer benchmark utilities. Do not use for end-to-end server throughput or request latency.
Establish baseline-vs-candidate evaluation plans, delegate missing evaluations, compare validated results, and decide quantization feasibility. Use when the user asks to compare baseline vs quantized runs, explain an accuracy drop/regression, verify whether a quantized checkpoint is acceptable, or compare NEL/MLflow evaluation outputs. Do NOT use for generic single-model evaluation without comparison intent (use evaluation), live NEL status/debugging (use launching-evals), or generic MLflow browsing without a comparison goal (use accessing-mlflow).
Deterministic end-to-end driver for day-0 quantized-checkpoint releases — chains PTQ → evaluation → comparison with enforced gates between stages (the evaluation stage deploys the checkpoint itself), and returns a publish decision (ACCEPT / REGRESSION / ANOMALOUS / INFEASIBLE). Use when the user asks to "release a model at day-0", "quantize and validate model X is within N% of baseline and tell me if it's publishable", or "run the full day-0 workflow". Do NOT use for single-stage requests — quantizing only (use ptq), serving only (use deployment), evaluating only (use evaluation), or comparing two existing runs (use compare-results).
Run commands inside a remote Docker container via the file-based command relay (tools/debugger). Use when the user says "run in Docker", "run on GPU", "debug remotely", "run test in container", "check nvidia-smi", "run pytest in Docker", or needs to execute any command inside a Docker container that shares the repo filesystem. Requires the user to have started server.sh inside the container first.
Serve a quantized or unquantized LLM checkpoint as an OpenAI-compatible API endpoint using vLLM, SGLang, or TRT-LLM. Use when user says "deploy model", "serve model", "start vLLM server", "launch SGLang", "TRT-LLM deploy", "AutoDeploy", "benchmark throughput", "serve checkpoint", or needs an inference endpoint from a HuggingFace or ModelOpt-quantized checkpoint. Do NOT use for quantizing models (use ptq) or evaluating accuracy (use evaluation).
> Add a new model to the EAGLE3 offline pipeline. Generates an hf_offline_eagle3.yaml launcher config for a new model checkpoint, choosing the right hidden state dump backend (TRT-LLM / HF / vLLM) and GPU configuration. Use when user wants to run EAGLE3 on a model that does not yet have a YAML in tools/launcher/examples/ or asks how to configure the pipeline for a new checkpoint.
> Review EAGLE3 pipeline experiment logs from the launcher's experiments/ directory. Summarizes pass/fail status for all 4 tasks, diagnoses failures with root causes and fixes, and flags warnings. Use when the user asks to review job logs, check experiment results, or diagnose why a specific task failed.
> Triage a failed EAGLE3 pipeline run. Identifies which step failed (data synthesis, hidden state dump, training, or benchmark), diagnoses root cause from logs, and suggests fixes. Use when user reports an EAGLE3 pipeline failure or asks why a specific step failed. Also helps debug new model support issues.
> Validate that an EAGLE3 pipeline run completed successfully end-to-end. Checks all 4 steps produced expected artifacts, verifies acceptance rate meets threshold (>= 2.1), and produces a summary report. Use when user wants to verify a pipeline run or check benchmark results.
Evaluates accuracy of quantized or unquantized LLMs using NeMo Evaluator Launcher (NEL). Triggers on "evaluate model", "benchmark accuracy", "run MMLU", "evaluate quantized model", "run nel". Handles deployment, config generation, and evaluation execution. Not for quantizing models (use ptq), deploying/serving models (use deployment), or comparing completed baseline-vs-quantized results (use compare-results).
Run, monitor, analyze, and debug LLM evaluations via nemo-evaluator-launcher. Covers running evaluations, checking status and live progress, debugging failed runs, exporting artifacts and logs, and analyzing results. ALWAYS triggers on mentions of running evaluations, checking progress, debugging failed evals, analyzing or analysing runs or results, run directories or artifact paths on clusters, Slurm job issues, invocation IDs, or inspecting logs (client logs, server logs, SSH to cluster, tail logs, grep logs). Do NOT use for creating or modifying evaluation configs.
Monitor submitted jobs (PTQ, evaluation, deployment) on SLURM clusters. Use when the user asks "check job status", "is my job done", "monitor my evaluation", "what's the status of the PTQ", "check on job <slurm_job_id>", or after any skill submits a long-running job. Also triggers on "nel status", "squeue", or any request to check progress of a previously submitted job.
>- Use when the user asks to "quantize a model", "run PTQ", "post-training quantization", "NVFP4 quantization", "FP8 quantization", "INT8 quantization", "INT4 AWQ", "quantize LLM", "quantize MoE", "quantize VLM", or needs to produce a quantized HuggingFace checkpoint from a pretrained model using ModelOpt. Do NOT use for multi-candidate recipe exploration or optimization (use quant-recipe-search).
>- Run explicitly requested ModelOpt Quantization-Aware Distillation (QAD) on Slurm through Megatron Bridge to recover a measured BF16-to-PTQ accuracy gap. Use only when the user explicitly asks for QAD, including its topology, data preparation, Slurm launch, resume, checkpoint export, or recovery decisions.
>- Use when the user asks to find, search for, or optimize the best quantization recipe for a model, including direct requests like "find the best quantization choose compute-vs-memory success metrics, select ModelOpt recipe baselines, design AutoQuant/manual recipe deltas, interpret sensitivity, and decide next candidates. Do NOT use for a single known PTQ recipe run (use ptq), serving (use deployment), creating/running evals (use evaluation or launching-evals), monitoring jobs (use monitor), MLflow browsing (use accessing-mlflow), or comparing completed baseline-vs-candidate scores only (use compare-results).
Cherry-pick merged PRs labeled for a release branch into that branch, then open a PR and apply the cherry-pick-done label. Use when asked to "cherry-pick PRs for release/X.Y.Z", "pick PRs to release branch", or "cherry-pick labeled PRs".
Official NVIDIA-authored guidance for navigating PhysicsNeMo — pick the model, datapipe, or example for a SciML/AI4Science task (surrogates, forecasting, downscaling, physics-informed, inverse, generative). Points at existing files via live repo search; never writes code. Do NOT use for installation or environment setup, training-loop or other code authoring/scaffolding, contributor/CI/packaging questions, repo-specific questions in physicsnemo-sym/-cfd/-curator, or general (non-physics) ML/PyTorch.
Use when the user wants to search, query, extract, transcribe, describe, quote, filter, or aggregate across documents — PDFs, scanned forms / images (`.jpg` `.png` `.tiff`), Office (`.docx` `.pptx`), text (`.html` `.txt`), audio (`.mp3` `.wav` `.m4a`), or video (`.mp4` `.mov`). Prefer this over native Read / Grep for multi-file or non-PDF corpora. Not for: editing files, web browsing, single-file plain-text lookups, fine-tuning.
Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and multi-GPU DataFrame workloads.
| Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.
| Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.
>-
Install and verify cuPyNumeric for Python — requirements, commands, verification. Source builds are out of scope.
Pre-migration readiness assessor for porting NumPy to cuPyNumeric. Use BEFORE substantial porting work begins when the user asks whether code will scale on GPU, whether they should migrate to cuPyNumeric, which NumPy patterns transfer cleanly, what must be refactored before porting, or mentions pre-port assessment, scaling analysis, or refactor planning. Inspect the user's source code, look up NumPy usage, cross-reference the cuPyNumeric API support manifest, and distinguish distributed-scaling-friendly patterns from blockers such as unsupported APIs, scalar synchronization, host round-trips, Python/object-heavy control flow, shape/data-dependent branching, and in-place mutation hazards. Produce a verdict of READY, LIGHT REFACTOR, SIGNIFICANT REFACTOR, or NOT RECOMMENDED, with concrete refactor pointers.
Load a sharded, on-disk dataset (sharded .npy, Parquet/Arrow, raw binary, sharded HDF5, custom layouts) into a distributed cuPyNumeric ndarray via a manual partition + leaf @task launch with CPU/OMP/GPU variants. Use when no single-call loader fits, including when per-shard row counts differ across files. Prefer cupynumeric.load or legate.io.hdf5.from_file when they apply.
Use when the user wants to create a dataset, generate synthetic data, or build a data generation pipeline.
Used for extracting selected metadata from one DICOM file and flagging standard-tag PHI presence. Not for anonymization or clinical use.
Used for header-only preflight of one DICOM series folder before conversion or inference. Not for de-identification or clinical clearance.
Used for converting one CT DICOM series folder to a HU NIfTI volume with affine evidence. Not for multi-frame DICOM or clinical use.
Stage 2 of the Clinical ASR Flywheel. Use when curating clinical terms, tagging IPA, and synthesizing a NeMo manifest. NOT for scoring (use /digital-health-clinical-asr-eval).
Stage 3 of Clinical ASR Flywheel. Score a NeMo manifest, produce the five-section KER leaderboard (by-ipa_source diagnostic). Not for ASR auth (/riva-asr).
Stage 4 of the Clinical ASR Flywheel. Use when priority KER is above 0.3 to run stock NeMo SFT on Parakeet TDT v2 and offline cycle N+1 re-eval. NOT for generic word boosting (use /finetune-asr).
Stage 1 of Clinical ASR Flywheel. Use when bootstrapping a cycle: NVCF+MW disclosure, NVIDIA_API_KEY check, deps install, TTS+ASR smoke test.
> Use this skill when the user is doing hands-on DOCA AES-GCM work on a BlueField DPU or ConnectX NIC — configuring `doca_aes_gcm_task_encrypt` / `_task_decrypt`, querying `doca_aes_gcm_cap_*` for per-key-type (only `DOCA_AES_GCM_KEY_128` / `_256` — AES-192 not supported) and per-task support, sizing plaintext against the max-buf cap, setting source / destination mmap permissions, validating with a NIST GCMVS or RFC 5288 vector, or debugging DOCA_ERROR_* including the security-critical tag-verification-failed outcome on decrypt. Trigger even when the user does not explicitly mention "DOCA AES-GCM" or IO_FAILED", "auth tag isn't verifying", "NOT_PERMITTED on my encrypt buffer", "is AES-192-GCM on this BlueField" (no), or "encrypted record came back tampered". Refuse and route elsewhere for non-GCM AES modes (CBC / CTR / XTS — CPU OpenSSL), key management (KMS / HSM / rotation), SHA (doca-sha), or general AEAD background.
> Use this skill for hands-on DOCA Arg Parser CLI work on a shipped sample or new DOCA-using app — adding / removing / renaming flags; wiring `doca_argp_init` → register params → `doca_argp_start` → `doca_argp_destroy` in order; picking a parameter type from the full public enum (`DOCA_ARGP_TYPE_STRING`, `_INT`, `_BOOLEAN`, `_DEVICE`, `_DEVICE_REP`, `_DOUBLE` — six values, not three); preserving the standard `--device` / `--representor` / `--json` (`-j`; real flag is `--json`, NOT `--json-config`) / `--sdk-log-level` surface; or debugging `DOCA_ERROR_BAD_STATE` / `INVALID_VALUE` / `NOT_SUPPORTED` / `IO_FAILED` from `doca_argp_*`. Trigger on implicit use getopt here", "BAD_STATE registering a new param", "my JSON config key is rejected", or "my sample's --json is ignored". Refuse and route elsewhere for variadic-flag / subcommand / shell-completion features, DOCA Core context, or DOCA Log internals.
> Use this skill when the user is deploying or operating the DOCA Argus Service — the packaged BlueField-side runtime-security container that watches the BlueField and attached host for suspicious activity, integrity violations, and operational anomalies, and forwards findings to a SIEM (Splunk / ELK / Sentinel / syslog). Covers the four-axis config (detection policy, forwarding, sampling, host coverage), running the NGC container on BlueField Arm, and wiring the forwarder. Trigger even without "DOCA Argus" by arrive", "false-positive flood in Splunk", or "runtime security on a fleet of BlueField-3s". Refuse and route elsewhere for installing DOCA, SIEM-side ingest stanzas, pre-baked detection-rule packs, and metrics observability (DOCA Telemetry). Argus is NVIDIA's currently- promoted runtime-security framework, superseding the older App Shield library; name it first for new runtime-security work.
> Use this skill for launching, supervising, debugging, OR platform lifecycle on a BlueField — BFB install, RShim/TMFIFO, host PF rebind, post-BFB recovery — taking a DOCA-linked binary to a healthy run directly on hardware (host x86 + BlueField NIC over PCIe, or BlueField Arm bare-metal). No container, no kubelet. Covers launch mode (direct, tmux, systemd), PCI/NUMA/ CPU/IRQ binding, co-tenant isolation (cgroup-v2/netns/numactl), a seven-layer error taxonomy, and a six-state BlueField lifecycle classifier. Trigger even when user does not say "bare-metal" — implicit phrasings include "binary exits 1 right after launch", "systemd keeps restarting it", "no matching device on the BF", "bfb-install exited 0 but DPU is dead", "ping 192.168.100.2 works but ssh fails", "host PFs aren't showing netdevs". Destructive firmware burn / mlxconfig set requires explicit confirmation via doca-hardware-safety; containers, library APIs, env prep, and build use other skills.
> Run `doca_bench` (DOCA 2.7.0 or newer) to measure throughput, bulk latency, precision latency, or maximum bandwidth for RDMA, Compress, AES-GCM, SHA, DMA, EC, Ethernet, Comch, or GPUNetIO on a host or BlueField Arm. Use it to discover enabled benchmark libraries, capture a reproducible command/version/device/environment baseline, compare stable runs against a declared tolerance, or diagnose configuration, device-binding, workload-precondition, and measurement failures. Trigger for requests such as measuring BlueField compression speed, NIC RDMA throughput, crypto latency, or a pre-upgrade baseline. Do not use for application end-to-end timing, custom benchmark code, DOCA installation, or binary patches.
> Use this skill when the operator is authoring, building, loading, or debugging a custom doca-bench plug-in — a versioned shared library with DOCA_EXPERIMENTAL-marked C entry points that doca-bench loads to measure a workload class its built-in modes do not cover, with doca_bench_cuda as the shipped reference exemplar. Trigger even when the user does not say "doca-bench-extension" or "doca_bench_cuda" — typical implicit phrasings include "no built-in doca-bench mode fits my workload", "how do I benchmark a CUDA GPUNetIO RX/TX kernel", "doca-bench cannot find or load my custom .so", "extension exported symbols do not match what the parent expects", "soversion mismatch after a DOCA upgrade", or "my GPU kernel hangs because stop_flag was never set". Refuse and route elsewhere for questions about which built-in doca-bench mode to pick, DOCA GPUNetIO programming semantics, CUDA toolkit installation, or contributor work on in-tree extensions — those belong to other skills.
> Use this skill for BlueField-3 (BF3) day-1 platform bring-up via the DPU over RShim with bfb-install from the host, the host-to-DPU TMFIFO management channel (tmfifo_net0, the 192.168.100.x convention), RShim daemon state and console-over-rshim, DPU mode selection (DPU/embedded-function vs separated-host/NIC mode) via mlxconfig, post-BFB recovery, a six-state BlueField-state classifier, and verifying the install (cat /etc/mlnx-release plus version checks). Trigger even when the user does not say "BF3" — typical phrasings include {push a BFB to my BlueField-3}, {bfb-install exited 0 but the DPU never came back}, {ping 192.168.100.2 works but ssh fails}, or {is DOCA on the host or the Arm side?}. BFB reflash, mlxconfig set, mode changes, and confirmation and load doca-hardware-safety. App launch, container deploy, env install, and the BF4 BMC-Redfish path route elsewhere.
> operations (PLDM firmware burns, ISO reflashes, power cycles, BMC factory resets) that can brick firmware, corrupt boot media, or cause outages — a maintenance window and rollback plan are required, and every mutating step is governed by doca-hardware-safety, loaded alongside. Use this skill for BlueField-4 (BF4) day-1 platform ISO onto the DPU (Grace, the Arm complex) over UEFI HTTP Boot, PXE, or Redfish Virtual Media; the PLDM firmware-update flow (BMC, NIC firmware, SBIOS, ERoT) via the Redfish UpdateService and pldmtool; and a Grace Ubuntu image with optional cloud-init. Trigger my new BlueField-4}, {the BlueField ISO will not boot over HTTP from the BMC}, {attach BF4 virtual media via Redfish}, {BF4 firmware Task stuck at Running}. BF3 bring-up, application launch, and library APIs belong to other skills.
> Use this skill when the user wants to invoke the read-only doca_caps CLI to ask what DOCA sees on this host — listing DOCA devices and PCIe addresses, listing representor devices, asking which DOCA libraries are available on the current OS, checking per-device per-library capabilities, scoping output to a specific PCIe address, or capturing a side-effect-free capability snapshot for a debug session or install smoke-test. Trigger even when the user does not explicitly mention "doca_caps" or "capabilities print tool" — typical implicit phrasings include "what does DOCA actually see on this box", "is my BlueField PF visible to DOCA", "is Flow available on my RHEL host", "enumerate VF representors for RDMA, is the tool broken". Refuse and route elsewhere for DOCA installation, library-internal capability matrices (Flow pipe creation, RDMA verbs features), streaming telemetry / DTS, or modifying the shipped binary — those belong to other skills.
> Use this skill to deploy and operate a CollectX (clx) based DOCA telemetry collector on a host or BlueField — wiring providers / counters into the collector, running the collection daemon, and shaping its exporters (Prometheus pull, Fluent Bit push, NetFlow, file / IPC) so the metrics actually leave the box. Trigger even when {collector emits nothing downstream}, {add a provider to the clx collector}, {turn on the Prometheus endpoint}, {ship counters to Fluent Bit from the DPU}, {daemon starts but no schema rows appear}. This skill owns the CollectX collection mechanism plus the operator's own doca-telemetry / doca-telemetry-exporter usage; it ROUTES the productized DOCA Telemetry Service (DTS) to public docs (AGENTS.md Non-goal #7), the reader API to doca-telemetry, and the publisher API to doca-telemetry-exporter. Refuse to invent clx symbols, provider names, schema fields, flags, or config paths — describe the class and route to the live source.
> Use this skill when the user is doing hands-on DOCA Comch work on a host + BlueField pair — bringing up host ↔ DPU PCIe control-plane messaging, picking server (DPU) vs client (host) roles, choosing slow-path send-task / recv-callback vs fast-path producer / consumer, querying max-msg-size or max-clients capabilities, registering connection callbacks, or debugging DOCA_ERROR_* returns from the Comch API. Trigger even when the user does not explicitly mention "DOCA Comch" or "Comm Channel" (renamed in DOCA 2.5) — typical implicit phrasings include "send a control message from host to BlueField over PCIe", "DPU can't see the host representor", "DOCA_ERROR_NOT_PERMITTED on server_create", "DOCA_ERROR_AGAIN on task_send submit", "connect callback never fires", or "stream bulk data from a host driver to a DPU agent". Refuse and route elsewhere for installing DOCA itself, BFB / firmware bring-up, non-Comch DOCA libraries, or deploying Comch apps at scale — those belong to other skills.
