Jetson Diagnostic

A unified, agent-friendly view of a running Jetson device. Replaces the need to remember which of tegrastats, jtop, procrank, /sys/kernel/debug/nvmap, nvpmodel, free, swapon, df, and systemctl list-units produces which slice of the truth.

Purpose

Capture a read-only health snapshot from the Jetson host so agents can answer device identity, memory, GPU, thermal, power, storage, and service-state questions using live data instead of guesses.

When to use

Activate when the user asks:

• “What is this Jetson? What SKU? How much memory?”
• “What’s running on this Jetson right now?”
• “Why is my Jetson slow / hot / out of memory?”
• “Give me a snapshot of GPU / CPU / power usage.”
• “What does my tegrastats output mean?”
• “Which services are running that I could turn off?”
• The user has installed jetson-memory-audit, jetson-headless-mode, jetson-inference-mem-tune, jetson-llm-benchmark, jetson-llm-serve, or jetson-package and needs a baseline measurement before running them.

Do not use this skill to change power modes, drop caches, stop services, install packages, serve models, or tune inference flags. Report the observed state, then hand off to the action-oriented skill.

Prerequisites

• Run on the Jetson host, or in a sandbox/container with host-visible Jetson system paths and process data.

Available Scripts

If your agent runtime supports run_script, use it to run scripts/snapshot.sh or scripts/mem_summary.sh and summarize the returned output. Otherwise run the scripts with bash from the repository root.

Instructions

1. Run scripts/snapshot.sh for the all-in-one JSON view (preferred default).
2. For a quick human-readable memory line, run scripts/mem_summary.sh.
3. To explain a single tegrastats line the user has pasted, see references/tegrastats-fields.md.
4. To explain the NvMap clients output, see references/nvmap-clients.md.

Reporting guidance

Run the matching helper script before summarizing device state, and report only fields returned by that script. If direct execution is blocked by the runtime, run it with bash {baseDir}/scripts/<script-name> rather than trying to chmod files.

• For “what is this Jetson” questions, quote product_model or sku, variant, l4t_version, and mem_total_gb.
• For “slow and hot” questions, run snapshot.sh and summarize both sides of the symptom: thermal_c for heat, plus top_processes, gpu_processes, nvmap.top_clients, or gpu_source for load. End with a concrete handoff such as jetson-memory-audit, jetson-headless-mode, or jetson-inference-mem-tune.
• For “which process is using memory” questions, run snapshot.sh and name the leading process as pid <number>, cmd, and its pss_kb / MiB value. If NvMap GPU memory is the relevant signal, also quote gpu_source and the top nvmap.top_clients or gpu_processes entry.

If your agent runtime does not automatically execute helper scripts relative to this skill directory, resolve script paths with the AgentSkills {baseDir} placeholder:

{baseDir}/scripts/snapshot.sh
{baseDir}/scripts/mem_summary.sh

Do not call jetson-diagnostic as a tool name unless the runtime explicitly registers skills as callable tools; Agent Skills are normally instructions plus files, not direct tool functions.

All scripts source the canonical platform detector at skills/jetson-diagnostic/scripts/detect_jetson.sh (exports JETSON_SKU, JETSON_GENERATION, JETSON_PRODUCT_LINE, JETSON_VARIANT, JETSON_MEM_GB, JETSON_L4T_VERSION, JETSON_PRODUCT_MODEL). Other skills may source this detector rather than duplicating Jetson identification logic. Exits 2 with a remediation message off-platform.

Limitations

• Seeing this skill file does not guarantee access to Jetson host hardware. If /proc/device-tree/model, /etc/nv_tegra_release, tegrastats, nvpmodel, nvidia-smi, or /sys/kernel/debug/nvmap are missing inside a NemoClaw/OpenClaw sandbox, say the sandbox lacks Jetson host visibility and ask the user to run on the Jetson host or relaunch with a host-visible sandbox profile.
• NvMap debugfs often requires root, so unprivileged runs may report gpu_source: "none" or incomplete nvmap fields.
• This skill reports observed state only. Do not fabricate memory, GPU, thermal, service, or reclamation data when a tool is missing or inaccessible.

Error handling

• If a helper exits off-platform, report that the current environment is not a Jetson host or lacks host visibility; do not substitute generic Linux values.
• If tegrastats, nvpmodel, nvidia-smi, or NvMap debugfs are unavailable, preserve the corresponding null, false, or empty fields from the JSON and explain which signal is limited.
• If snapshot.sh emits malformed JSON, report the raw failure and rerun after fixing the helper output; do not hand-edit a synthetic device snapshot.

