v0.3.0 - Hardware dashboard + knob context + Explain context + Open WebUI link

Hardware dashboard:
- New hardware.py module: SSH probes each Spark for hostname, uptime, load+cores, RAM, disk, GPU (name, util, temp, power) + per-process GPU memory sum
- DGX Spark uses unified memory (nvidia-smi memory.total returns N/A); fall back to per-process compute memory and compute fraction against system RAM. Marks with gpu_unified_memory=true.
- 4s TTL cache in HardwareProbe to avoid hammering
- /api/hardware returns per-Spark snapshot
- UI: 'Spark hardware' section at the top with per-Spark cards (CPU load, RAM, GPU mem (unified), GPU util + temp + power, disk) — bars with warn threshold styling
- Polls every 8s

Knob context (tied to live hardware):
- Each Advanced knob now shows plain-English help text
- 'GPU memory %' shows '~N GB allocated · ~M GB left for OS/buffers' computed from actual Spark RAM
- 'Max context' shows '~N pages of text'
- Toggles show tradeoff descriptions

Explain context:
- '✨ Explain context' button on the update banner
- /api/explain-updates POST: forwards pending commits to the loaded vLLM model and streams its response back as SSE
- Renders into an expandable 'Explained by the loaded LLM' section under Pending commits
- Reasoning tokens shown italicized when the model emits them

Open WebUI integration:
- New 'Open WebUI URL' optional field in Configure Sparks
- /api/config exposes it; UI shows 'Open chat ↗' button in the top bar if set

Downloads:
- Third radio option: Spark 1 only / Spark 2 only / Both Sparks
- Backend picks SSH target based on mode
- HF repo link icon next to the input
- Helper line about NVFP4 for Blackwell

Model cards:
- Repo name is now a clickable link to its Hugging Face page

Package: bump 0.3.0:0

image/app/config.py

@@ -42,6 +42,7 @@ class Settings:
     parakeet_port: int
     magpie_port: int
     bind_port: int
+    open_webui_url: str
 
     @classmethod
     def from_env(cls) -> "Settings":
@@ -66,6 +67,7 @@ class Settings:
             parakeet_port=int(_env("PARAKEET_PORT", "8000")),
             magpie_port=int(_env("MAGPIE_PORT", "9000")),
             bind_port=int(_env("BIND_PORT", "9999")),
+            open_webui_url=_env("OPEN_WEBUI_URL", ""),
         )
 
     @property

image/app/download.py

@@ -19,7 +19,7 @@ from .config import Settings
 from .ssh import ssh_stream, StreamHandle
 
 
-Mode = Literal["solo", "cluster"]
+Mode = Literal["spark1", "spark2", "cluster"]
 
 
 _TQDM_RE = re.compile(
@@ -113,17 +113,26 @@ class DownloadManager:
 
     async def _do(self, job: DownloadJob) -> None:
         s = self.settings
-        if not s.spark1_host or not s.spark1_user:
+        # Pick the SSH target and hf-download flags from the mode.
-            raise RuntimeError("spark1 not configured")
+        if job.mode == "spark2":
+            target_host, target_user = s.spark2_host, s.spark2_user
+            flags = ""
+        elif job.mode == "cluster":
+            target_host, target_user = s.spark1_host, s.spark1_user
+            flags = "-c --copy-parallel"
+        else:  # spark1
+            target_host, target_user = s.spark1_host, s.spark1_user
+            flags = ""
+        if not target_host or not target_user:
+            raise RuntimeError(f"{job.mode} host not configured")
 
-        flags = "-c --copy-parallel" if job.mode == "cluster" else ""
         cmd = f"cd ~/spark-vllm-docker && ./hf-download.sh {job.repo} {flags}".strip()
         job.append(f"$ {cmd}")
         job.state = "downloading"
         job.progress.phase = "Connecting to Hugging Face…"
 
         handle = StreamHandle()
-        async for line in ssh_stream(s.spark1_host, s.spark1_user, cmd, s, handle=handle):
+        async for line in ssh_stream(target_host, target_user, cmd, s, handle=handle):
             job.append(line)
             self._update_progress(job, line)
 