> Use this skill to enumerate host↔DPU DOCA comch (formerly Comm Channel) servers and connections via the shipped doca_comm_channel_admin binary — listing comch-capable devices and decoding the per-device server / connection table (server name, PID, in-use / max, PCIe address). The shipped binary is a SINGLE-SHOT SCAN-AND-PRINT tool with no registered arguments — NO list / inspect / drain / restart subcommands; one inventory pass over every comch-capable doca_dev on this side. Channel reset / drain / restart go to doca-comch (program side), doca-setup / doca-hardware-safety (driver reload), or BFB / RShim — NOT to this binary. Trigger on phrasings like "list comch servers", "which channels are active on this BlueField", or "verify admin tool sees same channel as program." Refuse and route elsewhere for the comch programming API, library install, protocol design, channel reset, or general orientation.
> Use this skill whenever the user is doing hands-on DOCA programming on a BlueField DPU or ConnectX NIC and needs the foundation primitives every per-library context rests on — walking the doca_ctx lifecycle, discovering doca_dev / doca_devinfo and gating on doca_*_cap_* before trusting a feature, wiring doca_mmap / doca_buf_inventory / doca_buf for zero-copy I/O across libraries, driving doca_pe for completions, or DOCA Log's two-tier (--sdk-log-level vs app-side) model. Trigger even when the user does not say "DOCA Common" — typical implicit phrasings include "my tasks submit but nothing completes", "DOCA_ERROR_BAD_STATE from doca_ctx_start", "--sdk-log-level does nothing for my DOCA_LOG_DBG lines", "share a buf between doca_dma and doca_rdma", or "crashes far from the offending line". Refuse and route elsewhere for per-library questions in isolation (load doca-flow / doca-rdma / doca-eth alongside), installing DOCA (doca-setup), or doc lookup (doca-public-knowledge-map).
> Use this skill for hands-on DOCA Compress programming on a BlueField DPU, ConnectX NIC, or host with DOCA — enabling compress-deflate, decompress-deflate, decompress-lz4-stream, or decompress-lz4-block tasks on a doca_compress context (the hardware supports DEFLATE both directions plus LZ4 decompress; LZ4 encode is NOT supported), sizing source / destination doca_buf against the per-task cap query, setting mmap permissions, deciding offload vs CPU zlib / zstd, validating with a round-trip smoke, or debugging DOCA_ERROR_* from a Compress call. Trigger on phrasings like "offload this gzip", "decompress incoming network data", "compress task returns INVALID_VALUE on alloc_init", "submitted a task but no completion arrives", or "decompress LZ4 on the BlueField." Refuse and route elsewhere for non-DEFLATE / non-LZ4 algorithms (zstd / Snappy / brotli), LZ4 encode (route to a CPU LZ4 library), pure mmap-to-mmap copies (doca-dma), or DOCA Core lifecycle internals.
> Use this skill when the user is hands-on deploying an in-bundle DOCA service container (Argus, DMS, Firefly, or UROM service) on a BlueField — kubelet standalone watching a static-pod manifests directory, YAML pod-spec drop, kubelet status / ENTRYPOINT logs / per-service liveness, smoke-before-bulk, and the layered error taxonomy (pod-spec, scheduling, image pull, runtime, mount, network, version, host). Trigger even when the user does not say "container deployment" — typical implicit phrasings include "how do I run my built service on the BlueField?", "where do I drop the pod-spec YAML?", "pod stuck in Pending / ImagePullBackOff / CrashLoopBackOff", "container Running but service isn't ready", "pod restart-loops after edit", or "DMS and Firefly together". Refuse and route elsewhere for per-service config schemas, DOCA install, library-API questions, external NVIDIA services (BlueMan, HBN, SNAP, Virtio-net), or full Kubernetes-cluster ops — those belong to other skills.
> Use this skill when the user is debugging any DOCA symptom — a build that won't compile, a link step that can't resolve a doca_* symbol, a runtime call returning DOCA_ERROR_*, a silent service or tool, or a stack trace / valgrind / core dump — and needs the layered ladder (install → version → build → link → runtime → program → driver), verbosity controls (--sdk-log-level, DOCA_LOG_LEVEL, the doca-{lib}-trace flavor), container-debug constraints, or how to capture state for a Developer Forum post. Trigger even when the user does not say "DOCA debug" — implicit phrasings include "undefined reference to doca_*", "how do I get more logs", "packets aren't reaching the wire", "doca_caps returned nothing", or "hugepages empty in the container". Refuse and route elsewhere for library-specific debug (Flow pipe trace, RDMA QP, Comch stats), env-class pkg-config or hugepages symptoms, the DOCA_ERROR_* taxonomy and lifecycle interpretation, and performance or incident-response work — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA Device Emulation on a BlueField DPU — exposing a custom emulated PCIe device the host sees as a real peripheral while DPU-side code runs the backend, picking the sub-library (PCI Generic, virtio-net, virtio-fs), wiring the per-sub-library Core context plus doorbell / DMA primitives, querying `doca_devemu_*_cap_*`, or debugging DOCA_ERROR_* from a `doca_devemu_*` call. Trigger even when the user does not say "devemu" — typical implicit phrasings include "expose a custom PCIe device from BlueField to the host", "host should see a virtio NIC backed by my DPU code", "lspci does not show my emulated device", "device enumerated but no driver binds", "DPU sees nothing when host kicks the queue", or "virtio feature negotiation failed at bind". Refuse and route elsewhere for the packaged DOCA SNAP / Virtio-net Services, host-side virtio kernel drivers, backend body design, or standard BlueField NIC behavior — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA DMA programming — bringing up a doca_dma context, configuring the single doca_dma_task_memcpy task type, sizing buffers via the doca_dma_cap_task_memcpy_* queries, setting LOCAL_READ_ONLY / LOCAL_READ_WRITE permissions on source / destination doca_mmap regions (plus doca_mmap_export_* for cross-peer copies), driving the progress engine, or debugging DOCA_ERROR_* returns. Trigger even when the user does not explicitly mention "DOCA DMA" or "doca_mmap" — typical implicit phrasings include "memcpy host buffer to BlueField without using the CPU", "offload a bulk copy to the DPU", "copy returns NOT_PERMITTED on first submit", "buffer too big for one DMA task", "task submitted but no completion", or "scatter-gather copy between two memory regions". Refuse and route elsewhere for cross-network copies (DOCA RDMA), producer/consumer messaging (DOCA Comch), DOCA Core / progress-engine internals, or DOCA install — those belong to other skills.
> Operate NVIDIA DOCA Management Service (`dmsd` + `dmspe`) on a authentication, configure `-allowed_users` and `dmsgroup`, use gNMI Get/Set/Subscribe, run supported gNOI workflows, and debug frontend/backend failures. Trigger even without "DMS" for "manage a remote BlueField over gRPC", "gNOI reboot from orchestrator", or fleet-management requests. are destructive and require target-bound explicit confirmation; never invoke them speculatively. Route installation and library/API build questions elsewhere, and route turnkey aggregation to the externally-productized DOCA Telemetry Service.
> Use this skill when the user is doing hands-on DOCA DPA host-side work on a BlueField — creating the `doca_dpa` Core context, loading a DPACC-compiled DPA app image (`doca_dpa_app`), creating DPA threads, launching kernels via `doca_dpa_kernel_launch_update_*`, draining `doca_dpa_completion`, running `doca_dpa_cap_*` discovery, choosing between the DPA comm component (inter-DPA messaging) and the DPA verbs component (in-kernel RDMA), or debugging `DOCA_ERROR_*` from `doca_dpa_*`. Trigger even without "DOCA DPA" or "Data-Path hangs, no completion", "DOCA_ERROR_DRIVER on launch", "DOCA/DPACC version skew", or "does this BlueField expose a DPA". Route elsewhere for DPA-side kernel programming itself, DPACC compiler internals, host↔DPU messaging (doca-comch), host-side RDMA (doca-rdma), and GPU-initiated networking (doca-gpunetio).
> Use this skill when the user runs doca_dpa_hl_tracer to capture/decode DPA-side traces at the programming-events layer (kernel entry/exit, sync points, comm primitive calls, RDMA WR submission, completion drain) — picking TRACE vs CRIT, tuning the JSON config (file-size limits + file_size_limit_policy, thread priorities/cores), decoding against the matching DPA-side ELF, or diagnosing empty/noisy captures. Trigger even when the user does not explicitly mention "DOCA DPA tracer" or "high-level tracer" — typical implicit phrasings include "DPA kernel returns wrong result but host completions look clean", "kernel-entry to first-comm latency is huge", "RDMA WR to drain gap on the DPA", "trace file truncated mid-run", "TRACE doubled my DPA latency", or "tracer wrote a file but parser shows zero events". Refuse and route elsewhere for writing DPA kernels, DPA-Comms/DPA-Verbs programming, raw per-cycle DPA profiling, host-side doca-dpa debugging, or production DPA telemetry — those belong to other skills.
> Use this skill when the user has an existing DPDK application and is adding DOCA capabilities in-place — most commonly DOCA Flow hardware binding a DPDK port id to a `doca_dev` (`doca_dpdk_port_probe` / `doca_dpdk_port_as_dev`), converting `rte_mbuf` ↔ `doca_buf`, querying `doca_dpdk_cap_is_rep_port_supported`, or debugging `DOCA_ERROR_*` from a bridge call. Trigger even without "DOCA DPDK visible to DOCA", "the bridge loads but every operation returns errors", "pkg-config --exists doca-dpdk-bridge fails", or "DOCA_ERROR_NOT_FOUND on port registration". Route elsewhere for fresh DOCA-native packet I/O (doca-eth), flow-rule programming (doca-flow), DOCA or DPDK install (doca-setup), or RDMA data movement (doca-rdma).
> Use this skill when the user is doing hands-on DOCA Erasure Coding programming on a BlueField DPU, ConnectX NIC, or host — bringing up a doca_ec context, picking among the create / recover / update tasks, choosing matrix type / N / K / block size, querying doca_ec_cap_* before sizing, setting doca_mmap src/dst permissions, or debugging DOCA_ERROR_* returns from doca_ec_task_*. Trigger even when the user does not name "DOCA Erasure Coding" or "Reed-Solomon" — typical implicit phrasings include "one data block changed, how do I refresh parity without re-encoding", "a disk failed and 2 parity blocks are gone, can I rebuild", "RAID-6 resilience across 12 disks", "my doca_ec_task_create returns NOT_PERMITTED", or "is this N+K layout still recoverable". Refuse and route elsewhere for non-Reed-Solomon codes (fountain / LDPC / raptor), pure-replication designs, network FEC, or other DOCA accelerator libraries (SHA / Compress / AES-GCM / DMA) — those belong to other skills.
> Use this skill for hands-on DOCA Ethernet packet-queue work on a BlueField DPU or ConnectX NIC — bringing up a `doca_eth_rxq` or `doca_eth_txq` on a port / representor / SF, picking among the four `enum doca_eth_rxq_type` values (`_REGULAR` / `_CYCLIC` / `_MANAGED_MEMPOOL` / `_SHARED_MEMPOOL`), sizing burst or scatter-gather length against the `_cap_*` queries, submitting `doca_eth_txq_task_send` / `_lso_send` (carrying packet `doca_buf`s — no `doca_eth_frame` struct exists), or debugging DOCA_ERROR_* from an Ethernet call. Trigger on arrive", "send-task returns AGAIN at line rate", "which queue type for fixed-MTU ingress", "device open fails without sudo", or "is L3 checksum offload available here". Refuse and route elsewhere for installing DOCA, flow-rule / steering programming, host↔DPU control messaging, or RDMA data movement.
> Use this skill when the user is operating the DOCA Firefly Service container on BlueField — picking the four PTP configuration axes (role / profile / domain / interface), wiring the BlueField PHC + host follower + consumer workload pairing, deciding whether PTP-grade time is even needed (vs. chrony / NTP), or debugging a Firefly deployment where PTP isn't syncing or the host clock isn't following. Trigger even when the user does not explicitly mention "DOCA Firefly" or "PTP" — typical implicit phrasings include "container green but PTP never advances past LISTENING", "Firefly says synced but the host clock still drifts", "sync acquired but offset is tens of microseconds", "my Rivermax SMPTE workload needs PTP", or "is chrony good enough". Refuse and route elsewhere for installing DOCA, host-side chrony / ptp4l config bodies, PTP topology / boundary-clock design, building DOCA apps that read the disciplined PHC, or other DOCA services (DMS, Flow-Inspector, HBN) — those belong to other skills.
> define match/action pipes, initialize ports and representors, choose forwarding targets, validate pipes before hardware programming, read counters, match the Flow version to the installed DOCA release, and diagnose Flow API errors. Trigger on DOCA packet steering, classifier, representor, rule-matching, hairpin, or 5-tuple-to-queue questions even when "DOCA Flow" is not named. Route plain DPDK `rte_flow`, kernel TC, OVS, BFB bring-up, and DPU OS installation elsewhere. DPU OS installation is destructive and always requires explicit confirmation.
> Use this skill when the user is invoking doca_flow_dpa_perf on DPA-capable hardware (ConnectX-7 minimum supported, ConnectX-8 recommended, or BlueField-3) to measure rule update / disable rates on the DPA-offloaded DOCA Flow path — picking the active / passive device split, choosing workload-shape axes (burst, queue, completion threshold, workers, hash pipe algo, PSL tables), or reading Kops/sec iteration stats and the optional self-test. Trigger even when the user does not explicitly mention "doca_flow_dpa_perf" or "DPA Provider" — typical implicit phrasings include "how fast can the DPA program path-selector entries", "baseline rule-update rate on ConnectX-8", "tool reports zero ops on my BlueField", "self-test sentinel never shows on tcpdump", or "is my BlueField-2 DPA-capable". Refuse and route elsewhere for the host / DPU-CPU Flow path (doca-flow-perf), Flow pipeline tuning (doca-flow-tune), writing doca-flow / doca-dpa applications, or DOCA install — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA Flow DPA Provider work — exporting a `doca-flow` pipe or external resource (index-selector/memory) into BlueField DPA address space so a DPACC-built kernel can read counters, mutate hash-pipe entries, and update/read memory or index-selector resources inline with Flow. Covers per-port `doca_flow_dpa_ctx`, three queue types (general/resources-write/resources-read), the order-sensitive export handshake (`_export_prepare` → add entries → `_export` → `_get_device_addr`), and DPA-side device API. Trigger even when the user does not say "DOCA Flow DPA Provider" — implicit phrasings include "DPA kernel never sees entries in the exported pipe", "BAD_STATE from `_pipe_export`", "how do I disable a hash entry from a DPA kernel", "DPA memory read returns no value", or "DPA-side post keeps returning AGAIN". Refuse and route elsewhere for `doca-flow` pipe construction, generic host-side DPA (`doca-dpa`), or DPA-side kernel-writing — those belong to other skills.
> knob on the binary — transport security must come from external infrastructure (e.g. an mTLS proxy / sidecar) on a trusted segment. Use this skill when bringing up, configuring, hardening, or debugging `doca_flow_grpc` — the DOCA-shipped gRPC remote-control surface in front of `doca-flow` that lets non-C++ clients (Python, Go, Rust, Java) program Flow pipes and entries over RPC instead of linking `libdoca_flow.so` directly. Trigger even when the user doesn't say 'doca-flow-grpc-server' or 'gRPC' — e.g. 'program Flow rules from Python on another host', 'remotely configure pipes on the BlueField', 'client times out connecting to the Flow server', 'where is the .proto for Flow', 'UNAUTHENTICATED / FAILED_PRECONDITION on a Flow RPC'. Route elsewhere for the underlying doca-flow API, generic gRPC tooling (protoc, language bindings), or DOCA install / BFB bring-up.
> Use this skill when the user is measuring the host or DPU-CPU control-plane rate of a DOCA Flow pipeline with doca_flow_perf — picking a JSON policy from configs/, choosing the DPDK or DOCA backend, running the single-iteration smoke then the iterative eval loop, interpreting per-iteration CPU cycles and num_pushed / num_failed, or capturing the four-tuple (DOCA version, BlueField/firmware, JSON policy, worker/queue/burst config) that makes a Kops/sec number defensible. Trigger even when the user does not explicitly mention "doca-flow-perf" — typical implicit phrasings include "how many rules per second can my BlueField insert", "5-tuple hairpin rule rate", "Kops/sec for steering", "flow-perf number does not match release notes", "DPDK vs DOCA benchmark", or "rule-install variance too high". Refuse and route elsewhere for optimizing a live Flow app (doca-flow-tune), the DPA-offloaded path (doca-flow-dpa-perf), dataplane throughput or latency, or library-internal pipe semantics — those belong to other skills.
> Use this skill when the user is tuning a live or captured `doca-flow` pipeline with `doca_flow_tune` — snapshotting pipe / counter / KPI state, picking a tuning axis (rule placement, resource hints / table sizing, HW-offload mode) and a matching measurement (rule-install rate, lookup latency, hardware-counter delta), running offline or online (read-only or state-changing) modes, reading the dumper CSV / analyze JSON / visualize mermaid, or applying a recommendation back into the Flow program. Trigger even when the user does not explicitly mention "doca_flow_tune" — typical implicit phrasings include "Flow rule-install rate is low on BlueField", "table sizing looks wrong for this pipe", "tune visualize step is empty", "before/after counters don't move", or "which doca-flow knob does this recommendation hit". Refuse and route elsewhere for measuring baseline numbers (doca-flow-perf, doca-flow-dpa-perf), writing the doca-flow application, DOCA install, or streaming Flow telemetry — those belong to other skills.
