dgx-spark-cli

Agent and CLI for operating an NVIDIA DGX Spark (Grace-Blackwell) workstation — device setup, health/monitoring, and local AI/ML workload management.

What you get

• An agent-first CLI cited from teken (afi-cli) — the runtime package has no third-party dependencies.
• A mesh identity — culture.yaml (suffix + backend) and the matching prompt file (CLAUDE.md for backend: claude).
• The canonical guildmaster skill kit (11 skills) under .claude/skills/, vendored cite-don't-import. See docs/skill-sources.md.
• A build + deploy baseline — pytest, lint, the agent-first rubric gate, and PyPI Trusted Publishing wired into GitHub Actions.

Quickstart

uv sync
uv run pytest -n auto                 # run the test suite
uv run dgx-spark-cli whoami  # identity from culture.yaml
uv run dgx-spark-cli learn   # self-teaching prompt (add --json)
uv run teken cli doctor . --strict    # the agent-first rubric gate CI runs

CLI

Machine scope (DGX Spark host telemetry)

The Spark is the system, so these read-only verbs sit at the top level:

They have zero runtime dependencies — kernel telemetry is read from /proc and /sys, while nvidia-smi, docker, and ip are shelled out and degrade gracefully (a missing tool reports available: false and still exits 0). doctor remains the health gate. Because the GB10 has no discrete VRAM, nvidia-smi reports aggregate GPU memory as [N/A]; gpu instead sums per-process compute-app memory so you can see how much of the shared pool the GPU holds.

Every command supports --json. Results go to stdout, errors/diagnostics to stderr (never mixed). Exit codes: 0 success, 1 user error, 2 environment error, 3+ reserved.

Swap management

Swap inspection is read-only; the grow command is a dry-run by default — re-run with --apply to mutate. All commands support --json.

Per-process history collection via systemd timer

Per-process memory and swap history accrues only when spark swap sample runs on a schedule. Here's a systemd service and timer pair:

# /etc/systemd/system/spark-swap-sample.service
[Unit]
Description=Sample per-process swap/memory for dgx-spark-cli
[Service]
Type=oneshot
ExecStart=/usr/bin/env spark swap sample

# /etc/systemd/system/spark-swap-sample.timer
[Unit]
Description=Periodic spark swap sample
[Timer]
OnBootSec=2min
OnUnitActiveSec=10min
Persistent=true
[Install]
WantedBy=timers.target

Enable with sudo systemctl enable --now spark-swap-sample.timer. Note that system-level CPU/mem/swap trend already comes from sysstat/sar; this timer only adds the per-process layer.

Monitoring (monitor) — AI-free webhook watchdog

monitor turns the collectors into a deterministic, always-on watchdog. It evaluates the same numbers against configurable thresholds and POSTs to a generic webhook when a catastrophe condition crosses — and again when it clears (edge-triggered, so a standing condition doesn't spam). No AI, no new dependencies (urllib does the POST).

dgx-spark-cli monitor config --init     # scaffold ~/.config/dgx-spark/monitor.json
export DGX_SPARK_WEBHOOK_URL=https://…  # or put webhook_url in the config
dgx-spark-cli monitor check             # dry run: what's firing right now
dgx-spark-cli monitor test              # POST a synthetic alert
dgx-spark-cli monitor install           # write the systemd --user unit
dgx-spark-cli monitor enable            # start it always-on (+ linger)
dgx-spark-cli monitor status            # service + currently firing alerts

Watches memory %, swap %, disk %, hottest sensor, GPU temp, load-per-core, I/O contention (iowait % + blocked processes), container health, and subsystem availability. Thresholds live in the config (null disables a check); webhook_format is generic (default), slack, or discord.

When monitor run starts (the systemd ExecStart), it POSTs a one-shot "started watching" liveness alert — so a watchdog that silently fails to come up is noticed by the absence of its heartbeat, not just the absence of an alert. A failed startup POST is logged, never fatal. Set notify_on_start: false in the config to disable it.

Make it your own

1. Rename the package spark/ and the dgx-spark-cli CLI/dist name throughout pyproject.toml, the package, tests/, and sonar-project.properties.
2. Edit culture.yaml with your suffix and backend.
3. Rewrite CLAUDE.md for your agent and run /init.
4. Re-vendor only the skills you need from guildmaster (see docs/skill-sources.md).

See CLAUDE.md for the full conventions (version-bump-every-PR, the cicd PR lane, deploy setup).

License

MIT — see LICENSE.