troubleshoot-zymtrace-profiler

What problem does it solve?

This Skill troubleshoots the zymtrace profiler agent when it misbehaves, preventing reliable CPU/GPU profile collection.

Core Features & Use Cases

• Symptom-to-root-cause routing: Guides diagnosis for CrashLoopBackOff, ImagePullBackOff, OOMKilled/restart cycles, missing NVML/GPU metrics, broken PC sampling, and profiler-side license/auth failures.
• Agent-focused verification: Checks DaemonSet health, pod readiness, recent agent logs, and workload interception signals, while explicitly handing off to backend troubleshooting when the issue is downstream.
• Kubernetes-first workflows: Uses kubectl and helm workflows to triage GPU vs CPU expectations, container configuration mismatches, and common cluster/runtime constraints.
• Use Case: When the UI shows no GPU traces but the CPU path appears partially healthy, it walks the agent-side implant and GPU metrics prerequisites (cudaProfiler enabled, NVML discoverability, and PC sampling conditions) to determine whether the failure is agent-side or workload/backend-side.

Quick Start

Tell the Skill what you see (for example, “profiler pods CrashLoopBackOff” or “CPU profiles arrive but no GPU traces”) and confirm the profiler namespace and Helm release to get a targeted diagnosis and fix path.