> Use this skill for hands-on DOCA GPI programming — wiring a GPU-Packet-Initiator context so a CUDA kernel drives RDMA queues directly from GPU memory without host CPU mediation. Covers picking GPI vs doca-gpunetio, the doca_gpi / domain / channel object model, the GPU-side handle handoff (doca_gpu_gpi_channel*), attaching GPU memory to a GPI domain, the domain and channel attribute objects, and debugging DOCA_ERROR_* from doca_gpi_* calls. Trigger even when the user does not explicitly mention "DOCA GPI" — implicit phrasings include "my CUDA kernel needs to post RDMA directly from GPU memory", "DOCA_ERROR_* from doca_gpi_gpu_channel_get", "how do I hand a GPU handle to my CUDA kernel", "how many channels can a GPI domain hold", or "GPU kernel driving RDMA without the host CPU on the path". Refuse and route elsewhere for the doca-gpunetio Send/Receive surface, the doca-rdma queue lifecycle, DPA-side initiation (doca-rdmi), or the CUDA programming model — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA GPUNetIO programming — wiring a CUDA kernel on an NVIDIA GPU to a doca-eth queue via doca_gpu_eth_rxq / doca_gpu_eth_txq, standing up the per-CUDA-device doca_gpu context, designing the persistent CUDA kernel that drains the GPU-visible queue, running the dual capability check (DOCA cap-query plus cudaGetDeviceProperties), registering cudaMalloc pools via doca_buf_arr_create_*, or debugging DOCA_ERROR_* returns from the GPUNetIO API. Trigger even when the user does not explicitly mention "DOCA GPUNetIO" or "persistent kernel" — typical implicit phrasings include "CUDA kernel reading packets directly from the NIC", "GPU-initiated networking on BlueField", "DOCA_ERROR_DRIVER on doca_gpu_create", "nvidia_peermem not loaded", "kernel-per-packet is too slow", or "which GPU supports GPU-side packet I/O". Refuse and route elsewhere for general CUDA programming, DOCA Ethernet queue bring-up, DOCA DPA, or DOCA install — those belong to other skills.
> Use this skill when the user is building, running, or interpreting the doca/tools/gpunetio_ib_write_bw client+server benchmark — a CUDA kernel on the client posts RDMA WRITE work requests through the doca-gpunetio device-side surface to measure sustained GPU-driven WRITE bandwidth on a GPU+IB-device pair. Trigger even when the user does not explicitly mention "doca-gpunetio-ib-write-bw" or "GPUNetIO" — typical implicit phrasings include "measure WRITE BW when the GPU posts the WRs", "BW swings between runs on the same flags", "is the NIC saturated or am I CPU-bound on the CUDA kernel", "meson compile fails for the GPUNetIO bw tool", "nvidia_peermem isn't picking up my GPU buffer", or "GPU-initiated WRITE throughput vs CPU-initiated perftest". Refuse and route elsewhere for general doca-gpunetio library work, DOCA install, the GPU-initiated WRITE latency analog, the CPU-initiated upstream perftest, or application-level end-to-end throughput — those belong to other skills.
> Use this skill when the user is measuring GPU-kernel-initiated RDMA WRITE latency through doca-gpunetio — building and running the `gpunetio_ib_write_lat` client + server pair under `doca/tools/gpunetio_ib_write_lat/`, checking GPU-NIC pairing, reading the half-iter / full-iter / CUDA-side usec columns, characterizing median / p99 / jitter for a real-time control loop, picking GPUNetIO vs GPI vs CPU-initiated `perftest`, or weighing the latency-vs-batching trade-off. Trigger even without 'GPUNetIO' or CUDA kernel post a WRITE', 'p99 RDMA latency on H100 + ConnectX', 'kernel-launched WR tail latency', or 'compare GPU-init vs CPU-init perftest'. Route elsewhere for bandwidth runs (doca-gpunetio-ib-write-bw), the GPI surface (doca-gpi), library debugging (doca-gpunetio), or DOCA install.
> Use this skill whenever the agent is about to recommend or apply a change that touches DPU / NIC hardware state on a live system — mlxconfig firmware-parameter write, NIC firmware burn, BFB reflash, NIC ↔ DPU mode flip, SR-IOV or device-emulation slot enable, kernel boot-parameter change (IOMMU, hugepages, VFIO), PCIe rebind / rescan / link-state flip, or BlueField cold reboot. Wraps the change in pre-flight inventory, OOB reachability, a maintenance window, the mlxconfig cold-power-cycle rule, replica rehearsal, and rollback. Trigger even when the user does not say "hardware safety" SR-IOV and reboot", "burned firmware but mlxconfig shows old value", "reflashed BFB and lost representors", "reflash during business hours", "vendor says this is one-way". Refuse for general DOCA orientation (doca-public-knowledge-map), install or env debug (doca-setup), and program-side debug (doca-debug, doca-programming-guide) — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA Management programming against BlueField / ConnectX devices — standing up a management or representor context (doca_mgmt_dev_ctx / doca_mgmt_dev_rep_ctx), querying device caps (data-direct, caps-general), toggling congestion-control global status, modifying diagnostics-data, setting ICM quotas, or issuing a raw firmware command via doca_mgmt_raw_cmd with the right scope (CONFIGURATION / DEBUG_READ_ONLY / DEBUG_WRITE / DEBUG_WRITE_FULL). Trigger even when the user does not say "DOCA Management" — typical implicit phrasings include "fleet tool that walks every BlueField and reads device state", "toggle data-direct on a VF", "set an ICM quota per representor", "send a raw firmware command from C", "DOCA_ERROR_IO_FAILED from raw_cmd", or "fwctl ioctl is failing". Refuse and route elsewhere for mlxconfig direct operation, BFB / firmware reflash, streaming telemetry, doca_caps CLI snapshots, or DOCA install itself — those belong to other skills.
> Use this skill when the user is doing hands-on host-side DOCA PCC work to load a CUSTOM Programmable Congestion Control algorithm onto a BlueField DPU — creating per-port `doca_pcc` contexts, loading a `dpacc`-compiled `doca_pcc_app` onto the `doca_dev` for the RoCE-bearing port, parameterizing it, walking triple-axis capability discovery (DOCA cap-query + DPA-capable BlueField + firmware custom-PCC slot enabled), or debugging `DOCA_ERROR_*` from `doca_pcc_*`. Trigger even without explicit "DOCA PCC" phrasing — implicit forms include "loading my own congestion control onto a BF port", "DOCA_ERROR_NOT_PERMITTED on algorithm load", "DOCA_ERROR_DRIVER when I attach my custom algorithm", "my custom rate-update isn't affecting RoCE traffic", or "load succeeds but no on-wire change". Refuse and route elsewhere for DPA-side algorithm-body design, the `pcc_counters` CLI, default factory PCC in ConnectX firmware, or setting up the RDMA / RoCE traffic — those belong to other skills.
> Use this skill when the user is invoking the DOCA PCC Counters tool — the `pcc_counters.sh` bash script under the DOCA tools directory — to arm and read the fixed firmware/hardware PCC (Programmable Congestion Control) diagnostic counters (CNP, RTT, WRED-drop, etc.) on a ConnectX / BlueField device via mst + the mlx5 debugfs `diag_cnt` interface. The script takes two positional args — `set | query` and an mst device path — with no `--help` or subcommands. Trigger even without "pcc_counters.sh" or "PCC "PCC counter stuck at zero", "the script says Bad Device", or "is congestion control dropping packets on this port?". Route elsewhere for writing a custom PCC algorithm (doca-pcc), factory firmware PCC config, DOCA install, or fleet-wide CC tuning.
> Use this skill when the user is doing hands-on deployment, tuning, or evaluation of the DOCA-shipped Zero-Touch RoCE RTT-based Congestion Control (ZTR RTTCC) reference algorithm on a BlueField-3 DPA — wiring `doca_pcc_dev_ztr_rttcc_algo` into the shipped DOCA PCC sample, picking a variant (vanilla / PM / RX-rate / multipath / window-probeless) at DPACC build time, tuning host-set parameters, or diagnosing `DOCA_PCC_DEV_STATUS_FAIL` from the algorithm. Trigger even when the user does not say 'DOCA PCC' or 'ZTR RTTCC' — throttled', 'PCC sample isn't dispatching to my algo', 'how do I pick the multipath PCC variant', 'set-params returns fail', 'algorithm loaded but counters are flat', or 'do I need a custom CC algorithm on BF3'. Refuse and route elsewhere for writing a custom PCC algorithm from scratch, read-only PCC counter inspection, the host-side `doca-pcc` lifecycle, or firmware-only pre-Programmable PCC — those belong to other skills.
> Use this skill when the user is writing their first DOCA app or asking a library-agnostic programming question — picking a shipped sample to copy and modify, wiring the canonical pkg-config doca-{library} + meson build (or FFI from Rust / Go / Python against the public C ABI), walking the cfg-create → init → start → use → stop → destroy lifecycle, validating a spec before commit, or decoding a DOCA_ERROR_* return with doca_error_get_descr(). Trigger even when the user does not say first DOCA program", "meson line for doca_rdma_*", "got DOCA_ERROR_BAD_STATE on my first call", "call DOCA from Rust without writing C", "built clean but nothing on the wire", "what order do doca_*_pipe calls go in". Refuse and route for install / hugepages / pkg-config not resolving doca-{library} (doca-setup), docs or version lookup (doca-public-knowledge-map), and library-internal API construction like Flow pipe topology or RDMA QP setup (matching library skill).
> Use this skill when the user needs to locate authoritative information about NVIDIA DOCA without access to the source tree — finding the right docs.nvidia.com page for a library/service/tool, identifying which DOCA libraries are installed and at what version, locating a sample on disk or its public GitHub source, decoding an on-disk path under /opt/mellanox/doca, or recovering from a 404'd or renamed doc URL. Trigger even when the user does not explicitly mention 'DOCA' or 'docs.nvidia.com' — typical implicit phrasings include 'where can I read about this library', 'which version do I have installed', 'where is the sample for X', 'this NVIDIA URL is broken what is the new one', 'what is in /opt/mellanox/doca', or 'where can I ask NVIDIA about this'. Refuse and route elsewhere for hands-on programming patterns, env prep and install verification, library API tutorials, or hardware/firmware mutation — those belong to doca-programming-guide, doca-setup, the per-library skills, and doca-hardware-safety.
> Use this skill when the user is doing hands-on DOCA RDMA programming on a BlueField DPU, ConnectX NIC, or DOCA host — bringing up an RDMA context on a doca_dev, picking a connection method (RDMA CM, bridge/OOB, or gRPC exchange of doca_rdma_export()), enabling one of the eleven task types (Send/Receive/Send-Imm, Read/Write/Write-Imm, Atomic CmpSwap/FetchAdd, Get/Set/Add Remote Sync Event), setting matching mmap + RDMA permissions, sizing queues and connections, querying doca_rdma_cap_*, or debugging DOCA_ERROR_* from an RDMA call. Trigger even when the user does not mention "DOCA RDMA" — typical implicit phrasings include "one-sided read returns permission denied", "completions never arrive after submit", "connection callback never fires", "how do I do atomic compare-and-swap over RoCE", or "send queue hits DOCA_ERROR_FULL under burst". Refuse and route elsewhere for general RDMA / ibverbs theory (queue pairs, MRs, RoCE vs IB), installing DOCA itself, or non-RDMA DOCA libraries.
> Use this skill when the user is doing hands-on DOCA RDMI (RDMA Initiator) programming — picking doca-rdmi vs doca-rdma for an accelerator-initiated one-sided RDMA flow, standing up a doca_rdmi_connection or doca_rdmi_poster, attaching a doca_dpa_completion or doca_verbs_cq before doca_ctx_start(), retrieving the DPA-side handle for a DPA kernel, auditing whether a doca_rdmi_* symbol is EXPERIMENTAL on this DOCA, or debugging DOCA_ERROR_* returns from RDMI calls. Trigger even when the user does not say "DOCA RDMI" or "initiator" — implicit phrasings include "my DPA kernel needs to post RDMA writes to a remote responder", "DPA kernel sees no completions", "function not found on doca_rdmi_* at link time", "DOCA_ERROR_BAD_STATE from completion attach", or "the DPA posted but the work request never arrived". Refuse and route elsewhere for two-sided or host-CPU RDMA, the DPA programming model, GPU-side RDMA initiation, or general RDMA/IB/RoCE concepts — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA Rivermax work on a BlueField DPU or ConnectX host — standing up `doca_rmax_in_stream` (receive) sessions for timing-precise media-over-IP (SMPTE ST 2110 video/audio, market data, scientific feeds), confirming the Rivermax SDK + license precondition before any DOCA-side code, running `doca_rmax_get_*_supported` capability queries, pairing with `doca-eth` queues and `doca-flow` steering, or debugging `DOCA_ERROR_*` from a Rivermax call. Trigger even when the user does not explicitly mention "DOCA Rivermax" or "rmax" — implicit phrasings include "ST 2110 receive isn't getting frames", "sub-microsecond jitter on BlueField", "NOT_SUPPORTED from doca_rmax_init", "no recv events after stream start", or "license check failing on a media receiver". Refuse and route elsewhere for installing the Rivermax SDK or its license, programming the underlying queue (`doca-eth`), steering rules (`doca-flow`), or best-effort packet I/O — those belong to other skills.
> Use this skill when the user is dealing with the DOCA environment around their workload — verifying an install is healthy, preparing the build env (pkg-config, headers, LD_LIBRARY_PATH, hugepages, devlink, representors), debugging env-class failures, deciding container-vs-bare-metal deployment shape, or reaching a DOCA install from a host that doesn't have one yet via the NGC DOCA container Stage-1 fallback. Trigger even when the user does not explicitly mention "DOCA setup" — typical implicit phrasings include "I just got a BlueField, what now", "my code is built, how do I run it", "pkg-config can't find doca-flow", "no free 2048 kB hugepages", "representor X not found", "I'm on a Mac and want to learn DOCA". Refuse and route elsewhere for library API specifics (Flow pipes, RDMA queues), the modify-a-sample first-app workflow or DOCA_ERROR_* program-side debugging, and "where is X documented" knowledge-map questions — those belong to other skills.
> Use this skill when the user is doing hands-on DOCA SHA programming — offloading SHA-1, SHA-256, or SHA-512 hashing onto a BlueField DPU or ConnectX accelerator, picking between one-shot `doca_sha_task_hash` and incremental `doca_sha_task_partial_hash`, querying `doca_sha_cap_*` for algorithm support and min destination / max source buffer sizes, setting source / destination `doca_mmap` permissions, or decoding DOCA_ERROR_* returns from the SHA API. Trigger even when the user does not explicitly mention "DOCA SHA" or "doca_sha_task" — typical implicit phrasings include "hash a multi-GiB file on the DPU", "offload SHA-256 to the BlueField", "streaming hash over chunks", "partial hash returns BAD_STATE", "destination buffer too small for digest", or "is SHA-512 available on this card". Refuse and route elsewhere for general cryptographic-hash theory (collision resistance, SHA-3 selection), other DOCA crypto libraries (AES-GCM, Compress, DMA), or DOCA install / BFB bring-up — those belong to other skills.
> Use this skill when wiring the DOCA SHA Offload Engine (an OpenSSL ENGINE) into an existing OpenSSL pipeline to offload one-shot SHA-1, SHA-256, or SHA-512 (EVP_Digest) onto DOCA SHA hardware without rewriting against doca-sha. Covers engine load mechanics (`openssl engine dynamic`, `set_pci_addr` ctrl, `-engine_impl`), the SHA-224 negative test that proves offload engaged, the message-size window where offload beats CPU SHA, and engine-vs-library selection. Trigger even when the user does not say "DOCA SHA Offload Engine" or "OpenSSL ENGINE" — typical implicit "offload SHA without code changes", "is openssl using the accelerator or falling back to software", "prove DOCA SHA actually ran", "openssl dgst hashed but I'm not sure it was offloaded". Refuse and route elsewhere for new SHA pipelines (use doca-sha), MD5 / SHA-3 / SHA-224 / HMAC-SHA offload, incremental hashing via chained `EVP_DigestUpdate`, or OpenSSL PROVIDER authoring.
> Use this skill when the operator is driving the DOCA Socket Relay to bridge a socket-oriented host application onto a BlueField DPU peer without rewriting it — picking the deployment shape (in-process, sidecar, or BlueField service container), configuring the host-side socket and the DPU-side forwarding endpoint, walking the bind → connect → round-trip → admit-fleet smoke, or diagnosing a stuck/silent relay. Trigger even when the user does not explicitly mention "DOCA Socket Relay" — typical implicit phrasings include "move my socket app onto the BlueField without rewriting it", "host app gets ECONNREFUSED on the relay", "relay accepts the connection but bytes never arrive on the DPU side", "first round-trip works, the rest hang", "bridge an AF_UNIX (UDS) socket to a DPU peer over Comch", or "I want a sidecar that forwards my socket to the BlueField". Refuse and route elsewhere for the comch programming API, line-rate raw packet I/O via doca-eth, and DOCA install/bring-up — those belong to other skills.
> Use this skill when the user is invoking `doca_spcx_cc` (the host-side CLI under /opt/mellanox/doca/tools/) to load, parameterize, start, observe, or stop a Programmable Congestion Control (SPCX) algorithm on a BlueField with a DPA processor against a live RDMA / RoCE fabric, or picking SPCX vs the established `doca-pcc` surface. Trigger even when the user does not say "DOCA SPCX" or "doca_spcx_cc" — typical implicit phrasings include "I want to write a custom RTT-based CC algorithm for my RoCE fabric", "my SPCX session loaded but throughput / latency didn't change", "doca_pcc status shows Active but factory CC seems to still be in charge", "DOCA_PCC_PS_ERROR on start", "is the programmable-CC surface available on my install", or "DPA-side algorithm image won't load". Refuse and route elsewhere for DPA-side algorithm authoring detail, factory PCC firmware configuration, read-only PCC counter inspection, raw DPA cycle profiling, RDMA library programming, or general DOCA install — those belong to other skills.
> Use this skill when the user is doing hands-on NVMe-over-Fabrics storage-target work on a BlueField DPU or ConnectX NIC with DOCA STA — standing up a doca_sta DOCA Core context that accelerates the target-side NVMe-oF data path over RDMA, defining doca_sta_subsystem targets (NQN + namespaces) backed by local NVMe-PCI backend disks (doca_sta_be), checking device support via doca_sta_cap_is_supported, sizing the per-connection I/O queues, or debugging DOCA_ERROR_* from a STA call. Trigger even when the user does not say "DOCA STA" — typical implicit phrasings include "my NVMe-oF Connect never completes", "Identify Controller times out over RoCE", "16 I/O queues at depth 1024 — does this BlueField support that", "offload the nvmf target onto the DPU", or "DOCA_ERROR_IO_FAILED on an NVMe read". Refuse and route elsewhere for DOCA install, raw RDMA data movement, raw packet I/O, flow-rule programming, or initiator-side / host NVMe stack work — those belong to other skills.
