4090-probe - SKILL.md Agent Skill

name: 4090:probe description: > Run the W0 Substrate hardware probe on the 4090 node. Windows: deploy/provision/glances-autodetect.ps1 Linux: deploy/provision/glances-autodetect.sh Outputs structured JSON: gpu, cpu, nics, nic_collisions, unifi_topology, system specs. Pipe through json-to-profile.py to write pmoves/config/profiles/.yaml.

4090:probe — W0 Substrate Hardware Probe

Runs the W0 Substrate hardware probe to capture the 4090 node's system profile: GPU, CPU, NIC stats (including collision counters), Unifi network topology, and system memory. Output feeds into the node TAC tree and ghost detector pipeline.

Run (Windows)

# Requires glances running: http://localhost:61208
# Run as Administrator for full NIC collision data
powershell -ExecutionPolicy Bypass -File deploy/provision/glances-autodetect.ps1

Optional: write JSON output to file for piping to the profile writer:

powershell -ExecutionPolicy Bypass -File deploy/provision/glances-autodetect.ps1 `
    --json-file probe.json
python deploy/provision/json-to-profile.py --json probe.json --node-id pmoves-4090

Run (Linux)

bash deploy/provision/glances-autodetect.sh

Pipe directly to profile writer:

bash deploy/provision/glances-autodetect.sh --json-file /tmp/probe.json
python deploy/provision/json-to-profile.py --json /tmp/probe.json --node-id pmoves-4090

Profile Auto-Write

json-to-profile.py (W0-PR6, merged in PR #1486) converts the probe JSON into a canonical node profile YAML at pmoves/config/profiles/<node-id>.yaml.

USAGE:
  python deploy/provision/json-to-profile.py \
      --json probe.json \
      --node-id pmoves-4090 \
      [--out-dir pmoves/config/profiles/] [--dry-run] [--force]

The profile includes hardware tags, GPU list, Tailscale role, unifi_topology (null if Unifi probe was skipped), ghost adapter warnings, and compose overrides.

Expected Probe Output Shape (JSON)

{
  "hostname": "PMOVES-4090",
  "timestamp": "2026-05-18T15:00:00Z",
  "arch": "x86_64",
  "suggested_node_type": "gpu-4090",
  "suggestion_confidence": "high",
  "cpu": { "model": "Intel Core i9-13980HX", "cores_physical": 24, "cores_logical": 32 },
  "ram_gb": 64,
  "gpus": [{ "vendor": "NVIDIA", "model": "RTX 4090 Laptop GPU", "vram_gb": 16 }],
  "nic_collisions": [],
  "unifi_topology": null,
  "os": { "distro": "Windows", "version": "11 Pro" }
}

Ghost Detector Use

The nic_collisions field is the primary ghost detector signal. A rising collision counter on an idle interface indicates phantom traffic. Compare across two probe runs:

# Run probe twice, 60s apart — diff the nic_collisions values
powershell -File deploy/provision/glances-autodetect.ps1 --json-file probe1.json
sleep 60
powershell -File deploy/provision/glances-autodetect.ps1 --json-file probe2.json
# Compare nic_collisions values between probe1.json and probe2.json

Prerequisites

Glances running with JSON API: pip install glances[web] then glances -w
Windows: PowerShell 5.1+ with Get-NetAdapterStatistics available
Linux: glances + ip -s link
Profile writer: pip install pyyaml

Notes

W0-PR4 (ghost detector/Shift Crew) merged as PR #1591
W0-PR5 (Unifi probe) merged as PR #1588
W0-PR6 (json-to-profile.py) landed in PR #1486
See node-4090-sitrep for quick health check (doesn't require glances)
See pmoves/configs/tac_trees/node-4090-laptop.tac.yaml Phase 6 for probe TAC entries
TAC entry n4090.shift-crew.probe-wire status: done (updated 2026-05-24)