image/app/hardware.py

@@ -0,0 +1,118 @@
+"""Per-Spark hardware snapshots: RAM, disk, GPU memory + utilization, CPU load, uptime.
+
+Drives via a single SSH command per Spark that runs `free`, `df`, `nvidia-smi`,
+`/proc/loadavg`, and `uptime -p` and prints labeled lines back. We parse those
+labels in `_parse`.
+"""
+from __future__ import annotations
+import asyncio
+import time
+from typing import Any
+
+from .config import Settings
+from .ssh import ssh_run
+
+
+_PROBE = r"""
+set -e
+echo HOSTNAME=$(hostname)
+echo UPTIME=$(uptime -p 2>/dev/null || uptime)
+echo LOAD=$(awk '{print $1, $2, $3}' /proc/loadavg)
+echo CORES=$(nproc 2>/dev/null || echo 0)
+echo MEMORY=$(free -b 2>/dev/null | awk '/^Mem:/ {print $2, $3}')
+echo DISK=$(df -B1 / 2>/dev/null | awk 'NR==2 {print $2, $3}')
+echo GPU=$(nvidia-smi --query-gpu=name,utilization.gpu,temperature.gpu,power.draw,memory.total --format=csv,noheader,nounits 2>/dev/null | head -1)
+echo GPU_MEM_USED_MIB=$(nvidia-smi --query-compute-apps=used_gpu_memory --format=csv,noheader,nounits 2>/dev/null | awk '{s+=$1} END {print s+0}')
+""".strip()
+
+
+def _parse_int(s: str) -> int | None:
+    try: return int(s)
+    except (TypeError, ValueError): return None
+
+
+def _parse(out: str) -> dict:
+    info: dict[str, Any] = {}
+    for raw in out.splitlines():
+        if "=" not in raw:
+            continue
+        k, v = raw.split("=", 1)
+        info[k.strip().lower()] = v.strip()
+    parsed: dict[str, Any] = {}
+    parsed["hostname"] = info.get("hostname")
+    parsed["uptime"] = info.get("uptime")
+    parsed["cores"] = _parse_int(info.get("cores", ""))
+    # Load average -> (1m, 5m, 15m)
+    if info.get("load"):
+        loads = info["load"].split()
+        try:
+            parsed["load"] = [float(x) for x in loads[:3]]
+        except ValueError:
+            parsed["load"] = None
+    # Memory: total used in bytes
+    if info.get("memory"):
+        mem = info["memory"].split()
+        if len(mem) == 2:
+            tot, used = _parse_int(mem[0]), _parse_int(mem[1])
+            parsed["ram_total_bytes"] = tot
+            parsed["ram_used_bytes"] = used
+    # Disk: total used in bytes
+    if info.get("disk"):
+        dk = info["disk"].split()
+        if len(dk) == 2:
+            parsed["disk_total_bytes"] = _parse_int(dk[0])
+            parsed["disk_used_bytes"] = _parse_int(dk[1])
+    # GPU: "name, util_gpu, temp_C, power_W, memory_total_MiB"
+    if info.get("gpu"):
+        parts = [p.strip() for p in info["gpu"].split(",")]
+        if len(parts) >= 5:
+            name, ug, temp, power, mt = parts[0], parts[1], parts[2], parts[3], parts[4]
+            parsed["gpu_name"] = name
+            parsed["gpu_util_pct"] = _parse_int(ug)
+            parsed["gpu_temp_c"] = _parse_int(temp)
+            try: parsed["gpu_power_w"] = float(power)
+            except ValueError: parsed["gpu_power_w"] = None
+            # memory.total may be "[N/A]" on unified-memory systems (DGX Spark)
+            parsed["gpu_mem_total_mib"] = _parse_int(mt)
+            parsed["gpu_unified_memory"] = parsed["gpu_mem_total_mib"] is None
+    # Sum per-process compute memory (works even on unified-memory systems)
+    if info.get("gpu_mem_used_mib"):
+        parsed["gpu_mem_used_mib"] = _parse_int(info["gpu_mem_used_mib"])
+    return parsed
+
+
+class HardwareProbe:
+    """Caches results briefly to avoid hammering the Sparks."""
+
+    def __init__(self, settings: Settings, ttl_sec: float = 4.0) -> None:
+        self.settings = settings
+        self.ttl_sec = ttl_sec
+        self._cache: dict[str, tuple[float, dict]] = {}
+        self._locks: dict[str, asyncio.Lock] = {}
+
+    def _lock(self, key: str) -> asyncio.Lock:
+        if key not in self._locks:
+            self._locks[key] = asyncio.Lock()
+        return self._locks[key]
+
+    async def fetch(self) -> dict:
+        return {
+            "spark1": await self._one("spark1", self.settings.spark1_host, self.settings.spark1_user),
+            "spark2": await self._one("spark2", self.settings.spark2_host, self.settings.spark2_user),
+        }
+
+    async def _one(self, key: str, host: str, user: str) -> dict:
+        if not host or not user:
+            return {"reachable": False, "configured": False}
+        async with self._lock(key):
+            now = time.monotonic()
+            cached = self._cache.get(key)
+            if cached and (now - cached[0] < self.ttl_sec):
+                return cached[1]
+            rc, out, err = await ssh_run(host, user, _PROBE, self.settings, timeout=8)
+            if rc != 0:
+                result = {"reachable": False, "configured": True, "host": host, "error": err.strip() or out.strip() or f"rc={rc}"}
+            else:
+                result = {"reachable": True, "configured": True, "host": host, **_parse(out)}
+            self._cache[key] = (now, result)
+            return result