> Use this skill whenever another DOCA skill says "prefer the structured tool per doca-structured-tools-contract", or when the user wants a one-shot answer that consolidates info multiple manual commands would produce — DOCA env / version / devices / capabilities / validate / host vs DPU state. Trigger even when the user does not explicitly mention "structured tool" or "doca-env --json" — typical implicit phrasings include "is there one command that tells me everything about my DOCA install", "what version is X capability available since", "every PF/VF/SF visible on this BlueField with PCIe address", "will this pipe pass validate before commit", "diff host vs DPU state", or "why does the agent give a one-line answer on host A and five commands on host B". Refuse and route elsewhere for general DOCA orientation, specific library API how-to, or install-from-scratch guidance — those belong to the per-library skill, doca-public-knowledge-map, or doca-setup.
> Use this skill to read DOCA hardware-counter events from a `doca_telemetry_pcc`, `_dpa`, `_diag`, `_adp_retx`, `_phy`, and `_pci`. It covers capability checks, context creation, startup, and per-domain reads or samples. Trigger for implicit requests such as "read PCC counters from my BlueField app", "sample DPA counter exports", or "expose PHY, PCI, or DIAG counters from this doca_dev". This is the counter-reader surface, not a NetFlow, IPFIX, or local-socket collector. Route publishing and export to `doca-telemetry-exporter`; route deployed DOCA Telemetry Service (DTS), collectors, and plain stdout logging elsewhere.
> Use this skill when the user is doing hands-on DOCA Telemetry Exporter programming on a host where DOCA is installed — defining a doca_telemetry_exporter_schema, creating sources, picking counter/gauge/event types, running capability queries before assuming limits, registering schemas before the first emit, or debugging DOCA_ERROR_* failures from the exporter API. Trigger even when the user does not explicitly mention "DOCA Telemetry Exporter" or "doca_telemetry_exporter_*" — typical implicit phrasings include "publishing counters from my DOCA app", "emit returns AGAIN under bulk load", "consumer sees nothing but emit reports success", "NOT_FOUND on first submit", or "should I link the exporter or the telemetry service". Refuse and route elsewhere for the receiving DOCA Telemetry Service, plain stdout logging via doca_log, non-DOCA Prometheus scrape sinks, or real-time event subscription back into the app via doca-comch — those belong to other skills.
> Use this skill when the user is invoking `doca_telemetry_utils` on a host with DOCA installed — discovering the diagnostic-counter schema, translating counter names to binary Data IDs, validating per-device counter support before committing a DOCA Telemetry exporter config, or reverse-resolving a captured Data ID. Trigger even when the user does not explicitly mention "doca_telemetry_utils" or "Data ID" — typical implicit phrasings include "my exporter ships but the collector sees nothing", "this metric silently drops downstream", "which counters does this BlueField expose", "translate this 0x... back to a counter name", "what do node / pcie_index / depth mean here", or "is this counter supported on this device before I commit it". Refuse and route elsewhere for developer-side collector / exporter library programming, DTS deployment, or DOCA install / repair — those belong to doca-telemetry, doca-public-knowledge-map, and doca-setup.
> Use this skill when the user is contemplating a DOCA upgrade or downgrade — moving a host to a newer DOCA release, refreshing the BlueField BFB, bumping the NGC DOCA container tag, or rolling back. detect what is installed, discover what newer release exists, report the gap, then STOP and ask for explicit confirmation — never upgrade newer DOCA", "should I move to the next release", "I want the latest features", "my component is being deprecated, what now", or "roll me back". Route elsewhere for version detection (doca-version), first-time install (doca-setup), any hardware/firmware/reboot step (doca-hardware-safety), and public-docs / sunset routing (doca-public-knowledge-map).
> Use this skill when the user is doing hands-on DOCA UROM library work from the host side — wiring doca-urom under an HPC / UCX / MPI stack to OFFLOAD remote memory operations (puts, gets, atomics, collectives) onto a BlueField DPU, creating a UROM Service context (doca_urom_service_*) and Worker contexts (doca_urom_worker_*) that run plugins on the DPU, discovering plugins via doca_urom_service_get_plugins_list, progressing completions, or debugging DOCA_ERROR_* from a doca_urom_* call. Trigger even without onto the BlueField", "first doca_urom call returns NOT_PERMITTED", or "host library and DPU service look out of sync". Route elsewhere for UROM Service deployment on the DPU side, MPI / UCX collective algorithm design, and RDMA / RoCE / IB substrate bring-up.
> Operate the DOCA UROM Service container on BlueField Arm for remote memory operations (puts, gets, atomics, collectives) enqueued by a component, size queues, configure Comch pairing, and align host and control; Comch pairing and RDMA permissions are the boundary. Pair only intended hosts, expose least-privilege memory regions, and verify both views before start. Trigger for slow UCX collectives, unexpected NOT_PERMITTED, or missing completions. Do not use for host application code, MPI/UCX integration design, or DOCA install.
> Use this skill when the user is dropping below the higher-level DOCA libraries (doca-rdma / doca-eth / doca-rmax) into the raw-verbs escape hatch — managing QP / CQ / PD / MR / SRQ / AH / CC-group / Ethernet-SQ-RQ primitives inside DOCA Core, porting libibverbs code into the DOCA Core model, capability-querying a specific verb / opcode / WR flag / QP attribute via doca_verbs_query_device, or debugging DOCA_ERROR_* from doca_verbs_* calls. Trigger even when the user does not say "doca-verbs" — implicit phrasings include "raw QP attribute the task API doesn't expose", "keep my ibv_* code next to doca_* on the same QP", "IO_FAILED on WR submit", "QP state transition rejected", "attach a congestion- control group", or "porting my libibverbs code". The skill's first job is to route MOST users back UP to the higher-level library. Refuse and route elsewhere for general doca-rdma / doca-eth / doca-rmax workloads, DOCA install, Core internals, and general libibverbs theory — those belong to other skills.
> Use this skill when the user is doing DOCA version handling — detecting the installed release, validating the four-way match across pkg-config doca-common, applications/VERSION, doca_caps --version, and bfver/mlnx-release on BlueField, reasoning about NGC container tags, looking up whether a capability is on the installed release, or diagnosing build-vs-runtime drift. Trigger even when the user does not explicitly say "DOCA version" or "four-way match" — typical implicit phrasings include "program built but does nothing on the wire", "undefined reference to a symbol the docs claim exists", "DOCA_ERROR_NOT_SUPPORTED at runtime", "counter didn't increment", "what does `latest` mean for this tag", or "is my LTS still supported". Refuse and route elsewhere for installing or choosing DOCA packages (doca-setup), per-library API/capability questions (matching library skill), the cross-library DOCA_ERROR_* taxonomy (doca-programming-guide), or the general debug ladder (doca-debug) — those belong to other skills.
Validate that a Dynamo deployment's NIXL/UCX/NCCL interconnect is ready for disaggregated serving over RDMA/NVLink. Use after recipe-runner brings a deployment up (especially disagg/multi-node) to confirm the KV transport is correct; use troubleshoot for diagnosing already-failed pods.
Select, validate, patch, and deploy existing NVIDIA Dynamo Kubernetes recipes. Use for model/backend/GPU/deployment-mode recipe bring-up; use router-starter for router-only mode work and troubleshoot for broken deployments.
Diagnose failed or unhealthy Dynamo deployments. Use when pods, model-cache jobs, PVCs, workers, frontend/router health, endpoints, or benchmark jobs fail; use recipe-runner/router-starter before this for normal bring-up.
Install Holoscan SDK v4.3+ via Conda in a CUDA 13 environment. Use for Conda installs; redirect CUDA 12 hosts to container/wheel.
Install Holoscan SDK via the NGC Docker container. Use for container-based installs; not for native apt/pip/Conda installs.
Install Holoscan SDK natively on Ubuntu via apt. Use for C++ installs on Ubuntu; pair with /holoscan-install-wheel for Python.
Build Holoscan SDK from source via the in-tree ./run script. Use only when published packages don't meet the user's needs.
Install Holoscan SDK Python wheel via pip into a venv. Use for Python installs; not for native C++/apt or Conda installs.
Guides Holoscan SDK installation: inspects the host, assesses platform compatibility, recommends an install method, and delegates to the matching install skill.
Discover and run Holoscan Sensor Bridge example applications on a connected devkit. Filters available apps by the user's platform, HSB software version, board type, and sensors. Supports timed execution, failure analysis, code-edit suggestions, and iterative re-runs.
Flash the FPGA on an HSB board connected to an NVIDIA devkit. Supports HSB Lattice boards (FPGA versions 2407, 2412, 2507, 2510) and Leopard Imaging VB1940 "all-in-one" cameras (FPGA versions 2507, 2510). Uses release-specific YAML manifests and board-type-specific program commands. Lattice and VB1940 commands must never be mixed.
Create or explain fixed-format HSB FPGA_top.sv wrappers from validated HOLOLINK_def.svh files. Do not use for def generation or validation.
Generate, validate, compare, or explain HSB HOLOLINK_def.svh macros. Do not use for FPGA_top.sv wrappers or packetizer-only derivation. Generation runs bundled Python scripts locally through shell commands and writes validated .svh output files after user-confirmed paths.
Choose or explain HSB Sensor RX packetizer fields for HOLOLINK_def.svh. Do not use for full defs, validation, or runtime APB programming.
Clone the latest NVIDIA Holoscan Sensor Bridge repo, ask which supported devkit is being used, configure the host per platform, build the correct demo container, run it, and verify HSB connectivity by pinging 192.168.0.2. Use for Holoscan Sensor Bridge setup, build, container launch, and first-connectivity bring-up.
Execute QA test plans on Holoscan Sensor Bridge hardware. Reads a user-provided test document, filters tests by the user's setup, determines which tests can run automatically, executes them with pass/fail evaluation, and produces a structured test results report.
Overview of `workflows/catheter_navigation/` (fluorosim DRR, XPBD physics, vasculature digital twin). Use when the user asks what the catheter navigation workflow is, what's supported, or where to start.
Build a patient vasculature digital twin from CT (preprocess + segment). Use when asked to preprocess CT, segment vessels, extract centerline, or prepare ct_cache for viewport/DRR.
End-to-end smoke for catheter navigation covering setup, digital twin, DRR, and unit tests. Use when asked to run the full catheter workflow smoke or demo the v0.7 pipeline.
Render a single DRR fluoroscopy frame from a CT cache or synthetic phantom. Use when asked to render DRR, generate a fluoro image, or smoke-test the Slang renderer.
Verify host/GPU requirements and PYTHONPATH for the catheter navigation workflow. Use when asked to set up, install, or bootstrap catheter_navigation, or when hitting import/GPU/slangpy errors.
Run CPU-only fluorosim smoke tests (imports, preprocessing, CLI parsers). Use when asked to smoke-test catheter navigation in CI or without a GPU.
Launch the interactive Slang fluoroscopy viewport with XPBD catheter physics. Use when asked to open the viewport, teleop a catheter, or demo fluoro navigation.
Serve the LeRobot HTML visualizer for a converted dataset in a browser. Use when asked to visualize, inspect, or open a LeRobot dataset; not for converting HDF5 (use [[i4h-workflow-dataset-convert]]).
Overview of `workflows/agentic/` (IsaacLab-Arena + GR00T/openpi). Use when the user asks what i4h workflow is, what's supported, or where to start.
Create a new agentic env by forking an existing env. Use for new env/task scaffolding, not scene edits or baking.
Use a VLM to verify whether each episode satisfies the env's task description. Use when the user asks to annotate, label episodes, filter demos, or gate finetuning on a success classifier.
Convert an agentic HDF5 recording into a LeRobot dataset (parquet, meta, videos). Use when asked to convert HDF5, prepare for training, or export to LeRobot; not for viewing — use [[i4h-lerobot-viz]].
Expand an HDF5 recording by cloning trajectories with action/state noise. Use when asked to mimic, expand, or augment a dataset; not for recording new demos (use [[i4h-workflow-dataset-teleop]]).
Replay a recorded HDF5 episode inside Isaac Sim for visual verification. Use when the user asks to replay, play back, or step through an HDF5 recording.
Record episodes for an agentic env via teleoperation (keyboard, SO-ARM leader, or VR) into HDF5. Use when the user wants to teleop or record human demos.
Run the full end-to-end agentic pipeline (record → mimic → annotate → replay → convert → visualize → finetune → validate). Use when asked to run the whole pipeline or do an e2e, smoke, or demo run.
Fine-tune a GR00T or openpi PI0 policy on a LeRobot dataset. Use when asked to finetune, train, or post-train a policy on demos; not for evaluating a checkpoint (use [[i4h-workflow-validate]]).
Edit an env's scene in place — objects, cameras, task, success bounds, randomization. Use when asked to edit a scene or launch/run/open an env in edit mode (`--bridge`), incl. a just-created env.
Verify host requirements and run `workflows/agentic/setup.sh`. Use when asked to set up, install, or bootstrap the agentic workflow, or hits missing `.venv`, third-party checkout, or engine errors.
Validate, evaluate, or run i4h envs. Use for policy/checkpoint rollouts and scripted state-machine smoke runs.
>- Use when you need to rebuild the BSP overlay — DT, OOT modules, bsp, rebuild dtb, rebuild kernel.
>- Enable MIPI/GMSL camera sensors on a Jetson Thor or Orin custom carrier by rendering a kernel-DT overlay from the in-tree sensor DTSI. Do NOT use for UPHY lane allocation or ODMDATA edits.
Use to lock/cap Jetson CPU/GPU/EMC clocks, toggle EMC/CPU DVFS, or change cpufreq governors by editing BPMP DTB and nvpower.sh pre-flash. Do NOT use for live tuning or nvpmodel edits.
>- Use when you need to add, remove, edit, list, or change the boot default of an nvfancontrol fan profile on a Jetson/Tegra (Orin,
>- Enable Jetson Thor 25G/10G/1G MGBE QSFP via kernel-DT overlay. Do NOT use for UPHY lane allocation or ODMDATA edits.
>- Use when you need to add, remove, edit, list, or change the boot default of an nvpmodel power mode on a Jetson/Tegra (Orin, Thor)
>- Per-controller PCIe enable / disable / lanes / link-speed for a Jetson Thor or Orin custom carrier via ODMDATA + kernel-DT overlay. Do NOT use for UPHY lane allocation or endpoint-mode bring-up.
>- Per-pin SFIO / direction / initial-state configurator for a Jetson Orin or Thor custom carrier from the pinmux XLSM. Do NOT use for kernel-DT overlay or ODMDATA edits.
Configure Jetson UPHY lane allocation (uphy0/uphy1-config) on Orin/Thor custom carriers. Do NOT use for pinmux or PCIe-only edits.
Enable/disable Jetson USB2/USB3 SS ports via kernel-DT overlay. Do NOT use for UPHY lane allocation or ODMDATA edits.
>- Bootstrap a custom carrier board by forking carrier files and scaffolding a DT overlay from the reference devkit. Use after jetson-init-source; not for module-level or kernel-DTB changes.
Read-only Jetson health snapshot for identity, memory, GPU, thermal, power, storage, services, and top processes.
>- Download NVIDIA Jetson Linux BSP artifacts (BSP tarball, sample rootfs, public_sources, x-tools, guides) for the active target. Use for Auto Setup; not for extraction or profile edits.
Use to flash a promoted BSP image to a Jetson DUT in RCM mode via flash.sh or l4t_initrd_flash.sh. Do NOT use for BSP customization, image promotion, or carrier derivation.
>- Build a per-target knowledge-base markdown next to the active profile by walking the BSP root and source tree. Use after init-image / init-source; not for editing profile fields.
Plan and apply safe Jetson headless-mode changes to reclaim GUI and daemon memory.
Pick the serving stack and per-runtime memory flags (vLLM, SGLang, llama.cpp, TensorRT Edge-LLM) for an LLM/VLM workload on any NVIDIA Jetson.
>- Extract Jetson Linux + sample-rootfs tarballs and run apply_binaries.sh for the active target, then record bsp_image in the profile. Use after jetson-init-target; not for source-tree setup.
>- bsp_sources, Crosstool-NG toolchain. Use after jetson-init-image; not for fetching inputs.
>- Author a new Jetson target-platform profile (reference_devkit + optional custom_carrier) and update the active pointer. Use to create a target; not for switching existing profiles.
>- Bind pre-downloaded Jetson reference docs (developer guide, design guide, pinmux, schematics) into the active profile documents block. Use after staging docs on disk; not for downloading.
Benchmark Jetson LLM/VLM serving performance across vLLM, llama.cpp, and Ollama with structured JSON output.
Stand up vLLM or SGLang serving on Jetson, using upstream vLLM on Thor and Orin JetPack 7.2+, and NVIDIA-AI-IOT vLLM on older Orin.
Measure Jetson DRAM/NvMap usage and verify before/after memory reclamation with live audit data.
>- Reclaim DRAM by disabling unused subsystems across MB1 BCT, MB2 BCT, kernel reserved-memory, and SWIOTLB. Use for headless or no-camera Jetson deployments; not for CPU/GPU frequency tuning.
Pick Jetson-compatible containers, vLLM runtime images, and Jetson AI Lab PyPI indexes; maps Orin SM 8.7 vs Thor SM 11.0 and JetPack-specific package choices.
Use when you need to print Jetson BSP info (L4T version, board configs, rootfs state) from a Linux_for_Tegra root on the host PC. This is an example skill.
Use when you need to print Jetson device info (module model, L4T version, kernel, OS version, current power mode) from a running Jetson target. This is an example skill.
>- Use to promote overlay files and built artifacts into the staged image.
>- Entry skill for Jetson / IGX BSP customization. Asks one core click-to-select setup questionnaire and passes prefilled answers to downstream setup skills.
>- Switch the active Jetson target-platform pointer to an existing profile YAML. Use before customize/build/flash to change target; not for authoring profiles — use jetson-init-target instead.
Add EAGLE-3 or draft-model speculative decoding to a Jetson vLLM server when TPOT is the bottleneck.
>- Use after jetson-flash-image to run static BSP checks, on-target smoke/regression tests on a flashed DUT, or both. Not for build
Playbook for launching, monitoring, stopping, and debugging NeMo-RL recipes on a Kubernetes cluster via the nrl-k8s CLI. Covers ephemeral vs long-lived RayCluster modes, iterating on runs, and debugging hung or failed training jobs.
Guide for selecting and configuring distributed training strategies in NeMo AutoModel, including FSDP2, Megatron FSDP, DDP, and parallelism settings.
