Put the assistant on the dashboard, where the questions actually occur to you

Plugin/unraid/Tools/api_cache_writer.php

@@ -91,6 +91,21 @@ $t = microtime(true);
 // Call vv_api_data() once — result is static-cached for the rest of this process.
 vv_api_data();
 
+// Must stay in step with api/monitor.php's own block. This file is what the Monitor tab
+// normally reads — the endpoint only assembles a payload on a cache miss — so a key added there
+// and not here leaves the card that consumes it loading forever on every ordinary page load,
+// and working on the one request that happens to miss the cache.
+$_vv_ai = null;
+if (vv_ai_ui_on()) {
+    require_once $_base . '/include/ai.php';
+    $_vv_ai = [
+        'model'   => vv_ai_config()['model'],
+        'runtime' => vv_ai_runtime_stats(),
+        'index'   => vv_ai_index_stats(),
+        'tokens'  => vv_ai_token_stats()['today'] ?? null,
+    ];
+}
+
 $monitor = [
     'system'          => vv_system_info(),
     'fallback'        => vv_fallback_state(),
@@ -118,6 +133,7 @@ $monitor = [
     'vms'             => vv_get_vms(),
     'docker_folders'  => vv_get_docker_folders(),
     'remote_hosts'    => vv_remote_hosts_stats(),
+    'ai'              => $_vv_ai,
     '_api_status'     => vv_api_get_status(),
     'ts'              => time(),
 ];

Plugin/unraid/api/monitor.php

@@ -64,6 +64,7 @@
 //   A flat object of the keys listed in the assembly below, plus _api_status and ts.
 //
 // DEPENDS ON
+//   include/ai.php              AI residency and index figures, only when vv_ai_ui_on()
 //   include/config.php          vv_cache_read()
 //   include/common.php          raw hardware metrics — system, cpu, mem, net, gpu, disks,
 //                               ups, parity, containers, transcodes, remote roll-ups
@@ -92,6 +93,24 @@ require_once dirname(__DIR__) . '/include/docker_folders.php';
 // Pre-warm the API cache with one request (shared by all API-first functions below).
 vv_api_data();
 
+// AI residency, for the Monitor tab's AI row. Collected here rather than by that card polling
+// api/ai.php on its own cycle: vv_ai_runtime_stats() calls out to Ollama and shells nvidia-smi,
+// and this payload is written once a minute by the background writer — a card polling it
+// directly would pay both costs every five seconds on every open tab.
+//
+// Null on any host that is not the AI host or has AI_ENABLED false, which is also what makes the
+// row absent rather than empty there. Same shape as every other optional subsystem on this page.
+$_vv_ai = null;
+if (vv_ai_ui_on()) {
+    require_once dirname(__DIR__) . '/include/ai.php';
+    $_vv_ai = [
+        'model'   => vv_ai_config()['model'],
+        'runtime' => vv_ai_runtime_stats(),
+        'index'   => vv_ai_index_stats(),
+        'tokens'  => vv_ai_token_stats()['today'] ?? null,
+    ];
+}
+
 echo json_encode([
     'system'          => vv_system_info(),
     'fallback'        => vv_fallback_state(),
@@ -119,6 +138,7 @@ echo json_encode([
     'vms'            => vv_get_vms(),
     'docker_folders' => vv_get_docker_folders(),
     'remote_hosts'   => vv_remote_hosts_stats(),
+    'ai'             => $_vv_ai,
     '_api_status'    => vv_api_get_status(),
     'ts'             => time(),
 ]);

Plugin/unraid/pages/monitor.php

@@ -29,7 +29,15 @@
 //
 //   Missing subsystems simply do not render.
 //       No GPU, no UPS, no VMs — the corresponding card is absent rather than showing zeros
-//       or an error. The page is built to be correct on hardware lacking any given part.
+//       or an error. The page is built to be correct on hardware lacking any given part. The
+//       AI row is the same rule applied to a subsystem rather than a device: it exists only
+//       where vv_ai_ui_on() is true, which is the AI host with AI_ENABLED set.
+//
+//   The assistant here starts on General Chat, where the AI tab starts on Varaverk Assistant.
+//       Different jobs. The tab is where you go to interrogate the installation; this is the
+//       box you type an idle question into while watching the dashboard. The worker escalates
+//       anything genuinely about this install to the strict profile on its own, so starting
+//       loose costs nothing and starting strict would refuse ordinary questions.
 //
 // OPERATIONAL SAFEGUARDS
 //   The health roll-up must not default to healthy.
@@ -63,17 +71,22 @@
 // RENDERS
 //   System header, CPU per core, memory breakdown, GPU cards, storage pools and array disks,
 //   network, UPS, VMs, containers, transcode sessions, media now-playing, watchdog summary,
-//   partner node cards
+//   partner node cards, and — on the AI host only — model residency, saved conversations and
+//   an assistant
 //
 // DEPENDS ON
 //   include/monitor.php     required directly for initial render
+//   include/ai_chat.php     the AI row's chat and conversation list, shared with the AI tab
 //   api/monitor.php         full payload, slower cycle
 //   api/monitor_fast.php    fast-moving values, 1s
 //   api/media.php           now-playing sessions
 //   api/docker_action.php   container actions
 //   api/flag_toggle.php     toggles
+//   api/ai.php              the AI row's turns and conversation store
 // ═══════════════════════════════════════════════════════════════════════════════════════════════
 require_once dirname(__DIR__) . '/include/monitor.php';
+require_once dirname(__DIR__) . '/include/ai_chat.php';
+if (vv_ai_ui_on()) vv_ai_chat_assets();
 ?>
 <style>
 @keyframes vvRsPulse {
@@ -279,6 +292,59 @@ require_once dirname(__DIR__) . '/include/monitor.php';
       <div id="vv-array-body">Loading...</div>
     </div>
 
