Jetson Health Snapshot
Read-only diagnostics that consolidate a live Jetson’s hardware, resource, and service state.
The skill clearly limits itself to read-only operation, forbids power-mode changes, cache dropping, service stopping, and package installation, and requires reporting access limitations. The scripts primarily read local telemetry, with no observed exfiltration, credential theft, or destructive default. Five points are deducted because process, service, and kernel telemetry may contain sensitive runtime information without documented redaction or data-flow handling; least-privilege sudo guidance, user confirmation, and recovery procedures are also incomplete. Attribution and Apache-2.0 licensing are reasonably clear.
The documentation specifies script entry points, arguments, platform detection, null preservation, and failure handling, and explains cross-platform GPU data sources. Fourteen points are deducted because the supplied evidence omits the critical snapshot.sh implementation, preventing static verification of the default path; no committed test suite covers key paths; and mem_summary.sh uses a GPU-name heuristic for nvidia-smi availability that does not match the documented capability-probe description. Static calibration limits this score to 10 or below.
The intended audience, Jetson scenarios, trigger phrases, non-fit mutation tasks, host/container visibility requirements, and hardware limitations are clearly described, with explicit refusal to substitute generic Linux data off-platform. Four points are deducted because Chinese-language interaction and mainland-China environment fit are not addressed, and the skill depends on Jetson host access, debugfs, sudo, and local utilities.
The material includes purpose, triggers, prerequisites, script documentation, reporting guidance, limitations, error handling, an output contract, safety boundaries, and cross-platform notes. Version, author, tags, and Apache-2.0 metadata are present, and BENCHMARK records validation findings. Six points are deducted for the missing recommended Examples section, incomplete author format, lack of explicit maintenance ownership, changelog, and update path, plus implicit runtime assumptions in dependency handling.
The skill covers identity, memory, GPU, thermal, power, storage, processes, and services, and maps requests to either a JSON snapshot or compact memory summary. The BENCHMARK reports correctness and effectiveness results across eight tasks. Nine points are deducted because the evaluation dataset is unavailable, snapshot.sh is absent from the supplied evidence, and static review cannot confirm representative outputs remain directly usable across real environments. Static calibration limits this score to 7 or below.
The source code, output contract, platform-detection logic, official documentation references, and BENCHMARK results provide auditable material and distinguish observed state from inference. One point is deducted because the BENCHMARK lacks raw data and reproduction steps, while the key snapshot script is not included, preventing independent reproduction of the main path. Static calibration limits this score to 5 or below.
- Before publication or deployment, include and review snapshot.sh and its tests, and verify that GPU capability probing matches the documented behavior.
- Process, service, and debugfs output may expose runtime-environment information; document redaction, minimal sudo scope, and any required user confirmation.
- The supplied material lacks Chinese output guidance and mainland-China environment notes; validate AgentSkills path resolution and local-tool availability before cross-environment use.
What does this skill do, and when should you use it?
Jetson Diagnostic produces a unified health snapshot for a running NVIDIA Jetson host. It reads device identity, memory, GPU, thermal, power, disk, process, and candidate service state, returning JSON or a compact human-readable summary. The skill observes state only and does not change power modes, services, caches, or inference settings. It requires a Jetson host or a sandbox/container with host-visible system paths and process data.
Runs scripts/snapshot.sh to emit JSON covering identity, memory, GPU, thermal, power, disk, top processes, and candidate services; --human, --tegra-secs N, and --top-procs N adjust the output. Runs scripts/mem_summary.sh for a compact RAM/GPU/swap summary and scripts/detect_jetson.sh for canonical Jetson SKU, generation, and product-line fields. It selects nvidia-smi or NvMap data for per-process GPU-memory attribution when available, while preserving null, false, or empty fields for inaccessible signals.
- A Jetson developer needs the device SKU, variant, L4T version, and total memory before troubleshooting or benchmarking.
- An operator investigates a slow, hot, or out-of-memory Jetson by correlating thermal readings with top CPU or GPU-memory consumers.
- A user wants a current snapshot of GPU, CPU, memory, storage, or power-mode state.
- A user has pasted a tegrastats line and needs its fields explained.
- A baseline is needed before running Jetson memory auditing, headless-mode, inference-tuning, LLM-serving, or benchmarking workflows.
What are this skill's strengths and limitations?
- Combines identity, memory, GPU, thermal, power, disk, process, and service-state signals.
- Read-only behavior is explicit: it does not change nvpmodel, stop services, install packages, or tune inference flags.
- Recognizes Thor, AGX Orin, Orin NX, and Orin Nano variants and includes GPU-source fallback logic.
- Handles missing tools and insufficient privileges with explicit null, false, or empty fields.
- Requires Jetson host visibility; a normal sandbox may not expose the required system paths or process data.
- NvMap debugfs commonly requires root, so unprivileged runs may lack process-level GPU-memory data.
- It reports observations and hands off remediation rather than fixing performance, thermal, or memory problems.
- The supplied material does not describe a test suite or test coverage for this skill.
How do you install this skill?
Install the individual skill with the CLI command documented in the repository README:
npx skills add nvidia/skills --skill jetson-diagnostic --yes
The README does not specify a fixed per-client installation directory; the CLI selects or prompts for the destination.
How do you use this skill?
After the skill is loaded, ask the agent for a concrete snapshot, for example: "What is this Jetson, and show its current GPU, thermal, memory, and power state?" Prefer scripts/snapshot.sh for the all-in-one view, or scripts/mem_summary.sh for a short memory summary. If the runtime supports run_script, use it; otherwise run the scripts with bash from the repository root. If host access is unavailable, rerun on the Jetson host or with a host-visible sandbox profile.
How does this skill compare with similar options?
The skill consolidates the different slices produced by tegrastats, jtop, procrank, NvMap debugfs, nvpmodel, free, swapon, df, and systemctl list-units into one view; those tools remain the underlying data sources or references.