Configure NeMo AutoModel job launches for interactive runs, Slurm clusters, and SkyPilot cloud execution.
Guide for onboarding new model architectures into NeMo AutoModel, including architecture discovery, implementation patterns, registration, and validation.
Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow.
Use when the user wants to create a dataset, generate synthetic data, or build a data generation pipeline.
Use when working on the Evaluator plugin CLI, jobs, SDK-backed specs, metric types, or plugin-owned Evaluator skills.
Run Megatron-LM (MLM) and Megatron Bridge training with mock or real data. Covers correlation testing, available recipes, and multi-GPU examples.
Convert single-node scripts to multi-node Slurm sbatch jobs and debug common multi-node failures. Covers srun-native vs uv run torch.distributed approaches, container setup, NCCL timeouts, OOM sizing for MoE models, and interactive allocation.
Validate and use selective and full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute.
Validate and use CPU offloading in Megatron Bridge, including layer-level activation offloading and fractional optimizer state offloading with HybridDeviceOptimizer.
Validate and use CUDA graph capture in Megatron Bridge, including local full-iteration graphs and Transformer Engine scoped graphs for attention, MLP, and MoE modules.
Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.
Operational guide for enabling hierarchical context parallelism in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.
Operational guide for enabling Megatron FSDP in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.
Techniques for reducing peak GPU memory in Megatron Bridge — expandable segments, PEFT + SP input re-gather, parallelism resizing, activation recompute, CPU offloading constraints, and common OOM fixes.
MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.
Choose the right MoE token dispatcher (`alltoall`, DeepEP, or HybridEP) for the hardware, EP degree, and optimization stage. Summarizes patterns from DSV3, Qwen3, Qwen3-Next, and VLM bring-up work.
Representative MoE training playbooks by hardware platform and model family. Summarizes rounded throughput bands, parallelism patterns, and common tuning stacks.
Long-context MoE training guidance for Megatron Bridge. Covers CP sizing, selective recompute, dispatcher choices, and practical patterns from DSV3, Qwen3, and Qwen3-Next long-context experiments.
Systematic workflow for MoE training optimization in Megatron Bridge, based on the Megatron-Core MoE paper. Covers the Three Walls framework, parallel folding, recompute strategy, dispatcher choice, and CUDA-graph bring-up.
Practical guidance for training MoE VLMs in Megatron Bridge. Compares FSDP and 3D-parallel approaches, using rounded lessons from Qwen3-VL, Qwen3-Next, and other multimodal experiments.
Operational guide for choosing and combining parallelism strategies in Megatron Bridge, including sizing rules, hardware topology mapping, and combined parallelism configuration.
Validate and use packed sequences and long-context training in Megatron-Bridge, distinguishing offline packed SFT for LLMs from in-batch packing for VLMs, and applying the right CP constraints.
Operational guide for enabling TP, DP, and PP communication overlap in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.
Recommend and customize Megatron Bridge recipes for a user's model, GPU count, and training goal. Indexes library recipes (pretrain/SFT/PEFT) and performance recipes.
Resiliency features in Megatron Bridge including fault tolerance, straggler detection, in-process restart, preemption, and re-run state machine.
Autonomous NeMo-RL research agent workflow for directed hypothesis testing and open-ended discovery. Guides agents through the full experiment lifecycle: understanding recipes and environments, wiring RL or NeMo-gym runs, launching reproducible baselines and iterations, analyzing results, preserving human oversight, and using git plus TSV logs as the research ledger. Do NOT use for: bug fixes, code review, documentation, refactoring, dependency updates, or single-file changes.
Brev instance operating guidance for NeMo-RL agents working in /home/ubuntu/RL with limited workspace disk, a larger /ephemeral volume, and optional /home/ubuntu/RL/.env secrets. Use when running nemo-rl-auto-research campaigns, experiments, training jobs, model or dataset downloads, shared cache-heavy commands, log-producing runs, checkpoint generation, W&B or Hugging Face authenticated workflows, or any workflow that may create large files on Brev.
Documentation conventions for NeMo-RL. Covers docs/index.md updates and docstring format. Do NOT use for: bug fixes, test fixes, dependency bumps, refactoring, CI/CD changes, performance tuning, or any task that does not involve writing or updating documentation.
Manage durable working-session memory for coding agents. Use when a user asks to preserve or recover agent context across disconnects, VS Code restarts, long-running work, handoffs, or any session where important state should be written periodically under the repo's session directory. Do NOT use for: simple questions, short tasks, one-off commands, linting, or code review.
Orchestration skill for NVIDIA Nemotron Speech (Riva) / NeMo ASR domain and language adaptation. Given a goal like "improve/fine-tune ASR for my domain or language", it scopes the task, picks the cheapest sufficient path (word boosting → n-gram LM → fine-tuning), delegates each stage to the right sub-skill (data generation, training, evaluation, deployment), and answers cost/time/data questions along the way.
Plan, configure, and chain repo-native Nemotron customization steps into single-step or multi-step pipelines: curation, translation, SFT/PEFT (AutoModel or Megatron-Bridge), pretraining/CPT, RL alignment (DPO/RLVR/GRPO/RLHF), BYOB/MCQ benchmarks, checkpoint conversion, ModelOpt optimization, env profiles, and evaluation of trained checkpoints or existing/hosted endpoints. Use when a request names a Nemotron step or workflow, or asks to clean, translate, train, fine-tune, align, convert, optimize, evaluate, or compose these into a pipeline. Do NOT use for frontend/dashboard/visualization work, generic ML advice, billing/access, or non-Nemotron coding tasks.
Generates BYO custom safety policies for NVIDIA Nemotron content-safety guardrails — Nemotron-Content-Safety-Reasoning-4B (text) and multimodal Nemotron-3-Content-Safety. Produces a Markdown policy, JSON taxonomy, and drop-in inference prompts. Maps rough words or an existing policy to V2 categories, adding custom categories or topic-following rules.
Use when planning, debugging, tuning, evaluating, exporting, or deploying public Nemotron `embed`/`rerank` retrieval recipes.
Routes NVIDIA Nemotron Speech (Riva) NIM tasks — deploys, runs, and tests ASR, TTS, and NMT NIMs on build.nvidia.com or self-hosted.
Used for generating synthetic CT volumes and masks with NV-Generate-CTMR rflow-ct. Not for production training data without review.
Used for generating synthetic body MRI volumes with NV-Generate-CTMR rflow-mr. Not for paired masks or production training data.
Used for generating synthetic brain MRI volumes with NV-Generate-CTMR rflow-mr-brain. Not for production training data.
Used for finetuning NV-Generate-CTMR MR-brain diffusion UNet from a NIfTI datalist. Not for clinical or production data approval.
Used for finetuning the NV-Generate-CTMR MAISI VAE from CT/MRI NIfTI datalists. Not for clinical or production data approval.
>- Use for NVIDIA-related requests where an NVIDIA skill might help, even if the user did not ask for a skill. Trigger on NVIDIA products, hardware, software, SDKs, GPUs, Jetson/JetPack/L4T/BSP/SDK Manager/driver/flashing/setup, CUDA, NIM, NeMo, Omniverse/OpenUSD/SimReady, RAPIDS/cuDF, cuPyNumeric, cuOpt, Dynamo, Holoscan, TensorRT, DeepStream, VSS, TAO, NGC/NVCF. Do not use for generic non-NVIDIA route, optimize, deploy, AI, video, data, or infrastructure tasks.
Used for command-shape or live NV-Reason-CXR chest X-ray reasoning smoke tests. Not for diagnosis or clinical reporting.
Used for running NV-Segment-CT VISTA3D on CT NIfTI volumes and recording label-map evidence.
Used for smoke or dataset finetuning of NV-Segment-CT VISTA3D on CT NIfTI labels. Not for clinical validation.
Used for running NV-Segment-CTMR on CT or MRI NIfTI volumes and recording label-map evidence. Not for clinical interpretation.
Coordinate the end-to-end CAD/source-asset to SimReady workflow. Use for broad requests such as CAD to SimReady, source asset to simulation-ready USD, or prop packaging that require conversion, material/physics assignment, SimReady conformance, validation, and optional package creation; deploy or verify Content Agents services first when property assignment is enabled; route single-stage work through nested references.
Use as the top-level router for Omniverse Realtime Viewer USD app requests and focused viewer reference documents.
Top-level workflow skill for USD performance diagnosis and optimization. Handles slow loading, high memory, low FPS, and broad scene-optimization requests; delegates auth/runtime setup to Phase 0 owners.
>- Full PAIDF AnomalyGen pipeline — fine-tune on a new anomaly dataset, generate synthetic anomaly images (SDG), evaluate quality (nn_score), and search per-sample generate from an existing checkpoint). Use when the user asks to "fine-tune AnomalyGen", "generate anomaly images", "run PAIDF SDG", "evaluate SDG output quality", "run per-sample search", or run any part of the AnomalyGen pipeline, even if they only mention one phase.
>- Use when the user wants to orchestrate defect image generation with NVIDIA Cosmos AnomalyGen (Cosmos-Predict2-derived) on OSMO for PCBA, metal surface, and glass inspection. The Day 0 path handles cold-start with USD-to-ROI, image-edit augmentation, and AnomalyGen to create initial PCBA datasets. The Day 1 path performs inference and labeling on real images. This skill helps with first-time asset setup, creation of finetuning checkpoints, and configuring deployment.
>- Use when the user wants to set up, scale, validate, or harden NVIDIA physical AI infrastructure for synthetic data generation workflows across local MicroK8s or Azure AKS, including Kubernetes clusters, inference endpoint deployment, OSMO deployment, workload submission readiness, and infrastructure scaling, SDG infrastructure, microk8s, azure aks, NVCF deployment, summarization or workload-only operations unless infrastructure setup, scaling, validation, or recovery is requested.
Router for NVIDIA NuRec/NRE: USDZ rendering, NCore conversion, 3DGS, gRPC sensor sim, PhysicalAI HF datasets. Do NOT use for SimReady or infra setup.
>- Use when running people attribute search (PAS) image augmentation and attribute search, PAS, person augmentation, attribute search, person re-identification, clothing augmentation, person crop augmentation.
>- flow selection, preflight, submit-time interpolation, monitoring, and output labeling, VDA demo, OSMO workflow, pseudo labeling.
Use when a user asks to build, optimize, backtest, rebalance, or analyze a stock portfolio with Mean-CVaR, Mean-Variance/SOCP variance caps, efficient frontiers, scenario generation, or NVIDIA cuOpt.
NVIDIA RAG Blueprint — deploy, configure, troubleshoot, and manage. Handles any RAG action: deploy, install, start, enable, disable, toggle, change, configure, troubleshoot, debug, fix, shutdown, stop, or tear down any RAG feature or service (Agentic RAG, VLM, guardrails, query rewriting, models, search, ingestion, observability, summarization, reasoning, and more).
>- monitoring, latency/throughput benchmarking (use rag-perf), or evals outside this repo layout.
>- load test driven by a single YAML config. Not for accuracy / RAGAS scoring (use rag-eval) or for deploying / repairing services (use rag-blueprint).
Use only to generate or update a governance skill card for a specified existing agent skill directory. Do not use for explaining, listing, comparing, or discussing skill capabilities.
Performs deep Root Cause Analysis (RCA) on NVIDIA TAO Visual ChangeNet classification experiments with image-evidence-driven investigation. Use when analyzing ChangeNet model failures, investigating poor recall / FAR / PASS-NO_PASS metrics, auditing visual inspection pipeline quality, or running an RCA report for an AOI defect-detection model. Trigger phrases include "RCA on my ChangeNet model", "why is my AOI model failing", "audit ChangeNet predictions", "investigate FAR regressions", "root cause analysis on visual-changenet".
Performs gap analysis on NVIDIA TAO VCN Classify (Visual Component Net) experiments by invoking the data-services container (`tao_toolkit.data_services` from `versions.yaml`) directly via `docker run … gap_analysis vcn_aoi …` — picks the optimal decision threshold, ranks per-sample weakness, and emits a top-K weakest parquet expanded per-lighting for downstream augmentation. Use when analyzing VCN classification failures, picking SDA augmentation targets, or auditing PASS/NO_PASS boundary cases.
Extract false-positive and false-negative gaps from VLM binary-classification-question (BCQ, yes/no) predictions. Use when the user asks to "analyze VLM BCQ gaps", "extract VLM false positives and false negatives", or identify failure cases from a predictions JSON for DEFT root-cause analysis on a binary-classification VLM workflow.
Use when selecting, configuring, composing, or troubleshooting NeMo Agent Toolkit agents and control-flow components, including ReAct, tool-calling, ReWOO, reasoning, router, sequential, parallel, and sub-agent patterns.
Use when designing, configuring, running, or troubleshooting NeMo Agent Toolkit evaluations, datasets, evaluator selection, ATIF surfaces, quality gates, custom evaluators, and `nat eval`.
Use when installing or configuring NVIDIA NeMo Agent Toolkit, verifying the `nat` CLI, setting up optional extras, or creating a first hello-world workflow.
Use when serving NeMo Agent Toolkit workflows, exposing workflows through FastAPI, configuring MCP clients or servers, or troubleshooting transport and server setup.
Use when configuring or running NeMo Agent Toolkit optimization with `nat optimize`, including Optuna parameter tuning, prompt evolution, optimizer sizing, output interpretation, and optimizer datasets.
Use when fixing NeMo Agent Toolkit documentation path-check failures, especially failed `ci/scripts/path_checks.py` output, slash-delimited text mistaken for paths, relative path references, Markdown code escaping, and path-check allowlist decisions.
Use when adding, configuring, or troubleshooting NeMo Agent Toolkit logging, tracing, telemetry exporters, OpenTelemetry, Langfuse, LangSmith, Weave, Phoenix, profiling, or observability provider integrations.
Use when authoring, registering, composing, or testing custom NeMo Agent Toolkit tools, functions, function groups, Python components, custom agents, custom evaluators, or advanced extension patterns.
Use first for general NVIDIA NeMo Agent Toolkit coding-agent behavior, task routing, naming conventions, component discovery rules, and cross-skill guidance.
Use when creating, editing, validating, running, or troubleshooting NeMo Agent Toolkit workflow YAML, component discovery, LLM configuration, and common `nat` CLI commands.
Use before creating, editing, or deciding whether to update any AI coding agent skill in this repository, including corrections to existing skill behavior, references, or routing.
Use when editing or reviewing CCCL code for style conventions; read common CCCL guidance and the path-specific references named by this skill.
Use when writing, updating, reviewing, or validating CCCL tests; read common CCCL test guidance and the path-specific references named by this skill.
Use when asked to check for SASS (or PTX) changes between commits, branches, or a local changeset; guides normalization, comparison, and reporting of CUDA disassembly diffs.
Use before MR handoff or when reviewing DCGM Exporter changes.
Use when adding or selecting tests for DCGM Exporter changes.
Use when validating DCGM Exporter in a local GPU-backed k3d/Kubernetes environment.
Use when changing metric CSV files, exporter-owned counters, Prometheus labels, or metric rendering.
Deploy a Kubernetes GPU cluster with DeepOps (Kubespray + GPU Operator) and prove it schedules GPU pods. Use when asked to deploy or rebuild Kubernetes on GPU servers with this repository.
Deploy a Slurm GPU cluster with DeepOps and prove it works. Use when asked to deploy, install, or rebuild Slurm on one or more GPU servers with this repository.
Diagnose NVIDIA driver installation failures on DeepOps-managed nodes — nvidia-smi errors, "No devices were found", DKMS build failures, or GPU pods crash-looping. Use before reinstalling anything.
Check whether a DeepOps-deployed Slurm or Kubernetes GPU cluster is healthy and report a machine-readable verdict. Use for health checks, post-deploy verification, "is the cluster working?" questions, and after any node or driver change.
Teaches the analyst agent how to write correct, robust Python analysis code for FHIR clinical data using pandas, matplotlib, and scipy.
Prepare a complete clinical case summary for a patient from FHIR endpoints. Use when asked to summarize a patient, compile a case, or prepare for tumor board.
How to delegate clinical tasks to specialist agents. Always use sub-agent runtime with explicit agentId — never ACP. Never call FHIR via web_fetch.
Teaches agents clinical reference ranges, condition codes, quality measure definitions, drug classifications, and regulatory context so they can flag abnormal values and identify care gaps.
Analyze a cohort of patients from FHIR endpoints to find care gaps and patterns. Use when asked to compare patients, find quality gaps, or analyze a population.
Diagnose common DGX Station GB300 issues — CUDA crashes, wrong-GPU targeting, vLLM/SGLang container bugs, MIG state problems, NVLink/Fabric Manager errors, X/Vulkan failures, HuggingFace auth, and port conflicts. Use when the user reports a GPU error, inference server crash, MIG problem, or any unexplained DGX Station failure.
Inspect and guide NVIDIA DGX Station GB300 development using the local dgx-assist CLI and pinned NVIDIA playbooks. Use for general Station platform questions, Software 1.0 or 2.0 compatibility, GB300 or RTX GPU selection, UUID ordering, mixed ATS/HMM coherency, CDMM, general containers, CDI, CUDA visibility, or vsloshd power-sloshing behavior. Do not use for vLLM or SGLang container selection or tuning, serving a named model, changing MIG, or troubleshooting a reported failure when the dedicated Station skill applies.
Run and interpret the complete read-only dgx-assist diagnostic suite for NVIDIA DGX Station GB300, correlate findings with pinned NVIDIA playbooks, export a redacted support bundle, and apply one separately approved allowlisted fix. Use when the user reports a Station, CUDA, GPU health, coherency, vsloshd, Docker, CDI, MIG, cache, port, or owned inference-service failure.
Resolve, tune, preflight, launch, verify, inspect, and stop exact-model inference on NVIDIA DGX Station through dgx-assist. Use for vLLM or SGLang container selection, NGC versus upstream, GPU memory utilization, CPU or KV offload, HBM fit, KV-cache sizing, ISL or context length, prefix caching, chunked prefill, batching, concurrency, performance tuning, serving or deploying a named model, an OpenAI-compatible endpoint, Station recipe models, or an owned inference service. Require an exact model ID for recipe resolution or model-specific tuning, and never recommend or substitute a different model.
Inspect NVIDIA DGX Station GB300 MIG state and installed-driver profiles, create a digest-bound layout plan, disclose disruption and restoration, and apply an approved still-valid plan. Use when the user asks to enable, disable, partition, reconfigure, inspect, or troubleshoot MIG instances or needs MIG UUIDs. Never assume static profile IDs or terminate GPU clients.