+<?php if (vv_ai_ui_on()): ?>
+    <!-- Row 5: AI residency | Saved conversations | Assistant -->
+    <!-- Present only on the AI host with AI_ENABLED true, on the same footing as the GPU and UPS
+         cards above: a subsystem that is not here does not render an empty card explaining that
+         it is not here. api/ai.php refuses every action independently, so this is presentation
+         rather than access control. -->
+
+    <div class="vv-card" id="vv-ai-stats-card" style="grid-column:span 1;">
+      <h3>
+        <span style="display:flex;align-items:center;gap:5px;">
+          <span class="vv-ico"><svg width="12" height="12" viewBox="0 0 12 12" fill="none" stroke="#aaa" stroke-width="1.1" stroke-linecap="round" stroke-linejoin="round"><rect x="2.5" y="2.5" width="7" height="7" rx="1.2"/><circle cx="4.7" cy="5" r="0.7" fill="#888" stroke="none"/><circle cx="7.3" cy="5" r="0.7" fill="#888" stroke="none"/><line x1="4.5" y1="7.3" x2="7.5" y2="7.3"/><line x1="6" y1="2.5" x2="6" y2="0.8"/><line x1="2.5" y1="6" x2="0.8" y2="6"/><line x1="9.5" y1="6" x2="11.2" y2="6"/></svg></span>
+          AI
+        </span>
+        <a href="/plugins/varaverk/Varaverk.page?tab=ai" class="vv-card-cog" title="AI tab">⚙</a>
+      </h3>
+      <div id="vv-ai-stats-body">Loading...</div>
+    </div>
+
+    <div class="vv-card" id="vv-ai-chats-card" style="grid-column:span 2;">
+      <h3>
+        <span style="display:flex;align-items:center;gap:5px;">
+          <span class="vv-ico"><svg width="13" height="12" viewBox="0 0 13 12" fill="none" stroke="#aaa" stroke-width="1.1" stroke-linecap="round" stroke-linejoin="round"><path d="M1 2.2C1 1.5 1.5 1 2.2 1H10.8C11.5 1 12 1.5 12 2.2V6.8C12 7.5 11.5 8 10.8 8H4.5L2 10.5V8H2.2C1.5 8 1 7.5 1 6.8Z"/></svg></span>
+          Conversations
+        </span>
+      </h3>
+      <?php vv_ai_chat_list_markup('vv-mon-ai'); ?>
+    </div>
+
+    <!-- Named for what it is, not "dock" — the Scheduler tab's vv-ai-dock is a different
+         component and vvAiDockOn() there tests for that id by name. -->
+    <div class="vv-card" id="vv-ai-assistant-card" style="grid-column:span 5;">
+      <h3>
+        <span style="display:flex;align-items:center;gap:5px;">
+          <span class="vv-ico"><svg width="12" height="12" viewBox="0 0 12 12" fill="none" stroke="#aaa" stroke-width="1.1" stroke-linecap="round" stroke-linejoin="round"><circle cx="6" cy="6" r="5"/><path d="M4.4 4.6C4.4 3.7 5.1 3.1 6 3.1C6.9 3.1 7.6 3.7 7.6 4.5C7.6 5.9 6 5.7 6 7"/><circle cx="6" cy="8.8" r="0.6" fill="#888" stroke="none"/></svg></span>
+          Assistant
+        </span>
+      </h3>
+      <?php
+      // Starts on General Chat, unlike the AI tab. This is the box you type an idle question
+      // into while watching the dashboard, and the worker escalates anything about this install
+      // to the assistant on its own — so landing on the strict profile here would refuse
+      // ordinary questions to guard against a mistake the server already prevents.
+      vv_ai_chat_markup('vv-mon-ai', [
+          'profile'     => 'chat',
+          'compact'     => true,
+          'height'      => '300px',
+          'empty'       => 'Ask anything. Questions about this installation are handed to the '
+                         . 'Varaverk assistant automatically.',
+          'placeholder' => 'Ask the assistant…',
+      ]);
+      ?>
+    </div>
+<?php endif; ?>
 
