OverviewHistoryStatsSecurity
npx skills add ...
Documentation
SKILL.md
npx skills add nvidia/fleet-intelligence-client --skill nvfleetint
Query NVIDIA Fleet Intelligence with the nvfleetint CLI. Use for ad hoc questions about fleets, nodes, GPUs, node groups, compute zones, alerts, agent health, firmware, verification, inventory, errors, or authentication. For a fleet-wide HTML snapshot use fleet-health-report; for a single-node RCA/RCCA use node-rca-rcca.
npx skills add nvidia/fleet-intelligence-client --skill nvfleetint
Use live nvfleetint JSON; never answer fleet-state questions from memory. Read references/cli-contract.md before querying. Read references/auth.md only for setup/profile work.
--output json; give the user prose, not command dumps.--page-size 1 and read .total. For identities, add --view basic. Use --all only when every row is required.| Need | Command |
|---|---|
| Fleet totals/health/metrics | overview |
| Zones or node groups | computezone list, nodegroup list |
| Nodes/current detail | node list, node describe <uuid> |
| Node health history | node health <uuid> |
| Fleet alert records | alert list |
| Alert impact/investigation | alert summary, alert node, alert describe, alert options |
| Raw events/histogram | event list, event buckets |
| Customer tags | tag list |
| Inventory/error reports | report inventory, report error |
| Verify signed report | report verify |
| Filter/sort values a command accepts | node options, nodegroup options, alert options, xidburst options |
overview is one non-paginated object. Use lists for rows behind its counts.
Ask users for human-readable zone/group names, never IDs. Name filters are comma-separated partial matches. For exact scope, resolve with computezone list --view basic or nodegroup list --view basic, clarify ambiguous names with recognizable detail metadata, then use IDs internally. Accept an ID already supplied by the user.
Detailed node list and describe query both agent views by default and return
{inband: ..., oob: ...} in JSON. Use --agent-type inband|oob when only one
view is needed; node list requires --agent-type whenever any filter or
--sort-by/--order flag is set, since in-band and OOB nodes carry different
fields. OOB list supports --bmc-hostname; OOB describe includes full
inventory JSON and supports table sections managers, systems, chassis,
and firmware. Node basic rejects health, agent, verification, and firmware
filters and supports sorting by hostname, nodeUUID, or bmcHostname.
node health requires both absolute boundaries. It does not support --window.
Filter values:
| Flag | Values |
|---|---|
--health | Healthy, Degraded, Unhealthy, Unknown |
--agent-status | Online, Offline, Unknown |
--verification-check | Verified, Unverified, Degraded, Pending, Unsupported, Unknown |
--firmware-check | Passed, Failed, Unknown |
Node sort keys are hostname, nodeUUID, healthStatus, nodegroup, computezone, gpuType, gpuCount, verificationCheck, agentStatus, agentVersion, kernelVersion, gpuDriverVersion, gpuFirmwareVersions, and bmcHostname. Node-group sort keys are health and nodes.
Alert severity is Critical/Warning; state is Detected/Triggered/Resolved. alert summary, alert node, and alert options default to the active view; use --view historical for history. Summary returns impacted nodes plus fleet-wide alert aggregates. Node returns alerts for one node. Describe returns one alert's event history. In node-alert results, Critical/Warning values are active severity; Detected/Resolved are inactive audit values. Don't count every non-Resolved row as active.
Events require --window or both --start/--end. Durations use Go units through hours—no d. Event list paginates; buckets do not.
Use at most one tag scope flag: --node, --nodegroup, or --computezone; tag list has no pagination. tag set <node-uuid> --tags <list> is the one command here that writes: it replaces a node's tags rather than adding to them, so run it only when the user asks for that change by name, pass --yes, and never use it to gather data. Report-error list requires --group-by error|node; only list supports --all and CSV. Signed inventory requires CSV. Report filters accept compute zone IDs, node group IDs, and tags; error reports also accept --errors, --severities, and graph-only --step.
For “Are any H100 nodes having problems?”, query H100 nodes filtered to Degraded/Unhealthy/Unknown. Include Unknown, read the filtered total, list only returned problem nodes, and do not claim the remaining H100 fleet is healthy without a separate query. Drill into requested UUIDs with node describe and node-scoped alerts.