image/app/server.py

@@ -11,6 +11,7 @@ from typing import Literal
 
 from .config import Settings
 from .download import DownloadManager
+from .hardware import HardwareProbe
 from .health import check_magpie, check_parakeet, check_vllm
 from .models import load_catalog
 from .overrides import add_custom, delete_custom, extract_knobs_from_args, load_overrides, set_knobs
@@ -25,6 +26,7 @@ catalog = load_catalog(settings.models_yaml)
 swap_manager = SwapManager(settings, catalog)
 download_manager = DownloadManager(settings)
 update_manager = UpdateManager(settings)
+hardware_probe = HardwareProbe(settings)
 
 app = FastAPI(title="spark-control", version="0.1.0")
 
@@ -44,6 +46,7 @@ async def get_config() -> dict:
         "spark1_host": settings.spark1_host,
         "spark2_host": settings.spark2_host,
         "vllm_port": settings.vllm_port,
+        "open_webui_url": settings.open_webui_url or None,
     }
 
 
@@ -116,6 +119,12 @@ async def del_model(key: str) -> dict:
     return {"ok": True, "key": key}
 
 
+@app.get("/api/hardware")
+async def get_hardware() -> dict:
+    """Per-Spark hardware snapshot — RAM, disk, GPU mem + util, CPU load, uptime."""
+    return await hardware_probe.fetch()
+
+
 @app.get("/api/services")
 async def get_services() -> dict:
     """Lifecycle state of always-on support services (Parakeet, Magpie, …).
@@ -297,7 +306,7 @@ async def stream_swap(job_id: str):
 
 class DownloadRequest(BaseModel):
     repo: str
-    mode: Literal["solo", "cluster"] = "solo"
+    mode: Literal["spark1", "spark2", "cluster"] = "spark1"
 
 
 @app.post("/api/download")
@@ -376,6 +385,81 @@ async def get_updates() -> dict:
     return await get_update_status(settings)
 