 </div>
 
@@ -1817,10 +1883,63 @@ function vvPollMonitor(live) {
       dfData.vms = d.vms ?? { available: false, vms: [] };
       vvRenderDockerFolders(dfData);
 
+      // ── AI ──────────────────────────────────────────────────────────────────
+      if (d.ai) vvRenderAi(d.ai);
+
     })
     .catch(vvPollFailed);
 }
 
+// ── AI residency card ────────────────────────────────────────────────────────
+// Leads with offload, not with size. 100% on this card is the difference between ~62 tok/s and
+// roughly a quarter of that, and nothing else in the WebGUI surfaces it — a model that has
+// quietly fallen back to partial CPU offload is otherwise invisible until answers feel slow.
+function vvRenderAi(ai) {
+  const box = document.getElementById('vv-ai-stats-body');
+  if (!box) return;
+  const rt = ai.runtime ?? {}, ix = ai.index ?? {}, tok = ai.tokens ?? null;
+
+  const line = (label, value, cls) =>
+      `<div style="display:flex;justify-content:space-between;align-items:baseline;gap:8px;margin-bottom:6px;">`
+    + `<span style="font-size:11px;color:#666;">${vvEscHtml(label)}</span>`
+    + `<span style="font-size:11px;font-family:monospace;${cls || 'color:#999;'}">${vvEscHtml(value)}</span></div>`;
+
+  let h = '';
+  if (!rt.reachable) {
+    h += line('Ollama', 'unreachable', 'color:#e57;');
+  } else if (!rt.loaded) {
+    h += line('Model', 'not loaded', 'color:#ffb74d;');
+    h += `<div style="font-size:10px;color:#444;margin:-2px 0 7px;">loads on first question</div>`;
+  } else {
+    const p = rt.offload_pct;
+    const full = p === 100;
+    h += line('GPU offload', p === null ? '—' : p + '%', full ? 'color:#6fcf97;' : 'color:#ffb74d;');
+    h += `<div style="font-size:10px;color:#444;margin:-2px 0 7px;">`
+       + (full ? 'all on GPU' : 'layers on CPU — slow') + `</div>`;
+    if (rt.context) h += line('Context', rt.context.toLocaleString());
+  }
+
+  if (rt.gpu) {
+    h += line('VRAM', (rt.gpu.mem_used/1024).toFixed(1) + '/' + (rt.gpu.mem_total/1024).toFixed(1) + ' GB');
+    h += line('GPU util', rt.gpu.util + '%');
+  }
+
+  // Staleness is stated, not implied. An index older than the newest tracked file answers
+  // confidently out of code that has since changed — the one failure a grounded answer cannot
+  // reveal on its own.
+  if (!ix.exists) {
+    h += line('Index', 'not built', 'color:#e57;');
+  } else {
+    h += line('Index', ix.chunks.toLocaleString() + ' chunks',
+              ix.stale ? 'color:#ffb74d;' : 'color:#999;');
+    if (ix.stale) h += `<div style="font-size:10px;color:#8a6a3a;margin:-2px 0 7px;">source newer — reindex</div>`;
+  }
+
+  if (tok && tok.turns) h += line('Today', tok.total.toLocaleString() + ' tok');
+
+  box.innerHTML = h;
+}
+
 // 5s against a payload the cache writer refreshes once a minute. Polling faster cannot make the
 // data newer — it only decides how soon the page notices the writer's update.
 vvPollRunner(vvPollMonitor, 5000);
@@ -2481,4 +2600,26 @@ document.addEventListener('click', () => {
 // on this page ever called the function — there are no such buttons, on this or any other tab.
 // Removed rather than left as an unreachable handler for the platform's three most destructive
 // operations. The endpoint stays; see its header for why it is kept unwired.
+
+// ── AI row ───────────────────────────────────────────────────────────────────
+// Constructed only when the row rendered. The card markup is behind vv_ai_ui_on(), so on any
+// other host these ids do not exist and the factories are never called — the row is absent
+// rather than broken.
+//
+// Same store as the AI tab, so a conversation started here is the one you carry on there. The
+// instance keys itself on its prefix and tears down any predecessor, which matters on this tab
+// specifically: Unraid swaps tab content by AJAX without unloading the previous page's script,
+// and this page already has three poll loops that survive that.
+if (document.getElementById('vv-mon-ai-chat')) {
+  let vvMonChatList = null;
+  const vvMonChat = VvAiChat({
+    prefix:  'vv-mon-ai',
+    profile: 'chat',
+    empty:   'Ask anything. Questions about this installation are handed to the Varaverk '
+           + 'assistant automatically.',
+    onChats: id => { if (vvMonChatList) vvMonChatList.setActive(id); },
+  });
+  vvMonChatList = VvAiChatList({ into: 'vv-mon-ai-chats', chat: vvMonChat });
+  vvMonChatList.reload();
+}
 </script>