Teaches agents how FHIR R4 APIs work, what resources are available, how to query them with search parameters, and how to correctly parse all response formats including component Observations.
Configure NVIDIA MIG (Multi-Instance GPU) partitions on the DGX Station GB300, including enabling MIG mode, choosing a profile layout, creating instances, and retrieving MIG UUIDs. Use when the user asks to partition the GB300, set up MIG, run multiple models in isolation on one GPU, or reconfigure existing MIG instances.
Visualize drug-protein complexes using build_viewer.py, PubChem, and OpenFold3 NIM. Use when asked to show a molecular structure, drug target, or protein visualization.
Deploy an SGLang inference server on an NVIDIA DGX Station GB300 with the cu130 container, RadixAttention prefix caching, and structured JSON output support. Use when the user asks to serve a model with SGLang, start an SGLang endpoint, or needs structured-output inference on DGX Station.
Deploy a vLLM inference server on an NVIDIA DGX Station GB300 with validated container, GPU targeting, and tuning parameters. Use when the user asks to serve a model with vLLM, start a vLLM endpoint, or set up OpenAI-compatible inference on DGX Station.
Use for CUDA-Q setup, simulation targets, QPU access, and @cudaq.kernel authoring guidance.
Use when porting Qiskit Python circuits to CUDA-Q kernels while preserving algorithms and validation fidelity.
> Bump the Earth2Studio version on main to start a new development cycle. This unblocks commits on main after a release branch has been merged. Performs a CHANGELOG update (new blank section) and a hatch version bump. Use when main already has the release merge but still carries the old version string, or when the developer-release-rebase skill detects the version was already bumped during rebase and only the changelog/version housekeeping remains.
> Prepare a new minor alpha release of Earth2Studio by rebasing the release candidate branch onto main, bumping the version, updating the changelog, updating the README latest-news highlights, stripping example version tags, and pushing for PR. Use when releasing, cutting a release, preparing a release branch, rebasing a release, or bumping the version for a new development cycle.
> Create and validate Earth2Studio data source wrappers (DataSource, ForecastSource, DataFrameSource, ForecastFrameSource) from remote stores. Do NOT use for fetching data with existing sources, model inference, or installation tasks.
> Create Earth2Studio diagnostic model wrappers for single-step data transformations, including simple derived diagnostics, packaged AutoModel diagnostics, and generative or diffusion diagnostics. Do NOT use for prognostic time-stepping models, data sources, or installation.
> Create Earth2Studio prognostic (time-stepping forecast) model wrappers. Do NOT use for diagnostic models, data sources, or installation.
> Fetch weather/climate data via Earth2Studio data sources for specific variables and times. Do NOT use for inference pipelines, model discovery, or installation.
> Build deterministic forecast scripts with Earth2Studio (model, data source, IO, inference). Do NOT use for ensemble, diagnostics, data-only fetch, or install.
> Find Earth2Studio models, data sources, and examples for a weather/climate use case. Do NOT use for writing inference code, downloading data, or installation.
> Guide installing Earth2Studio via uv or pip, selecting model extras, and configuring the environment. Do NOT use for writing inference code, choosing models, or PhysicsNeMo questions.
Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Use for solver internals, PRs, DCO, and code conventions.
Install cuOpt for Python, C, or server via pip, conda, or Docker; verify the install. For building cuOpt from source, see cuopt-developer.
Trace, complete, and interpret the Pareto frontier across competing objectives using repeated single-objective cuOpt solves (weighted-sum and ε-constraint).
LP, MILP, and QP (beta) with cuOpt — Python, C, and CLI. Use when the user is solving LP, MILP, or QP with any cuOpt interface.
LP, MILP, QP — concepts, problem-text parsing, and formulation patterns (parameters, constraints, decisions, objective). Concepts only; no API.
Vehicle routing (VRP, TSP, PDP) with cuOpt — Python API only. Use when the user is building or solving routing in Python.
cuOpt REST server — start server, endpoints, Python/curl client examples. Use when the user is deploying or calling the REST API.
After solving a non-trivial problem, detect generalizable learnings and propose skill updates. Always active — applies to every interaction.
Base rules for end users calling NVIDIA cuOpt (routing/LP/MILP/QP/install/server). Not for cuOpt internals — use cuopt-developer for those.
Assists with benchmarking and profiling the performance of an Apache Spark UDF on the GPU. This is step 3 of 3 in the UDF conversion workflow (udf-gen-test -> udf-convert-to-* -> udf-benchmark). Use this skill when you have a CPU UDF and a RapidsUDF or SQL implementation, and need to benchmark the performance of the CPU UDF against the GPU implementation.
Assists with converting a non-aggregating Apache Spark UDF to a native CUDA RapidsUDF using JNI and libcudf. This is step 2 of 3 in the UDF conversion workflow (udf-gen-test -> udf-convert-to-cuda -> udf-benchmark). Use this skill when you have a CPU UDF with a unit test and need to convert it to a native CUDA implementation. Prefer udf-convert-to-cudf unless a CUDA implementation is necessary for performance or correctness, or if requested by the user.
Assists with converting an Apache Spark UDF to a GPU-accelerated RapidsUDF using cuDF Java APIs. This is step 2 of 3 in the UDF conversion workflow (udf-gen-test -> udf-convert-to-cudf -> udf-benchmark). Use this skill when you have a CPU UDF with a unit test and need to convert it to a RapidsUDF.
Assists with converting an Apache Spark UDF to a functionally equivalent Spark SQL expression. This is step 2 of 3 in the UDF conversion workflow (udf-gen-test -> udf-convert-to-sql -> udf-benchmark). Use this skill when you have a CPU UDF with a unit test and need to convert it to SQL for GPU acceleration.
Assists with generating a unit test for an Apache Spark UDF. This is step 1 of 3 in the UDF conversion workflow (udf-gen-test -> udf-convert-to-* -> udf-benchmark). Use this skill when you have a CPU UDF and need to create a unit test for the UDF before converting it into a GPU-compatible implementation.
Reviews generated UDF tests and GPU/SQL implementations for robustness, anti-cheating, and GPU execution integrity. Use when the user requests a judge/review-agent pass, or when manually reviewing a completed conversion.
Iteratively optimizes a cuDF RapidsUDF implementation for GPU performance. Use after testing and benchmarking with udf-benchmark. Runs a loop of profiling, optimizing, testing, and benchmarking until performance converges or the iteration budget is exhausted.
Help coding agents work on an NVFlare-based Auto-FL harness that follows an autoresearch-style loop. Use when the user wants to create, edit, debug, or extend program.md, task folders such as tasks/cifar10/ and tasks/vlm_med/, task-local job.py, client.py, model.py, shared custom_aggregators.py, mutation policies, results.tsv logging, or coding-agent prompts for a bounded federated-learning research loop. This skill is specifically for NVFlare harness work where the Client API loop, DIFF upload contract, and NUM_STEPS_CURRENT_ROUND metadata must stay intact unless the user explicitly asks for a protocol upgrade.
Generate and commit a markdown report after an Auto-FL NVFlare autoresearch experiment has been manually stopped. Use when the user asks to summarize a stopped campaign, report achieved improvements, explain implemented literature-derived ideas and sources, refresh progress plots, capture pasted agent model/effort/cost context when available, or commit the final report and progress plot to the current experiment branch.
Use for agent-assisted Auto-FL optimization of an existing NVFLARE job in simulation, POC, or production. Do not use for code conversion, diagnosis-only work, or deployment setup.
Generate a reproducible final report, literature-outcome synthesis, JSON summary, and refreshed progress plot for a stopped or interrupted NVFLARE Auto-FL campaign.
Convert existing Hugging Face Transformers Trainer or TRL SFTTrainer training code into an NVFLARE federated job using flare.patch(trainer), local validation, and job export; do not use for manual PyTorch loops, Lightning, inference-only pipelines, deployment, or experiment workflows.
Convert existing PyTorch Lightning training code into an NVFLARE federated job using the Lightning Client API patch, local validation, and job export; do not use for plain PyTorch, other frameworks, deployment, POC/production lifecycle, or experiment workflows.
Convert existing PyTorch training code into an NVFLARE federated job using Client API model exchange, local validation, and job export; do not use for other frameworks, deployment, POC/production lifecycle, or experiment workflows.
Use when the user asks why a reported NVFLARE job failure signal occurred: the job failed, stalled, timed out, lost clients, ended with EXECUTION_EXCEPTION, or produced suspicious errors. Diagnose in simulation, POC, or production by collecting bounded evidence and mapping failure patterns to recovery actions.
Compute federated statistics over tabular data (count, sum, mean, stddev, var, histogram, quantile, noise-protected min/max) and image data (count, failure_count, pixel-intensity histogram) across NVFLARE sites via FedStatsRecipe — automatic and non-interactive from the dataset, feature names (header or supplied), and optionally a README or notes declaring which statistics to compute; do not use for model training conversion, hierarchical statistics, deployment, POC/production lifecycle, or failed-job diagnosis.
Route open-ended or ambiguous NVFLARE requests by inspecting the local project and recommending one specific workflow skill without editing files; explicit conversions normally route directly to a converter, except when inspection reports unresolved Trainer ownership or active Lightning and Hugging Face Trainer entrypoints.
Shared NVFLARE conversion references and templates used by the other NVFLARE agent skills (conversion workflow, validation ladder, dependency install, model exchange, metrics/artifact reporting, and the custom aggregator template). Not a user-triggered skill; loaded via references from the conversion skills.
Use this skill when the user invokes /review-nvbench to review an NVBench pull request, PR branch, commit range, or diff. Supports context, feedback, and adversarial modes.
Use when integrating a CuTeDSL/CUTE DSL kernel into cuDNN Frontend as a frontend-only Python API, including APIBase wrappers, lazy cudnn exports, optional cutedsl dependencies, FE OSS documentation, and pytest coverage.
Add a new cuTile GPU kernel operator to TileGym. Covers dispatch registration in ops.py, cuTile backend implementation, __init__.py exports, test creation, and benchmark in tests/benchmark. Use when adding, creating, or implementing a new cuTile operator/kernel in TileGym, or when asking how to register a new cuTile op.
Converts cuTile Python GPU kernels (@ct.kernel) to cuTile.jl Julia equivalents. Handles kernel syntax translation, 0-indexed to 1-indexed conversion, broadcasting differences, memory layout (row-major to column-major), type system mapping, and launch API differences. Use when converting, porting, or translating cuTile Python kernels to Julia cuTile.jl, or debugging/optimizing existing Julia cuTile translations.
Converts cuTile GPU kernels (@ct.kernel) to Triton (@triton.jit). Handles standard in-repo conversion, debugging (cudaErrorIllegalAddress, shape mismatch, numerical mismatch), and mapping cuTile idioms (ct.load/ct.store, ct.Constant, ct.launch) to Triton equivalents. Covers dual-kernel layout flags (e.g. transpose=True/False + autotune grid via META) per translations/advanced-patterns.md. Use when converting, porting, or translating cuTile kernels to Triton, or debugging existing Triton translations.
| Use this skill to convert, port, or translate Triton-TileIR (nvtriton) or cuTile-Python GPU kernels to cutile-rs (Rust). The orchestrator runs scripts/preflight.sh, then drives a bounded Agent A -> B -> D -> E pipeline (Agent C is diagnostic, Agent F optional), delegating all kernel/host/correctness/perf work to sub-agents and routing by each stage's single-line VERDICT.
Use when adding, modifying, optimizing, or debugging CuTile autotuning code. Trigger signals: `exhaustive_search` / `replace_hints` / `hints_fn` / `cuda.tile.tune` in code, `autotune` in filenames, or correctness/performance issues in autotuned CuTile kernels. Covers: tune-once/cache/launch pattern, per-architecture configs (sm80–sm120), parameter space design (tile sizes, occupancy, num_ctas), and 7 common pitfalls with solutions.
Expert cuTile programming assistant. Write high-performance GPU kernels using cuTile's tile-based programming model with proper validation and optimization. Supports deep agent orchestration for complex multi-kernel tasks.
Iteratively optimize cuTile kernel performance through systematic profiling, bottleneck analysis, IR comparison, and targeted tuning. Covers tile sizes, occupancy, autotune configs, TMA, latency hints, persistent scheduling, num_ctas, flush_to_zero, and IR-level debugging. Use when asked to "optimize cutile kernel", "improve kernel perf", "tune cutile performance", "make kernel faster", or iteratively benchmark and refine a cuTile GPU kernel in the TileGym project.
Integrate TileGym kernels into Hugging Face `transformers` models by replacing the library's submodule(s) and certain class(es)' implementations, and patching certain class(es)' init/forward/load weight methods prior to instantiating models. Used when the user requires integrating TileGym kernels into `transformers` models.
> Expert guide for adding, optimizing, or debugging DISCO convolution kernels kernel variant or GPU arch path, wiring up a kernel through the Python dispatch layer, fixing bf16/fp16/autocast issues in DISCO, profiling or benchmarking the CSR or kpacked forward/backward, propagating kernel changes to the distributed DISCO path, or writing tests that cover all dispatch branches. Also use for debugging correctness issues in the DISCO fwd/bwd path or distributed a2a collectives.
> Guide for writing tests in torch-harmonics. Use this skill whenever the user asks to add tests, check or extend test coverage, write a new test class, test a new layer or kernel, or add distributed tests. Also use when the user asks about tolerance values, how to compare tensors, how to structure distributed test infrastructure, or how to test gradients. Covers serial tests in test_convolution.py and distributed tests in test_distributed_*.py.
> Index of project-local Claude Code skills for torch-harmonics. Use this to discover which skills are available for this codebase. Individual skills live in the .skills/ directory at the project root.
> Analyze NVIDIA Nsight Compute (ncu) profiling reports (.ncu-rep files). Extract metrics, performance data, SASS/CUDA source, and identify bottlenecks. Nsight Compute report, kernel performance/profiling data from ncu, or asks to generate/collect an ncu profile for a tilus kernel or example script.
> Convention and format for writing instruction docstrings and RST tutorials in documentation for tilus instructions, instruction groups, or tutorials.
Troubleshoot cuOpt LP/MILP problems including errors, wrong results, infeasible solutions, performance issues, and status codes. Use when the user says something isn't working, gets unexpected results, or needs help diagnosing issues.
Map interpreted optimization problems into cuOpt-native models for the fast path with minimal clarifying questions.
Run cuOpt in the NemoClaw sandbox — probe/smoke gates, prefer cancelable Python gRPC jobs, use legacy remote execution only when that API is unavailable, then vendored cuOpt skills.
> variable/constraint reference for the max-supply base model.
Coordinate uploaded data plus a natural-language question into interpretation, clarification, cuOpt solve, and a user-facing answer.
Classify whether a data-backed request is LP, MILP, QP, routing, or non-optimization analytics.
Choose fast direct-to-cuOpt solve versus replayable or auditable model artifact mode.
Infer optimization structure from uploaded tables and identify minimal clarifications before cuOpt modeling.
> Apply FlashDreams-style inference speedups to model integrations after a `torch.compile`, CUDA graph capture, attention backend checks, decoder layout or replacement, transfer/materialization changes, and ordered presentation tuning. Use when porting known optimizations into a runner, demo, serving adapter, or downstream integration while preserving quality and reset behavior.
Navigate the FlashDreams package layout and integration architecture: core vs infra vs recipes vs workspace integrations, Transformer/Encoder/StreamingDecoder/Pipeline config contracts, AR caches, CP, CFG, KV cache, CUDA graph wrapping, runner registration, and test placement. Use when adding or editing built-in recipes under flashdreams/flashdreams/recipes/, plugin packages under integrations/, model configs, pipeline wiring, or when the user asks where code belongs. The template recipe is the reference design.
Add or modify FlashDreams video post-processing processors, sessions, presets, and runner stream wiring. Use when implementing a new VideoPostProcessorConfig / VideoPostProcessor / VideoPostProcessorSession, registering a --postprocess.preset entry point, changing VideoPostprocessStream behavior, or reasoning about streaming buffering, layouts, per-view processing, distributed execution, or postprocess tests.
End-to-end workflow for porting an external video diffusion model into a flashdreams integration — scope the architecture, scaffold a workspace-member plugin, reuse an existing recipe, write the checkpoint key-remap, layer model-specific conditioners, wire the runner, and verify with checkpoint weight-equality + upstream parity + a GPU rollout. Use when integrating a new model (e.g. a HuggingFace/research release) into flashdreams or a downstream repo, porting upstream weights, or reproducing an existing integration. Pairs with the `flashdreams-integrations` skill (architecture map) — this skill is the ordered procedure; that one is the contract reference.
Maintain FlashDreams's OSS-release state — the LICENSE / NOTICE / THIRD-PARTY-NOTICES / REUSE.toml / LICENSES/ / CONTRIBUTING.md collateral that satisfies OSRB Bug 6107043, the per-file SPDX headers, the third-party dependency manifest in THIRD-PARTY-NOTICES, and the pyproject.toml + uv.lock dependency pins. Use when adding or upgrading a runtime dependency, vendoring third-party source into the repo, adding a new first-party source file (any .py / .pyx / .pyi / .c / .cc / .cpp / .h / .hpp / .cu / .cuh / .sh / .proto / Dockerfile), reviewing whether a change requires reopening an OSRB bug or filing a self-cert, or triaging a reuse-lint CI failure.
> Inspect and baseline performance for FlashDreams-style model integrations and build focused probes, and identify whether decode, model/denoise, cache, data transfer, or presentation dominates. Use when starting performance work on an existing model runner, demo, serving path, or downstream integration before implementing speedups. Pair with `apply-inference-optimizations` after the bottleneck is known and `validate-performance-quality` for benchmark and quality gates.
Write Python docstrings and inline comments matching the flashdreams house style — SPDX header, one-line module docstring, Google-style function docstrings (Args/Returns/Raises), PEP 257 attribute docstrings on dataclass/class fields *and on module-level constants*, double-backticks for code references, imperative first sentences, and signpost-style inline block comments (kept, not stripped, on a tightening pass). Use when authoring or editing any .py file under flashdreams/, when adding a new module/class/function/field/constant, when polishing comments, or when the user asks about docstring or comment style.
Design benchmark, quality, and documentation validation for FlashDreams-style performance changes. Use when adding or updating sweep commands, profiler probes, decoder-quality comparisons, compile/cache probes, manual GPU validation, performance summaries, model cards, or README guidance after optimizing a model integration, demo, or serving path.
> Navigate the Cosmos3 package codebase to find where parameters, configs, defaults, scripts, and documentation live. Use when the user asks "where is X in cosmos3", "how do I find the config for Y", "where are the defaults", "where do I change a parameter", or any question about locating files, modules, or settings. Also use when the user opens or edits files and needs orientation.