 
+@app.get("/api/explain-updates")
+async def explain_updates():
+    """Stream a layman's explanation of the pending commits from the currently-loaded vLLM model."""
+    import httpx
+    info = await get_update_status(settings)
+    if not info.get("ok"):
+        async def err_gen():
+            yield f"event: done\ndata: {json.dumps({'error': info.get('error', 'unknown')})}\n\n"
+        return StreamingResponse(err_gen(), media_type="text/event-stream")
+
+    vllm = await check_vllm(settings)
+    if not vllm.get("ok") or not vllm.get("current_model"):
+        async def err_gen():
+            yield f"event: done\ndata: {json.dumps({'error': 'no vLLM model loaded — swap to a model first'})}\n\n"
+        return StreamingResponse(err_gen(), media_type="text/event-stream")
+
+    commits = "\n".join(info.get("log", []))
+    if not commits.strip():
+        async def empty_gen():
+            yield f"event: done\ndata: {json.dumps({'error': 'no pending commits'})}\n\n"
+        return StreamingResponse(empty_gen(), media_type="text/event-stream")
+
+    prompt = (
+        "You are reviewing pending git commits to `eugr/spark-vllm-docker`, an upstream community project that "
+        "orchestrates vLLM on dual NVIDIA DGX Spark hardware (Blackwell GPUs, cluster via Ray, recipes per model). "
+        "The reader has a setup running models like Qwen3.6-35B-A3B-NVFP4 (daily driver, solo), Qwen3-VL 235B (cluster), "
+        "and Gemma 4 31B. The reader is technically literate but is NOT a vLLM expert.\n\n"
+        "For the commit list below: give a short overall verdict (Apply / Optional / Skip and why), then a brief "
+        "bullet per commit grouping similar ones. Call out anything that would break a working setup or that "
+        "requires re-downloading models. Avoid jargon. ~250 words max.\n\n"
+        f"Pending commits:\n{commits}"
+    )
+
+    async def gen():
+        try:
+            async with httpx.AsyncClient(timeout=httpx.Timeout(300.0, connect=5.0)) as c:
+                async with c.stream(
+                    "POST",
+                    f"{vllm['base_url']}/chat/completions",
+                    json={
+                        "model": vllm["current_model"],
+                        "stream": True,
+                        "messages": [{"role": "user", "content": prompt}],
+                        "max_tokens": 600,
+                        "temperature": 0.4,
+                    },
+                ) as r:
+                    r.raise_for_status()
+                    async for line in r.aiter_lines():
+                        if not line.startswith("data: "):
+                            continue
+                        data = line[6:].strip()
+                        if data == "[DONE]":
+                            break
+                        try:
+                            chunk = json.loads(data)
+                            choices = chunk.get("choices") or []
+                            if not choices:
+                                continue
+                            delta = choices[0].get("delta") or {}
+                            text = delta.get("content")
+                            reasoning = delta.get("reasoning")
+                            if text:
+                                yield f"data: {json.dumps({'content': text})}\n\n"
+                            elif reasoning:
+                                yield f"data: {json.dumps({'reasoning': reasoning})}\n\n"
+                        except json.JSONDecodeError:
+                            continue
+        except Exception as e:
+            yield f"data: {json.dumps({'error': f'{type(e).__name__}: {e}'})}\n\n"
+        yield f"event: done\ndata: {json.dumps({'ok': True})}\n\n"
+
+    return StreamingResponse(gen(), media_type="text/event-stream")
+
+
 class UpdateRequest(BaseModel):
     mode: Literal["solo", "cluster"] = "cluster"
 

image/app/static/app.js

@@ -13,6 +13,8 @@ const state = {
   swap_progress: 0,        // 0–1
   services: {},
   service_action_in_flight: null,  // e.g. "parakeet:restart"
+  hardware: {},
+  config: {},
   configured: true,
   timer_handle: null,
 };
@@ -63,7 +65,9 @@ function renderCards() {
         ${(m.capabilities || []).map(c => `<span class="tag cap">${escapeHtml(c)}</span>`).join('')}
       </div>
       ${desc}
-      <div class="muted small repo">${escapeHtml(m.repo)}</div>
+      <div class="muted small repo">
+        <a href="https://huggingface.co/${encodeURIComponent(m.repo)}" target="_blank" rel="noopener" title="View on Hugging Face">${escapeHtml(m.repo)} <span class="hf-icon">↗</span></a>
+      </div>
       <div class="spacer"></div>
       <div class="card-actions">
         <button class="btn ${isActive ? '' : 'primary'}" data-swap-key="${key}" ${isActive || isSwapping ? 'disabled' : ''}>
@@ -93,6 +97,107 @@ function renderCurrent(status) {
   c.innerHTML = `<strong>${label}</strong>`;
 }
 