Output contract for snapshot.sh

{
  "sku": "orin-nano",
  "generation": "orin",
  "product_line": "orin-nano",
  "variant": "orin-nano-8gb",
  "mem_total_gb": 8,
  "l4t_version": "36.4.0",
  "product_model": "nvidia jetson orin nano developer kit",
  "memory_kb": { "total": 8123456, "available": 4123456, "swap_total": 0, "swap_free": 0, "cached": 1234567 },
  "tegrastats_sample": "RAM 4011/8138MB (lfb 8x4MB) ...",
  "thermal_c": { "CPU": 52.3, "GPU": 49.0, "AO": 47.0 },
  "power": { "nvpmodel_id": 0, "nvpmodel_name": "MAXN" },
  "disk": [ { "mount": "/", "used_pct": 41 } ],
  "gpu_source": "nvmap:iovmm-clients",
  "gpu_devices": [],
  "gpu_processes": [],
  "nvmap": {
    "readable": true,
    "total_kb": 654321,
    "stats_total_bytes": 669985280,
    "top_clients": [ { "pid": 1234, "cmd": "vlm-server", "kb": 524288 } ]
  },
  "top_processes": [ { "pid": 4321, "cmd": "vllm", "pss_kb": 4000000 } ],
  "candidate_services": { "gdm3": { "active": "inactive", "enabled": "disabled" } }
}

gpu_source names the specific datum the skill used to attribute per-process GPU memory, so the caller can tell exactly what the numbers represent:

• "nvidia-smi:compute-apps" — per-process used_memory values from nvidia-smi --query-compute-apps. Used on the unified nvidia.ko stack (Thor family today). Note: on this stack nvidia-smi’s device-level memory.used query returns [N/A] on some BSPs, which is why the skill sums the per-process list rather than reading a top-level total. The summed total appears in gpu_processes[*].used_mib.
• "nvmap:iovmm-clients" — per-process sizes from /sys/kernel/debug/nvmap/iovmm/clients. Used on the nvgpu stack (Orin family today), where nvidia-smi is a stub that returns [N/A] for every compute/memory query. Per-process entries appear in nvmap.top_clients; the kernel-side total is in nvmap.total_kb and (when readable) nvmap.stats_total_bytes.
• "none" — no authoritative source reachable. Typical when running unprivileged on an nvgpu-stack Jetson (debugfs under /sys/kernel/debug/nvmap needs sudo); rerun with sudo to populate the nvmap fields.

The agent should present the salient parts back to the user (SKU, available memory, top GPU consumer per gpu_source, hottest zone, power mode) and offer to drill into specifics (top_processes, gpu_processes / nvmap, services).

Safety

This skill is read-only. It does not change nvpmodel, does not run jetson_clocks, does not modify services. To act on findings, hand off to:

• jetson-memory-audit — focused memory snapshot + drop_caches verify loop
• jetson-headless-mode — disable GUI + auxiliary daemons (safe, reversible)
• jetson-inference-mem-tune — pick runtime + memory flags (vLLM / SGLang / llama.cpp / TensorRT Edge-LLM)
• jetson-llm-serve — vLLM and related GHCR images with Jetson defaults
• jetson-llm-benchmark — reproducible latency / throughput benchmarks
• jetson-package — GHCR + Jetson AI Lab PyPI indexes vs generic ARM wheels

Cross-platform behavior

* On Jetsons whose GPU is driven by the in-tree nvgpu kernel driver, the nvidia-smi binary is present but most fields (Memory-Usage, power, utilisation, compute-process table) report Not Supported / N/A. To decide which source to trust at runtime, the script does a capability probe — nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits — and only uses nvidia-smi for per-process GPU memory when that query returns a real integer. When it doesn’t, the script falls back to /sys/kernel/debug/nvmap/iovmm/clients, which on nvgpu-stack Jetsons is the authoritative per-process GPU-memory source.

The script handles each tool’s presence gracefully and reports null / false for tools it cannot reach (typical when the agent isn’t running with the privilege needed for /sys/kernel/debug). Variant detection uses the /proc/device-tree/model string first (recognising names like T5000 / T4000) and falls back to memory-size heuristics when the model string is generic.