> Navigate the Cosmos3 package codebase to find where parameters, configs, defaults, scripts, and documentation live. Use when the user asks "where is X in cosmos3", "how do I find the config for Y", "where are the defaults", "where do I change a parameter", or any question about locating files, modules, or settings. Also use when the user opens or edits files and needs orientation.
> Diagnose and fix Cosmos3 environment, installation, and runtime errors. Use when the user encounters an ImportError, ModuleNotFoundError, CUDA error, Docker error, checkpoint download failure, or any traceback during setup or inference.
> Diagnose and fix Cosmos3 environment, installation, and runtime errors. Use when the user encounters an ImportError, ModuleNotFoundError, CUDA error, Docker error, checkpoint download failure, or any traceback during setup or inference.
> Guide users through running Cosmos3 inference — offline batch generation, online serving with Ray and Gradio, parallelism options, input formats, sampling parameters, and prompt upsampling. Use when the user asks "how do I run inference", "how do I generate a video", "how do I serve the model", "what parameters should I use", or any question about running the model to produce outputs.
> preparing the example dataset and Wan2.2 VAE, converting the base checkpoint to DCP, launching distributed training (paired launch shell recommended, raw `torchrun` as an alternative), running T2V/I2V/V2V inference with the trained DCP checkpoint, and optionally exporting it to Hugging Face safetensors. Use when the user asks how to post-train Cosmos3, fine-tune on a custom video dataset, export a trained checkpoint, or invoke one of the recipe launch shells (`launch_sft_vision_nano.sh`, `launch_sft_llava_ov.sh`, `launch_sft_videophy2_nano.sh`, plus the `_super` LoRA variant) — or any question about `cu130-train` / `cu128-train`, `convert_model_to_dcp` / `export_model` / `train`, or SFT output paths. For dataset captioning / JSONL assembly, see `docs/dataset_jsonl.md`.
> preparing the example dataset and Wan2.2 VAE, converting the base checkpoint to DCP, launching distributed training (paired launch shell recommended, raw `torchrun` as an alternative), running T2V/I2V/V2V inference with the trained DCP checkpoint, and optionally exporting it to Hugging Face safetensors. Use when the user asks how to post-train Cosmos3, fine-tune on a custom video dataset, export a trained checkpoint, or invoke one of the recipe launch shells (`launch_sft_vision_nano.sh`, `launch_sft_llava_ov.sh`, `launch_sft_videophy2_nano.sh`, plus the `_super` LoRA variant) — or any question about `cu130-train` / `cu128-train`, `convert_model_to_dcp` / `export_model` / `train`, or SFT output paths. For dataset captioning / JSONL assembly, see `docs/dataset_jsonl.md`.
> Guide users through Cosmos3 installation, environment setup, checkpoint downloading, and verification. Use when the user asks "how do I install cosmos3", "how do I set up the environment", "how do I download checkpoints", "how do I use Docker", or any question about getting the package running for the first time.
Define custom groups (Irrep subclasses), build segmented tensor products with CG coefficients, create equivariant polynomials and IrDictPolynomials, and use built-in descriptors (linear, tensor products, spherical harmonics). Use when working with cuequivariance group theory, irreps, or segmented polynomials.
Execute equivariant polynomials in JAX using segmented_polynomial (naive/uniform_1d), the ir_dict workflow with IrDictPolynomial and dict[Irrep, Array], and Flax NNX layers (IrrepsLinear, SphericalHarmonics, IrrepsIndexedLinear). Use when writing JAX code with cuequivariance.
Execute equivariant tensor products in PyTorch using SegmentedPolynomial (naive/uniform_1d/fused_tp/indexed_linear), high-level operations (ChannelWiseTensorProduct, FullyConnectedTensorProduct, Linear, SymmetricContraction, SphericalHarmonics, Rotation), and layers (BatchNorm, FullyConnectedTensorProductConv). Use when writing PyTorch code with cuequivariance.
| Use when analyzing an AICR snapshot YAML file, reviewing cluster state, comparing provider characteristics, extracting GPU/network topology insights, or generating a cluster assessment report from a snapshot. GPU topology, node health, snapshot report.
| Use when reviewing AICR's Markdown documentation for duplication, drift, bloat, and gaps — to keep docs high-value as the project evolves. Triggers on "audit the docs", "review documentation", "docs cleanup", "/aicr-auditing-docs", or any request to find redundant/stale/missing docs across README, docs/, demos/, and the root governance files. Produces a prioritized findings report (research, not edits) grouped by the five audit dimensions, anchored to the project's canonical sources of truth.
Scaffolds an interactive guided demo script (demos/*.sh), live or self-paced, with the Frame → Tell → Show → Close pattern. Triggers on "demo script", "guided walkthrough", "demos/*.sh", "live demo".
Use when building a self-contained HTML slide deck or visual talking-point for a technical concept or workflow (e.g. a demos/*.html) — shown full-screen or projected and narrated, opening in any browser with no build step or dependencies.
| Multi-agent PR review using Claude Code, Codex, and CodeRabbit. Runs parallel reviews with integration impact analysis, then one cross-review round to a 2-of-3 consensus, with every confirmed finding adversarially verified by a fresh agent. Never runs the reviewed commit's code, and never posts unless explicitly asked. Use when asked for a thorough cross-review or multi-reviewer analysis. Requires the Codex plugin; CodeRabbit is best-effort. Claude Code only — uses the Workflow and Agent tools, which are not available in other agents.
| Use when adding, updating, or removing CVE/GHSA suppressions in `.openvex.json` — the OpenVEX document consumed by the daily image vulnerability scan workflow. Triggers on "VEX", "OpenVEX", ".openvex.json", "suppress CVE", "ignore CVE", "vulnerability suppression", "aiperf-bench CVE", or any request to act on findings reported by `Daily Image Vulnerability Scan` for the aiperf-bench for new HIGH+ findings, drops statements that no longer apply (dependency upgraded past the fix, advisory recalled, package removed), and verifies suppressions actually land in the JSON output.
| Use when drafting the human-readable GitHub release notes summary for an upcoming AICR release. Triggers on "release notes", "draft release notes", "/aicr-release-notes", or any request to summarize commits since the last tag into a polished release announcement. Runs tools/changelog, groups commits into thematic highlights, mirrors the style of the previous release, and writes a Markdown draft to a temp file for hand-editing before publishing.
| Use when the user runs `/aicr-triage` or asks to triage, review, or clean up a GitHub org-level Projects v2 board (default NVIDIA AICR project 248). Reviews active non-Done issues, then promotes P2 issues to P1, demotes Ready items to Backlog, closes superseded issues, and classifies unclassified ones — applying only user-confirmed changes via `gh` CLI. Triggers on backlog hygiene before sprint planning or release prep, or when the user files a new issue and asks to classify it on the board.
| Use when reporting on UAT health across services and GPU targets — which service (EKS/GKE/AKS) x GPU (H100/GB200) x intent combinations are passing or failing in the UAT Run workflow (uat-run.yaml). Triggers on "UAT report", "/aicr-uat-report", "which UAT combos are failing", "UAT pass rate", "download the UAT debug bundle", "why did the UAT run fail", or RC/release-candidate validation prep that needs the combinations to test manually. Runs the bundled uat_report.py, classifies failures as product vs infra signal, prints a summary table plus an RC validation priority list, and can download the per-run cluster debug bundles for triage.
>- IsaacTeleop CMake target layout and #include conventions. Use whenever you create or edit a CMakeLists.txt, add or move a module/library/executable/test target, set up target_include_directories or target_link_libraries, decide where a header file goes (private vs public inc/), write or reorder #include directives in C/C++ sources, or restructure directories under src/ or deps/. Also use when reviewing diffs that touch CMakeLists.txt, header placement, include paths, or the overall project/directory structure.
> Closed-loop fault injection and attribution accuracy benchmark. Draws from a prioritized pool of (fault_type, rank, iter, nodes) experiments and submits them 2 at a time via sbatch — waiting for each pair to finish before submitting the next — to bound filesystem load. GPU-related faults are front-loaded in the pool. After all jobs complete, runs /log-analysis and /fr-analysis on every experiment, scores attribution vs. ground truth, aggregates gaps, and iterates on attribution modules to close them.
> Analyze PyTorch NCCL flight-recorder (FR) dumps to identify collective operation hangs and isolate the responsible ranks using CollectiveAnalyzer. Use when a distributed training job hangs due to an NCCL collective timeout and FR dump files are available. Detects the wavefront process group where collectives diverge and returns the root-cause suspect ranks.
> Analyze a SLURM job log file for failure root-cause attribution and restart decisions using NVRxLogAnalyzer. Use when you have a SLURM training job log and need to determine why the job failed and whether it should be restarted. Performs per-cycle chunking, fast-path pattern matching, and LLM-based classification.
> Orchestration layer over nvidia_resiliency_ext attribution modules. Provides log-analysis, fr-analysis, and a Megatron-LM-oriented fault-injection feedback loop for benchmarking attribution quality on SLURM workloads.
Draft and revise concise, human-focused GitHub issues for pytest-kind-ng. Use when asked to write, rewrite, review, or prepare a bug report, enhancement request, maintenance issue, or follow-up issue for this repository, including turning code-review findings into issues.
Use for Kaggle competition overview fetches, writeups, discussion/kernel research, submissions, and dataset uploads. Not for unrelated ML code.
Build or migrate infra-controller REST API endpoints that call on-site NICo Core through the generic Core gRPC proxy. Use when working on REST-to-Core operations, ExecuteCoreGRPC, coreproxy, forge.Forge methods, creating new proxied REST endpoints, or migrating bespoke workflows to the gRPC proxy.
> Use only for offline/local OSMO service-config admin requests involving explicit config roots or values files, or to ask for one when a file-specific config request omits it. Do not inspect the workspace to infer a root. Do not use for live workflow support, resource capacity, pod/node diagnostics, or cluster operations, except live service-config paths that must be refused.
> How to deploy OSMO to a Kubernetes cluster on Azure (AKS), AWS (EKS), MicroK8s (single-node), or any kubectl-reachable cluster (BYO). Use this skill whenever the user asks to install, deploy, set up, or stand up OSMO; whenever they ask to provision an OSMO cluster; whenever they mention deploy-osmo-minimal.sh, deploy-k8s.sh, or "OSMO helm install"; whenever they ask to wire up workflow storage (MinIO / Azure Blob / S3); or whenever they ask to add a GPU pool to an OSMO cluster, install KAI scheduler, install the NVIDIA GPU Operator, or run the post-install smoke tests. Targets OSMO 6.3 (ConfigMap mode).
> check resources, submit/monitor/debug/explain workflows, fetch logs and Grafana/Kubernetes links, inspect direct data storage, manage workflow apps, and set workflow credentials. Use whenever the user asks about OSMO pools, quota, GPUs, or nodes, or about submitting, listing, querying, monitoring, or troubleshooting workflows — including failed, PENDING, queued, stuck, or image-pull-blocked workflows, or even when they describe a workflow, cluster resource, or storage URI without saying "OSMO". Do not use for Kubernetes admin, server-side `osmo config` changes, OSMO install/deploy, non-OSMO compute, or general NVIDIA hardware questions.
>- em-dash, or en-dash. Use when editing public docs, READMEs, AGENTS.md, agent skills, plans, GitHub issues, or GitHub Pull Request descriptions.
Release NVCA Operator chart changes from the native monorepo source to the vendored Helm chart. Use when updating the vendored NVCA Operator chart, changing NVCA image refs, publishing helm-nvca-operator, or validating the chart against a self-managed control plane.
Install or validate the NVCA Operator chart against a self-managed NVCF control plane from the native monorepo. Use when the control plane comes from deploy/stacks/self-managed and NVCA must be installed with stack-derived image repository settings.
Customize NVCA Operator Helm chart values in the native monorepo. Use when modifying vendored defaults, changing stack-derived install values, adding deployment-time overrides, or updating scripts under deploy/helm/nvca-operator.
>- Navigate and explain the NVCF self-hosted stack inside the monorepo. Maps helmfile releases to their charts, image-source subtrees, helm hooks, deploy/stacks/self-managed/helmfile.d/*.yaml.gotmpl and deploy/stacks/nvcf-compute-plane/helmfile.d/*.yaml.gotmpl as the source of truth for ordering and versions. Use when a user or developer asks "what deploys X", "what does X depend on", "what hooks run for X", "walk me through deployment order", "which subtree do I edit to change X", "what namespaces does the stack use", or mentions stack topology, helmfile stages, deployment DAG, or dependency map.
| Install, operate, and tear down self-hosted NVIDIA Cloud Functions (NVCF) deployments with nvcf-cli. Use for control-plane or compute-plane install, status checks, cluster registration, function deploy/invoke, task create/list/cancel/delete, API keys, admin tokens, JWKS rotation, nvcf, nvcf-cli, self-hosted nvcf, self-managed nvcf, NVCFBackend, NVCA, NCP, ICMS, helmfile, control plane, compute plane, LLM function, OpenAI-compatible invocation, Responses API, embeddings, batch task, task monitor, cluster rotate, cluster delete, helm task, helm-based task, task secrets, update task secrets, retrieve task results, bulk fetch tasks, task results, NVCT, NVCT task, NVCT batch job.
>- Install and operate NVCF self-hosted control-plane and separate compute-plane stacks. Covers Helmfile values and CLI profile installation flows, teardown, values overrides, pull secrets, and troubleshooting. Use for nvcf-self-managed-stack, nvcf-compute-plane-stack, split compute-plane installation, control-plane installation, CLI-generated control-plane profiles, Helmfile, self-managed, or self-hosted deployments. Do NOT use for local k3d environments; use the local k3d development workflow instead.
>- Install the prerequisites the NVCA operator / compute plane needs before compute-plane stack's register-cluster step), KAI Scheduler (for the KAIScheduler feature gate), and the SMB CSI driver (for the sharedStorage Samba sidecar PVCs). The two cluster components are cloud-neutral helm installs at the NVCF-validated version pins; same install on AKS, EKS, GKE, k3d, or bare metal. Use when the user mentions NVCA prereqs, nvcf-cli, "nvcf-cli not found", ensure-nvcf-cli, register-cluster, KAI Scheduler, SMB CSI, csi-driver-smb, queue quotas, default-parent-queue, NVCA shared-storage PVCs stuck Pending, or asks how to prepare a cluster before installing the NVCA operator.
Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras; VIOS records clips, AMC ingests them, then runs calibration.
Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.
Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP/live streams, use amc-run-rtsp-calibration instead.
Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.
NVIDIA DeepStream SDK development with Python pyservicemaker API. Use when building video analytics pipelines, GStreamer-based video processing, TensorRT inference integration, object detection/tracking, or Kafka/message broker integration.
Build DeepStream GStreamer pipelines interactively. Use when the user asks about pipelines for video/image inference, detection, tracking, or streaming — including natural phrases like 'pipeline to infer on image', 'run inference on video', 'detect objects in stream', 'save inference output', 'deepstream pipeline', 'gst-launch pipeline', 'process video with detection', 'build a pipeline', or any request involving GStreamer/DeepStream elements (nvinfer, nvstreammux, nvtracker, etc.).
> Use this skill to bring any vision model from HuggingFace or NVIDIA NGC into SafeTensors export, TRT engine build, custom nvinfer bbox parser, multi-stream benchmark, and PDF report. Object detection models only.
Profile a DeepStream pipeline with Nsight Systems and derive its configs from the measurement. Use when the user asks for an efficient, performant, or profiled pipeline — or to benchmark, tune, or measure FPS.
Run and operate the DeepStream Multi-View 3D Tracking reference app, also known as MV3DT. Use when the user asks to set up prerequisites, run shipped MV3DT samples, run Multi-View 3D Tracking on custom synchronized MP4 datasets, import camera calibration, delegate missing calibration to AutoMagicCalib, inspect OSD or BEV visualization, consume MV3DT Kafka metadata, or clean up MV3DT run state in the DeepStream MV3DT app directory.
> Use this skill when building, deploying, evaluating, debugging, or measuring latency for the DeepStream SOP Inference Microservice — a GPU-accelerated FastAPI service that detects whether operators perform assembly-line steps in order via event boundary detection (GEBD) plus VLM classification. Trigger missing or out-of-order SOP steps, score factory/work-cell video for procedure compliance, run VLM-based SOP checking on industrial cameras, or call /v1/chat/completions with a file, RTSP, or Basler camera. Also trigger Triton CAPI, nvds_custom_postprocess, Cosmos Reason 1/2 vLLM, SSE streaming, Kafka NvProto/JSON output, Basler/Pylon camera + emulation, Docker compose, chunk-level latency. Do NOT trigger for generic DeepStream pipelines, object detection/tracking, NIM imports, or video summarization.
Interact with GitHub using the gh CLI restricted to REST API only (no GraphQL). Use when the user wants to work with GitHub issues, pull requests, repos, releases, or Actions — especially in sandboxed environments where the GraphQL endpoint may be blocked. Trigger keywords - github, gh, pull request, PR, github issue, github actions, workflow, release, gh api.
Install NVIDIA NIM Operator on Kubernetes with prerequisite checks, optional NVIDIA GPU Operator dependency installation, public or local Helm chart selection, optional Dynamo support, and optional KServe compatibility verification. Use when a customer wants to install or upgrade the NIM Operator itself, with or without Dynamo and KServe, but does not want to deploy a NIM inference model yet.
Safely uninstall NVIDIA NIM Operator from Kubernetes with inventory checks, explicit approval gates for destructive actions, optional custom resource cleanup, optional CRD removal, and post-uninstall validation. Use when a customer wants to remove or clean up the NIM Operator itself, not the GPU Operator or unrelated cluster dependencies.
>- Create a custom metric for the PhysicsNeMo CFD benchmarking workflow. Use when the user wants to add a new evaluation metric, implement a custom error measure, compute force coefficients, or extend the benchmark with domain-specific quantities.
>- Create a new dataset adapter for the PhysicsNeMo CFD benchmarking workflow. Use when the user wants to add a new CFD dataset, write a DatasetAdapter, integrate a new mesh format, or benchmark models on custom data.
>- Create a new model wrapper for the PhysicsNeMo CFD benchmarking workflow. Use when the user wants to add a new CFD model, write a CFDModel wrapper, integrate a new neural network architecture, or run a custom model through the benchmarking pipeline.
