From 00a7578d3735f4db3813838aef8c464991c05d95 Mon Sep 17 00:00:00 2001
From: ThinkOffApp <233579950+ThinkOffApp@users.noreply.github.com>
Date: Mon, 2 Nov 2026 21:19:03 +0200
Subject: [PATCH] user-intent-kit: publish the GPU's own load, temperature,
 clock and power from nvidia-smi

---
 .../user-intent-kit/src/host-telemetry.js     | 39 ++++++++++++++++++-
 .../test/host-telemetry.test.js               | 39 +++++++++++++++++++
 2 files changed, 77 insertions(+), 1 deletion(-)

diff --git a/packages/user-intent-kit/src/host-telemetry.js b/packages/user-intent-kit/src/host-telemetry.js
index 1a741d8..0a9f36a 100644
--- a/packages/user-intent-kit/src/host-telemetry.js
+++ b/packages/user-intent-kit/src/host-telemetry.js
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0
 
 /**
- * Host vitals for the UIK agent record (CPU load, temperature, power).
+ * Host vitals for the UIK agent record (CPU load, temperature, power, GPU).
  *
  * Contract: every field is OPTIONAL and a value we cannot read is OMITTED,
  * never zeroed. A zero renders in the dashboard as a real reading — an idle
@@ -373,6 +373,37 @@ function readLinuxTempC(sources) {
   return { tempC: Math.round(hottest * 10) / 10, limitC };
 }
 
+/**
+ * The GPU's own vitals, from `nvidia-smi`, Linux only.
+ *
+ * Board thermal zones say nothing about how busy the GPU is: a GPU box at 95%
+ * and one at 5% can read the same zone temperature, so a card needs the GPU's
+ * own load to explain its heat (petrus, 2 Nov 2026). Each field is checked on
+ * its own, since nvidia-smi prints "[N/A]" for whatever a board does not
+ * expose (GB10 has no memory clock, for one). The first GPU is reported; every
+ * box in the fleet has one.
+ *
+ * @returns {{gpu_pct?: number, gpu_temp_c?: number, gpu_mhz?: number, gpu_w?: number}}
+ */
+function readNvidiaGpu(sources) {
+  const line = sources.run('/usr/bin/nvidia-smi', [
+    '--query-gpu=utilization.gpu,temperature.gpu,clocks.sm,power.draw',
+    '--format=csv,noheader,nounits',
+  ]).split('\n')[0];
+  if (!line) return {};
+
+  const [pct, tempC, mhz, watts] = line.split(',').map((field) => {
+    const text = field.trim();
+    return /^-?\d+(\.\d+)?$/.test(text) ? Number(text) : NaN;
+  });
+  const out = {};
+  if (pct >= 0 && pct <= 100) out.gpu_pct = Math.round(pct);
+  if (tempC >= TEMP_MIN_C && tempC <= TEMP_MAX_C) out.gpu_temp_c = Math.round(tempC * 10) / 10;
+  if (mhz > 0) out.gpu_mhz = Math.round(mhz);
+  if (watts > 0) out.gpu_w = Math.round(watts * 10) / 10;
+  return out;
+}
+
 /**
  * Collect what this host can actually report.
  *
@@ -426,6 +457,12 @@ export function collectHostTelemetry({ machine, kind, model, sources = defaultSo
     // No thermal zones exposed; publish without a temperature.
   }
 
+  try {
+    if (sources.platform() === 'linux') Object.assign(host, readNvidiaGpu(sources));
+  } catch {
+    // No NVIDIA GPU, or no driver: publish without GPU fields.
+  }
+
   try {
     const hw = readHardware(sources);
     if (hw) host.hw = hw;
diff --git a/packages/user-intent-kit/test/host-telemetry.test.js b/packages/user-intent-kit/test/host-telemetry.test.js
index 2b9fa82..b8fbc96 100644
--- a/packages/user-intent-kit/test/host-telemetry.test.js
+++ b/packages/user-intent-kit/test/host-telemetry.test.js
@@ -394,3 +394,42 @@ test('a nonsense critical point is dropped rather than published', () => {
     assert.ok(!('temp_limit_c' in host), `published nonsense limit ${bad}`);
   }
 });
+
+// --- the GPU's own vitals, from nvidia-smi ---
+
+const NVSMI = '/usr/bin/nvidia-smi --query-gpu=utilization.gpu,temperature.gpu,clocks.sm,power.draw '
+  + '--format=csv,noheader,nounits';
+
+test('publishes GPU load, temperature, clock and power from nvidia-smi', () => {
+  const host = collectHostTelemetry({ sources: sources({ commands: { [NVSMI]: '95, 72, 2184, 27.64' } }) });
+  assert.equal(host.gpu_pct, 95);
+  assert.equal(host.gpu_temp_c, 72);
+  assert.equal(host.gpu_mhz, 2184);
+  assert.equal(host.gpu_w, 27.6);
+});
+
+test('a GPU field nvidia-smi cannot read is omitted, the rest still published', () => {
+  const host = collectHostTelemetry({ sources: sources({ commands: { [NVSMI]: '40, 61, 1500, [N/A]' } }) });
+  assert.equal(host.gpu_pct, 40);
+  assert.ok(!('gpu_w' in host), 'published an unreadable power draw');
+});
+
+test('an idle GPU publishes 0% load, a real reading', () => {
+  const host = collectHostTelemetry({ sources: sources({ commands: { [NVSMI]: '0, 45, 208, 4.1' } }) });
+  assert.equal(host.gpu_pct, 0);
+});
+
+test('a host without nvidia-smi publishes no GPU fields', () => {
+  const host = collectHostTelemetry({ sources: sources() });
+  for (const key of ['gpu_pct', 'gpu_temp_c', 'gpu_mhz', 'gpu_w']) assert.ok(!(key in host), `published ${key}`);
+});
+
+test('macOS never asks nvidia-smi', () => {
+  const host = collectHostTelemetry({ sources: sources({ platform: 'darwin', commands: { [NVSMI]: '95, 72, 2184, 27' } }) });
+  assert.ok(!('gpu_pct' in host));
+});
+
+test('implausible GPU readings are omitted', () => {
+  const host = collectHostTelemetry({ sources: sources({ commands: { [NVSMI]: '140, 999, -5, abc' } }) });
+  for (const key of ['gpu_pct', 'gpu_temp_c', 'gpu_mhz', 'gpu_w']) assert.ok(!(key in host), `published ${key}`);
+});