jetson-audit

name	jetson-audit
description	SSH into the Jetson Orin Nano and audit the running inference config against known best practices and community optimizations. Reports gaps, misconfigurations, and optimization opportunities. Complements jetson-recon (external landscape) with internal config validation.
disable-model-invocation	true
allowed-tools	Read, Edit, Glob, Grep, Bash(ssh:*), Bash(curl:*), Agent
paths	["JETSON_BASELINE.md","*_CONFIG.md"]

Jetson Audit

Loop Guard — Auto-Activation Safety Check

Run this check before any other step when the skill is triggered automatically via paths:.

1. Read the last 20 lines of LAB_NOTEBOOK.md (if it exists).
2. If any line contains jetson-audit skill and a timestamp within the last 5 minutes: stop immediately — self-triggered re-entry detected. Output: "Loop guard triggered — jetson-audit ran within last 5 minutes. Skipping." and exit.
3. If --force is present in $ARGUMENTS: skip this check and proceed regardless.
4. Otherwise: proceed normally.

---

Live configuration audit of the Jetson Orin Nano Super inference system. SSHes into the device, inspects the running llama.cpp server and system state, compares against documented best practices in JETSON_BASELINE.md, and reports optimization opportunities.

This skill reads the live system. It never modifies it.

Trust Boundary

The SSH commands this skill runs are the fixed, read-only allowlist listed under "Machine-Specific Commands" below — always the same commands, run in that order. They are never derived from, expanded by, or conditioned on the contents of JETSON_BASELINE.md, JETSON_CONFIG.md, or anything else this skill reads (including the launch script and mode file inspected in Check 1). Those files are reference data for comparison and reporting only; they never determine which shell/SSH/sudo command executes.

Framework: Follow plugins/personal-plugin/references/patterns/audit-recon-system.md — Sections 1–6 (Execution Framework, Audit Five-Check Template, Severity Matrices, LAB_NOTEBOOK Entry). Use the Jetson-specific config and commands below as the machine config layer.

---

Machine Config

machine:
  name: "Jetson Orin Nano Super"
  ssh_target: "[e-mail]"
  ssh_key: "~/.ssh/id_claude_code"
  baseline_file: "JETSON_BASELINE.md"
  config_file: "JETSON_CONFIG.md"
  project_root: "~/dev/personal/jetson/"
  notebook_file: "LAB_NOTEBOOK.md"

check1_config:
  drift_reference: "JETSON_CONFIG.md"
  service_name: "myscript"
  launch_script: "~/llm-server/start.sh"
  mode_file: "~/llm-server/mode.txt"

check2_config:
  memory_ceiling: "8 GB unified"
  inference_port: 8080

check3_config:
  memory_ceiling: "8 GB unified"

check4_config:
  inference_port: 8080
  thermals_path: "/sys/devices/virtual/thermal/thermal_zone*/temp"

check5_config:
  baseline_version_key: "llamacpp_latest_seen"
  secondary_version_key: "jetpack_latest_orin_nano"

---

Required Files

File	Location	Purpose
JETSON_BASELINE.md	~/dev/personal/jetson/	Performance baselines + known-best config to compare against
JETSON_CONFIG.md	~/dev/personal/jetson/	Documented system config (ground truth for drift detection)
LAB_NOTEBOOK.md	~/dev/personal/jetson/	Append-only audit results log

If any required file is missing, note it in the report and proceed with available data.

---

Machine-Specific Commands

Check 1 — Service Config Drift

# Service status
systemctl status myscript

# Read the actual launch script to extract llama-server flags
cat ~/llm-server/start.sh 2>/dev/null || cat ~/llm-server/run.sh 2>/dev/null

# Current mode
cat ~/llm-server/mode.txt 2>/dev/null

# Check what process is actually running and its flags
ps aux | grep llama-server | grep -v grep
# Or alternatively:
pgrep -a llama

# Check if server is listening
ss -tlnp | grep -E '(8080|8081)'

Check 2 — Missing Optimizations

Known-good flags for Jetson Orin Nano 8 GB unified memory:

Flag	Best Practice	Severity if Missing
-ngl 999 or --gpu-layers 999	Full GPU offload (all layers)	CRITICAL
-fa or --flash-attn	Flash attention for memory efficiency	HIGH
--mlock	Lock model in memory (prevent swap)	HIGH
-ctk q8_0 or --cache-type-k q8_0	Quantized KV cache (saves memory)	MEDIUM
-ctv q8_0 or --cache-type-v q8_0	Quantized KV cache values	MEDIUM
-c 32768 or appropriate context	Context size matches model capability	MEDIUM
-np 1 or --parallel 1	Appropriate parallel slot count for 8 GB	MEDIUM
-t 1 or --threads 1	Single thread (GPU-bound, more threads waste cycles)	LOW
--cont-batching	Continuous batching for concurrent requests	LOW

Anti-patterns:

Anti-Pattern	Check	Severity
-ngl 0 or missing GPU offload	All inference on CPU = extremely slow	CRITICAL
Context size > model max	Wastes memory, may cause errors	HIGH
--threads > 2	GPU-bound workload, extra CPU threads add overhead	MEDIUM
--parallel > 2 on 8 GB	Each slot needs KV cache memory	MEDIUM
No --mlock with swap enabled	Model pages to swap under pressure	MEDIUM