>- How to write, read, compose, and load atomic data using nvalchemi's composable Zarr-backed storage pipeline (Writer, Reader, Dataset, MultiDataset, DataLoader). Use when saving simulation outputs or trajectories to disk, converting structures (e.g. ASE / extxyz) into Zarr stores, assembling datasets for training or inference, or wiring a DataLoader to stream batches to the GPU.
>- How to use AtomicData and Batch, the core graph-based data structures for representing atomic systems and batching them for GPU computation. Use when building systems from positions, cells, and atomic numbers, converting from ASE Atoms, batching or unbatching structures, reading per-atom vs per-graph tensors, or debugging shape, dtype, or device errors in model inputs.
>- How to configure and run dynamics simulations, compose multi-stage pipelines (FusedStage, DistributedPipeline), use inflight batching, and manage data sinks. Use when writing any simulation script — molecular dynamics (NVE/NVT), structure relaxation or geometry optimization (e.g. FIRE), equation-of-state or adsorption scans — or orchestrating many structures through a batched GPU pipeline.
>- How to use and write dynamics hooks — callbacks that observe or modify batch state at specific points during each simulation step. Use when a simulation needs neighbor-list rebuilds, convergence checks or early stopping, temperature control, per-step logging or trajectory capture, or any custom per-step behavior attached to a dynamics run.
>- How to implement a dynamics integrator by subclassing BaseDynamics and overriding pre_update() and post_update() methods. Use when creating a custom integrator, optimizer, or sampler that the built-in stages do not provide; for configuring existing dynamics, see nvalchemi-dynamics-api.
>- How to fine-tune nvalchemi-compatible models with FineTuningStrategy, pretrained checkpoint initialization, module patches, trainable-parameter filters, conservative optimizer defaults, validation, restart checkpoints, and model-agnostic MACE, AIMNet2, custom BaseModelMixin, or PyTorch inputs. Use when adapting a pretrained MLIP (e.g. MACE-MP) to new reference data, freezing or patching submodules during training, or resuming an interrupted fine-tune from a checkpoint.
>- How to use built-in loss functions and implement custom losses using the BaseLossFunction template-method pattern — residual types, per-atom normalization, masking, and graph-balanced reductions. Use when choosing or weighting energy, force, or stress objectives for training or fine-tuning, masking atoms or graphs out of the loss, or writing a custom loss term.
>- How to wrap an arbitrary MLIP (Machine Learning Interatomic Potential) using the BaseModelMixin interface to standardize inputs, outputs, and embeddings. Use when integrating a model such as MACE or AIMNet2 (e.g. MACEWrapper, loading pretrained checkpoints) so dynamics, training, or fine-tuning stages can call it, or when exposing energies, forces, or embeddings from a custom PyTorch model.
>- How to add observability to nvalchemi dynamics and training workflows using ReportingOrchestrator, RichReporter, TensorBoardReporter, scalar extraction, custom reporter callbacks, and dynamics LoggingHook. Use when showing live progress, writing TensorBoard summaries, preserving dynamics CSV rows, adding rank-safe distributed reporting, previewing Rich dashboards, or deciding between logging and reporting for training or molecular dynamics runs.
>- How to configure nvalchemi training workflows with TrainingStrategy, custom training functions, standalone or composed losses, loss-weight schedules, optimizer and scheduler configs, validation, hooks, restartable checkpoints, and model-agnostic inputs. Use when training a model from scratch or setting up optimizers, schedulers, validation, or checkpointing for a training run; for adapting a pretrained model, see nvalchemi-fine-tuning.
> Performance tuning for nvalchemi's Zarr-backed Reader, Dataset, and DataLoader pipeline. Use when configuring AtomicDataZarrReader, Dataset, DataLoader, ZarrWriteConfig, or nvalchemi-io-test for training/inference throughput, especially shuffled access, graph-like random access, fused prefetch, pinned memory, validation overhead, or Zarr chunk/shard choices.
Work with the nvtripy compilation pipeline. Use when: using tp.compile, creating InputInfo or DimensionInputInfo, understanding the Trace → MLIR → TensorRT flow, configuring optimization levels, working with Executable objects, debugging compilation, using dynamic shapes or NamedDimension.
Author input/output constraints for nvtripy operations using the declarative constraint DSL. Use when: defining input_requirements or output_guarantees, writing @wrappers.interface decorators, auto-casting dtypes, using GetInput/GetReturn/OneOf/If/Equal, debugging constraint validation errors.
Debug and diagnose errors in nvtripy code. Use when: interpreting TripyException stack traces, enabling MLIR/TensorRT debug output, understanding error reporting with stack_info, using raise_error, configuring debug environment variables, tracing compilation failures.
Write API documentation for nvtripy following project conventions. Use when: writing docstrings for ops or modules, adding code examples, using @export.public_api document_under paths, creating Sphinx RST cross-references, understanding the docs build pipeline.
Add a new neural network module to nvtripy. Use when: creating an nn layer, implementing a Module subclass, adding a new layer like Linear/LayerNorm/Conv, defining parameters with DefaultParameter or OptionalParameter, using constant_fields decorator.
Add a new operation to nvtripy. Use when: implementing a new op, adding a frontend op, creating a trace op, registering an op in the API. Covers the full Frontend → Trace → MLIR pipeline including export decorators, constraint definitions, and init registration.
Write tests for nvtripy following project conventions. Use when: adding tests for ops, modules, trace operations, or compilation, using pytest parametrize, testing error cases with helper.raises, testing dtype combinations, understanding test directory structure.
> Use when writing the objective_function= passed to Search(). Covers the two legal signatures (compiler-only str vs mixed list), the baseline-knockout branch, per-eval cache busting, framework-specific --apply-controls injection (raw PTXAS, NVCC, Triton, Helion, cuTeDSL/FA4, FlashInfer), correctness-before-timing, INVALID_SCORE handling, and the Debug-pack O0/O3 ACF-injection canary that must pass before launching a search. Triggers on "objective function", "apply-controls", "INVALID_SCORE", "save_compiler_config", "baseline knockout", "BASELINE_CONFIG", "every config returns the same score", "TypeError fromhex".
> Use BEFORE running a full CompileIQ search. Walks through downloading a Booster Pack from NVIDIA/CompileIQ GitHub Releases, applying ACF candidates one at a time to the user's compiler (raw PTXAS, NVCC, Triton, Helion, FlashInfer), and keeping only candidates that compile, pass correctness, and beat the no-ACF baseline. Includes the mandatory Debug-pack O0/O3 ACF-injection canary that proves the ACF is reaching PTXAS. Triggers on "booster pack", "ACF", "apply-controls", "speed up without searching", "helion fp8", "flashinfer batch decode", "debug pack".
> Use when starting a fresh CompileIQ project, hitting a socket timeout, or before running any other compileiq-* skill. Verifies CUDA 13.3+, ptxas, GPU access, that `from compileiq.ciq import Search` and friends resolve, and that `PtxasSearchSpace().retrieve()` returns a real path. Documents the env vars that control timeouts, caching, and search-space mirroring. Triggers on "set up compileiq", "compileiq doesn't work", "socket timeout", "where do search spaces come from", "air-gapped compileiq".
> INVALID_SCORE, scores aren't improving, every config returns the same number, ptxas errors fill the log, CV% is too high, or a winning ACF candidate needs NCU profiling to explain. Symptom-indexed table on top. Triggers on "compileiq hang", "socket timeout", "INVALID_SCORE", "not converging", "every score is the same", "TypeError fromhex", "ncu profile", "register spill", "ptxas error", "not in expected format", "high cv".
> Use when composing the Search(...) call and calling .start(). Covers the four worker classes (MultiProcessWorker / IsoMultiProcessWorker / RayWorker / AsyncWorker) and when to pick each, SearchConfiguration sizing rules, dump_results checkpointing, tracker_config choice (Disabled / Loguru / MLflow), num_workers/task_timeout semantics, and GPU clock locking for stable measurements. Triggers on "Search()", "tuner.start()", "pool_size", "num_workers", "task_timeout", "IsoMultiProcessWorker", "RayWorker", "dump_results", "MLflow", "GPU clocks".
> Use when picking the search_space= argument for Search(). Covers the three provider classes (PtxasSearchSpace, NvccSearchSpace, LocalSearchSpaceBin), how to pin a version/variant/tag, the attention-focused 'att' variant for attention kernels (FlashAttention, GQA, MHA, MLA, FlashInfer Batch Decode), air-gapped mirroring via CIQ_SEARCH_SPACES_DIR, and custom user-defined search spaces built from compileiq.search_spaces.base primitives. Triggers on "search space", "PtxasSearchSpace", "NvccSearchSpace", "air-gapped compileiq", "offline compileiq", "CIQ_SEARCH_SPACES_DIR", "attention variant", "ptxas att".
> Use AFTER a Search has completed and BEFORE claiming any speedup or shipping an ACF. Loads the dump_results CSV, extracts top-K candidates (single-objective) or the Pareto front (multi-objective), re-measures each against the no-ACF baseline with 100+ trials on fresh caches, runs Welch's t-test plus Cohen's d, rejects three classic false-positive patterns (lucky-min / higher-variance / multiple-comparisons-of-N), and saves the validated winner as best.acf. Triggers on "validate result", "extract best config", "Welch's t-test", "is my speedup real", "save best ACF", "pareto front", "claim speedup", "ship config".
Add or change a public NeMo Relay API surface across the core runtime and every affected binding
Add a new guardrail or intercept type to the NeMo Relay middleware pipeline
Contribute a new NeMo Relay public API surface safely, with binding parity and docs in mind
Contribute documentation or example changes that stay aligned with NeMo Relay public behavior
Contribute a new or updated third-party framework integration for NeMo Relay
Compare NeMo Relay release branches and draft or update the documentation-site release notes. Use when preparing a release-notes update under docs/about-nemo-relay/release-notes, reviewing release-to-release changes, or gathering evidence for a current release summary.
Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria.
Maintain and review NeMo Relay GitHub Actions workflows with explicit per-job permissions, pinned action SHAs, deterministic caching, reusable workflow permission boundaries, and local validation
Maintain NeMo Relay dynamic plugin loaders, manifests, Rust native SDKs, gRPC worker protocol, Python worker SDK, docs, tests, and release workflow coverage
Maintain or extend NeMo Relay observability surfaces across ATIF and typed OpenTelemetry projections
Maintain or extend the NeMo Relay adaptive surface across config, plugins, docs, and bindings; use this when users still say optimizer
Maintain NeMo Relay package metadata, module paths, generated artifacts, and release-facing build surfaces
Use this skill when NeMo Relay is installed or imported but application-side runtime behavior is missing or incorrect, including load failures, inactive scopes, missing events, and plugin or adaptive wiring problems.
Use this skill when first-time NeMo Relay users want to try Relay, choose the least-complex supported quick start, or verify initial value through the CLI, a maintained integration, or direct Python, Node.js, or Rust instrumentation before production setup.
Use this skill when choosing or running NeMo Relay installation for the CLI, Python, Node.js, Rust, OpenClaw, Hermes, or maintained framework integrations before runtime configuration or quick-start setup.
Use this skill when an application owns tool or LLM/provider call sites and needs to wrap them with NeMo Relay scopes and managed execution APIs for lifecycle events, middleware, or guardrails.
Use this skill when concurrent requests, async tasks, threads, workers, goroutines, or agents need independent NeMo Relay scope stacks and correct ancestry propagation.
Use this skill when adding NeMo Relay typed wrappers, domain types, or provider codecs while preserving JSON middleware semantics and caller-visible behavior.
Use this skill when migrating applications, examples, integrations, documentation, manifests, or repository code from NeMo Flow to NeMo Relay across Python, Rust, Node.js, Go, C FFI, CLI, configuration, and observability surfaces.
Use this skill when baseline NeMo Relay instrumentation exists and the user wants to configure or evaluate adaptive plugin behavior, including telemetry, state, adaptive_hints, tool_parallelism, ACG, hint consumption, or measured rollout.
Use this skill when building or packaging reusable NeMo Relay runtime behavior as a configuration-activated plugin with deterministic validation and rollback-safe registration.
Use this skill when choosing or configuring NeMo Relay 0.6 or 0.7 observability through the built-in plugin, subscribers, or exporters, including raw ATOF events, ATIF trajectories, OpenTelemetry, OpenInference, or custom event handling.
Prepare a NeMo Relay code freeze by creating the release branch, updating nightly alpha branch config, bumping main to the next version, and opening the required PR
Prepare, open, create, publish, update, or edit a NeMo Relay pull request or PR body with the right tests, docs, contributor hygiene, and repository pull request template
Perform a coordinated repository, package, crate, module, or symbol rename across NeMo Relay
Review documentation, examples, and docs-heavy changes for NVIDIA technical writing style, terminology, and repo accuracy
Make a small, reviewable NeMo Relay bug fix without widening scope unnecessarily
Build and test the NeMo Relay FFI surface; use this for crates/ffi changes, header generation, or ABI-facing validation
Build and test the NeMo Relay Go binding; use this for go/nemo_relay changes or Go-facing integration checks
Build and test the NeMo Relay Node.js binding; use this for crates/node changes or Node-facing integration checks
Build and test the NeMo Relay Python binding and worker plugin SDK; use for python/nemo_relay, python/plugin, or crates/python changes
Build and test NeMo Relay Rust core, adaptive, and dynamic plugin crates; use for crates/core, crates/adaptive, crates/plugin, crates/worker, crates/worker-proto, crates/types, or shared runtime changes
Update the NeMo Relay project version across Cargo, Node, and lockfiles without leaving release surfaces out of sync
Choose and run the right NeMo Relay validation matrix for a change instead of using one fixed test list
>- Use for NVIDIA-related requests where an NVIDIA skill might help, even if the user did not ask for a skill. Trigger on NVIDIA products, hardware, software, SDKs, GPUs, Jetson/JetPack/L4T/BSP/SDK Manager/driver/flashing/setup, CUDA, NIM, NeMo, Omniverse/OpenUSD/SimReady, RAPIDS/cuDF, cuPyNumeric, cuOpt, Dynamo, Holoscan, TensorRT, DeepStream, VSS, TAO, NGC/NVCF. Do not use for generic non-NVIDIA route, optimize, deploy, AI, video, data, or infrastructure tasks.
Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.
> Build and verify a TensorRT engine from a Hugging Face model ID or ONNX file, with numerical parity checked against ONNX Runtime. Use when the user imports a non-LLM model to TensorRT, needs a verified engine from ONNX, hits trtexec "unsupported operator", must verify the engine matches ONNX numerically, debugs a polygraphy parity failure (large max abs diff at FP16), or configures multi-input TensorRT, trtexec onnx, trtexec unsupported operator, optimum-cli export, polygraphy parity check, polygraphy run --trt --onnxrt, parity check failed, max abs diff, verify engine matches ONNX, --minShapes, dynamic shapes trtexec, multi-input shape profile, FP16 engine, INT64 `trt-cpp-runtime-quickstart` (C++ engine load). LLM token generation belongs in TensorRT-LLM, not here.
> Compile a PyTorch model to a TensorRT engine via Torch-TensorRT — AOT or JIT — under the new strong-typing default. Use when the user compiles PyTorch to TensorRT without ONNX, hits "enabled_precisions should not be used when use_explicit_typing=True", sees Dynamo graph breaks or PyTorch fallback, debugs ABI errors at import torch_tensorrt, or needs the compatible torch / torch_tensorrt / tensorrt-cu13 version pins for torch.compile backend torch_tensorrt, pytorch to tensorrt, ExportedProgram, Dynamo graph break, use_explicit_typing, enabled_precisions, torch_tensorrt.Input, min_block_size, truncate_double, tensorrt-cu13, `trt-cpp-runtime-quickstart`. LLM token generation belongs in TensorRT-LLM.
>- Load and run a TensorRT engine (.plan / .engine) from C++ using the TensorRT 11 / 10.x **modern Runtime API**, avoiding the deprecated TRT 8.x binding-index APIs that older guidance still promotes. Use whenever the user asks about loading or running a TensorRT .plan/.engine from C++, even on "minimal example" requests — without this skill the default reply uses deprecated enqueueV2-style code. Also use when the user hits "Engine plan file is generated on an incompatible device", deserializeCudaEngine returns nullptr, gets an enqueueV2 / IStreamReader deprecation warning, or inference, load TensorRT plan C++, run .plan from C++, IRuntime example, deserializeCudaEngine, enqueueV3, enqueueV2 deprecated, setTensorAddress, getBindingIndex, IStreamReaderV2, libnvinfer C++. NOT for building engines (`trt-onnx-quickstart`), Python deploy, plugins, multi-GPU.
>- Migrate a TensorRT build from weak typing (deprecated 10.12, removed 11.0) to strong typing — across Python INetworkDefinition builders, the trtexec CLI, and kSTRONGLY_TYPED, weak typing deprecated, kFP16/kINT8 removed, setPrecision rejected, setComputePrecision deprecated, do I still need --stronglyTyped, how to add the kSTRONGLY_TYPED flag, ModelOpt autocast, INT8 on TRT 11. NOT for ONNX import (`trt-onnx-quickstart`), Torch-TRT (`trt-torch-quickstart`), or C++ deploy (`trt-cpp-runtime-quickstart`).
Validate and analyze TensorRT performance data from paired layer-info JSON and profile/latency JSON files. Use when asked to inspect TensorRT, TRT, torch-tensorrt, or ONNX-TensorRT perf reports, verify that layer/profile JSON files are valid and from the same model, infer basic model information, find likely fusion or latency optimization opportunities, and produce a concise Markdown performance report or structured JSON data.
Use when the user provides Warp commit SHA(s) and GitHub issue number(s) to assess, draft issue comments, post progress updates, or recommend whether issue threads should stay open or close.
Generate a starting-point GitHub release notes draft for an upcoming Warp release (feature or bugfix, auto-detected from the version). Run from a release branch, or pass a branch / git ref as the argument.
Audit Warp CHANGELOG.md before a release: recover lost entries, sort by user impact, refine entry language, line-wrap, and (release-branch mode) bump compare refs.
Generate a Warp release audit report (pre-release or RC, auto-detected from version string and head ref).
> Guide users through Cosmos3 installation, environment setup, checkpoint downloading, and verification. Use when the user asks "how do I install cosmos3", "how do I set up the environment", "how do I download checkpoints", "how do I use Docker", or any question about getting the package running for the first time.
> Guide users through running Cosmos3 inference — offline batch generation, online serving with Ray and Gradio, parallelism options, input formats, sampling parameters, and prompt upsampling. Use when the user asks "how do I run inference", "how do I generate a video", "how do I serve the model", "what parameters should I use", or any question about running the model to produce outputs.