runai-diagnose-failed-workload — Claude Code skill
runai-diagnose-failed-workload diagnoses crashed, stopped, OOMKilled, and startup-failed Run:ai workloads using structured evidence and read-only cluster queries.
no license ★ 4,106 updated 2026-10-23
npx skills add NVIDIA/skills --skill runai-diagnose-failed-workload -a claude-codeWhat runai-diagnose-failed-workload does
The skill investigates terminated or broken Run:ai jobs, focusing on issues like CrashLoopBackOff, container startup failures, volume mount issues, and ImagePullBackOff. It gathers context from workload specifications, metrics, and pod events. The workflow remains strictly read-only: it reports findings and suggests remedies without editing, stopping, or restarting resources.
Inspection operates on two tiers. Tier 1 requires an MCP setup with read access to the Run:ai service account for specifications, events, and metrics. Tier 2 is optional, relying on read-only Kubernetes credentials to pull prior container logs, node descriptions, and namespace events via kubectl.
Every run requires an explicit cluster_id. You must provide either a direct Run:ai UUID or both the workload name and project name. Normal completions, suspended jobs, and scheduling preemption fall outside this skill's scope.
What it helps with
- categorize terminal failures into root causes like app crashes or OOM terminations
- inspect failed pod startups caused by invalid images or init scripts
- trace storage issues from persistent volume claim mounts
- separate worker node outages from application errors
Example prompts for the agent
- Diagnose why workload
w-98765failed on clusterai-prod-1 - Check the CrashLoopBackOff cause for
finetune-llmin projectresearchon clusterk8s-gpu - Find why job
eval-batchstopped in projecttestingon clusterdev-clusterover the last48h
How to install runai-diagnose-failed-workload
1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill runai-diagnose-failed-workload -a claude-code2Manually: put the files in place
~/.claude/skills/runai-diagnose-failed-workload/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/runai-diagnose-failed-workload/
1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill runai-diagnose-failed-workload -a codex2Manually: put the files in place
.agents/skills/runai-diagnose-failed-workload/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill runai-diagnose-failed-workload -a cursor2Manually: put the files in place
.cursor/skills/runai-diagnose-failed-workload/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install NVIDIA/skills skills/runai-diagnose-failed-workload2Manually: put the files in place
.github/skills/runai-diagnose-failed-workload/SKILL.md1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill runai-diagnose-failed-workload -a gemini-cli2Platform CLI
gemini skills install https://github.com/NVIDIA/skills --path skills/runai-diagnose-failed-workloadManually: put the files in place
.devin/skills/runai-diagnose-failed-workload/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about runai-diagnose-failed-workload
How to install runai-diagnose-failed-workload in Claude Code?
Run npx skills add NVIDIA/skills --skill runai-diagnose-failed-workload -a claude-code. The definition will be placed in ~/.claude/skills/runai-diagnose-failed-workload/SKILL.md.
What does runai-diagnose-failed-workload do?
It evaluates logs, pod states, and Run:ai metrics to classify why a specific job terminated or failed to start, returning a quantified verdict without modifying the cluster.
What credentials does this skill require?
It needs Tier 1 MCP read access to Run:ai workloads, specs, and metrics. Optional Tier 2 access uses read-only Kubernetes credentials for deeper container log inspection.