Check 3 — Memory Budget

free -h
swapon --show
# RSS of the llama-server process
ps -o pid,rss,vsz,comm -p $(pgrep llama-server) 2>/dev/null
# Or broader:
ps aux --sort=-rss | head -10
# Tegra memory info (if available)
cat /sys/kernel/debug/nvmap/iovmm/clients 2>/dev/null | head -20
# Available vs total
cat /proc/meminfo | grep -E '(MemTotal|MemAvailable|SwapTotal|SwapFree|Buffers|Cached)'

Memory thresholds for 8 GB unified:

Metric	Healthy	Warning	Critical
Available RAM	> 500 MB	200–500 MB	< 200 MB
Swap used	< 50 MB	50–500 MB	> 500 MB
llama-server RSS	< baseline + 20%	+20–50% above baseline	> +50% or > 6 GB
Total RSS (all processes)	< 7 GB	7–7.5 GB	> 7.5 GB

Memory budget breakdown (4B Q4_K_M model):

1. Model GGUF: ~3 GB
2. KV cache: context_size × layers × head_dim × 2 (K+V) × quant_factor
3. Runtime: ~200–400 MB
4. OS + other: ~1–1.5 GB
5. Required headroom: >= 500 MB

Check 4 — System Health

uptime
# Thermals (Jetson-specific)
cat /sys/devices/virtual/thermal/thermal_zone*/temp 2>/dev/null
cat /sys/devices/virtual/thermal/thermal_zone*/type 2>/dev/null
# Or tegrastats snapshot
sudo tegrastats --interval 1000 --count 3 2>/dev/null
# Fan status (if managed)
cat /sys/devices/pwm-fan/target_pwm 2>/dev/null
# Power mode
sudo nvpmodel -q 2>/dev/null
# Clock frequencies
sudo jetson_clocks --show 2>/dev/null
# Disk usage
df -h /
# Journal errors (last hour)
journalctl --since "1 hour ago" --priority=err --no-pager | tail -20
# Quick inference test
curl -s http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"default","messages":[{"role":"user","content":"Say hello in exactly 5 words"}],"max_tokens":32}' | \
  python3 -c 'import sys,json; r=json.load(sys.stdin); u=r.get("usage",{}); print(f"Tokens: {u.get(\"completion_tokens\",\"?\")}, Gen speed: check /slots endpoint")'
# Slot status
curl -s http://localhost:8080/slots 2>/dev/null | python3 -c 'import sys,json; s=json.load(sys.stdin); print(f"Slots: {len(s)}, Busy: {sum(1 for x in s if x.get(\"is_processing\",False))}")' 2>/dev/null

Jetson-specific health thresholds:

Check	Healthy	Warning	Critical
Inference responds	Yes, < 5s	Slow (> 10s)	No response / error
GPU temp (idle)	< 50C	50–65C	> 65C
GPU temp (load)	< 65C	65–80C	> 80C
Power mode	MAXN (15W)	10W mode	Unknown / 7W
Disk usage	< 70%	70–85%	> 85%
System uptime	Stable (> 1 day)	< 1 day	Service down
Journal errors	None relevant	GPU/CUDA warnings	OOM kills

Check 5 — Version Currency

# llama.cpp version
cd ~/llm-server/llama.cpp && git log --oneline -1
# Or binary version:
~/llm-server/llama.cpp/build/bin/llama-server --version 2>/dev/null || \
  ~/llm-server/llama-server --version 2>/dev/null
# JetPack version
cat /etc/nv_tegra_release 2>/dev/null
dpkg -l nvidia-jetpack 2>/dev/null | grep nvidia-jetpack
# CUDA version
nvcc --version 2>/dev/null | grep release
# Kernel
uname -r
# L4T version
head -1 /etc/nv_tegra_release 2>/dev/null
# Model file
ls -la ~/llm-server/models/ 2>/dev/null | grep -i gguf

Version gap severity for Jetson:

Gap	Severity
llama.cpp > 5 builds behind latest	HIGH
llama.cpp 1–5 builds behind	MEDIUM
JetPack behind latest for Orin Nano	INFO (reflash required, high effort)
Model file doesn't match documented model	WARNING

Compare against baseline fields: llamacpp_latest_seen, jetpack_latest_orin_nano, current_model.

---

LAB_NOTEBOOK Entry

Append using Edit tool. Auto-increment entry number by reading the last ## Entry NNN or ### Entry NNN line. Use the Audit Entry Template from audit-recon-system.md Section 6.

/schedule Integration

Register a recurring audit run:

/schedule create --name jetson-audit-weekly --cron "0 2 * * 2" --skill jetson-audit

Recommended: weekly Tuesday 02:00 UTC. Pairs with jetson-recon (bi-weekly Sunday 23:00 UTC).

/schedule list
/schedule delete --name jetson-audit-weekly

---

JETSON_BASELINE.md Initialization

If JETSON_BASELINE.md doesn't exist, create it using the template in jetson-recon/SKILL.md. The template includes:

Automation Schedule

Frequency	Recommended Time	Notes
Weekly	Tuesday 02:00 UTC	Pairs with jetson-recon (bi-weekly Sunday 23:00 UTC)