+// ===================== hardware dashboard =====================
+
+function fmtBytes(n) {
+  if (!n && n !== 0) return '—';
+  const u = ['B', 'KB', 'MB', 'GB', 'TB'];
+  let i = 0; let v = n;
+  while (v >= 1024 && i < u.length - 1) { v /= 1024; i++; }
+  return v < 10 ? `${v.toFixed(1)} ${u[i]}` : `${Math.round(v)} ${u[i]}`;
+}
+function fmtMiB(n) {
+  if (!n && n !== 0) return null;
+  // n is in MiB; render in GB
+  const gb = n / 1024;
+  return gb < 10 ? gb.toFixed(1) : Math.round(gb).toString();
+}
+
+function bar(usedPct, warn) {
+  const pct = Math.max(2, Math.min(100, usedPct));
+  return `<div class="bar ${warn ? 'warn' : ''}"><span style="width:${pct}%"></span></div>`;
+}
+
+async function pollHardware() {
+  try {
+    state.hardware = await fetchJSON('/api/hardware');
+    renderHardware();
+  } catch (e) { console.warn('hardware poll failed', e); }
+}
+
+function renderHardware() {
+  const panel = el('#hardware-panel');
+  const grid = el('#hardware-grid');
+  const hw = state.hardware || {};
+  const keys = Object.keys(hw).filter(k => hw[k] && (hw[k].configured !== false));
+  if (keys.length === 0) { panel.classList.add('hidden'); return; }
+  panel.classList.remove('hidden');
+  grid.innerHTML = '';
+  for (const key of keys) {
+    const s = hw[key];
+    const card = document.createElement('div');
+    if (!s.reachable) {
+      card.className = 'hw-card unreachable';
+      card.innerHTML = `
+        <div class="head">
+          <span class="name">${escapeHtml(key)}</span>
+          <span class="meta">unreachable</span>
+        </div>
+        <div class="muted small">${escapeHtml(s.host || '')} — ${escapeHtml(s.error || 'no response')}</div>
+      `;
+      grid.appendChild(card);
+      continue;
+    }
+    const ramPct = s.ram_used_bytes && s.ram_total_bytes ? (s.ram_used_bytes / s.ram_total_bytes) * 100 : 0;
+    const diskPct = s.disk_used_bytes && s.disk_total_bytes ? (s.disk_used_bytes / s.disk_total_bytes) * 100 : 0;
+    const loadPct = (s.load && s.cores) ? Math.min(100, (s.load[0] / s.cores) * 100) : 0;
+    // GPU memory: on unified-memory systems (DGX Spark) total is N/A, so use system RAM as the pool.
+    const gpuMemTotalMiB = s.gpu_mem_total_mib || (s.gpu_unified_memory ? (s.ram_total_bytes / (1024 * 1024)) : null);
+    const gpuMemUsedMiB = s.gpu_mem_used_mib ?? null;
+    const gpuMemPct = (gpuMemTotalMiB && gpuMemUsedMiB != null) ? (gpuMemUsedMiB / gpuMemTotalMiB) * 100 : 0;
+    const gpuMemNote = s.gpu_unified_memory ? ' <span class="muted">(unified)</span>' : '';
+    const gpuExtras = [];
+    if (s.gpu_temp_c != null) gpuExtras.push(`${s.gpu_temp_c}°C`);
+    if (s.gpu_power_w != null) gpuExtras.push(`${s.gpu_power_w.toFixed(0)}W`);
+    const gpuExtrasStr = gpuExtras.length ? ` · ${gpuExtras.join(' · ')}` : '';
+    card.className = 'hw-card';
+    card.innerHTML = `
+      <div class="head">
+        <span class="name">${escapeHtml(s.hostname || key)}</span>
+        <span class="meta">${escapeHtml(key)} · ${escapeHtml(s.gpu_name || '')} · ${escapeHtml(s.uptime || '')}</span>
+      </div>
+      <div class="hw-metric">
+        <span class="label">CPU</span>
+        ${bar(loadPct, loadPct > 80)}
+        <span class="val">${s.load ? s.load[0].toFixed(2) : '—'} / ${s.cores || '?'} cores</span>
+      </div>
+      <div class="hw-metric">
+        <span class="label">RAM</span>
+        ${bar(ramPct, ramPct > 85)}
+        <span class="val">${fmtBytes(s.ram_used_bytes)} / ${fmtBytes(s.ram_total_bytes)}</span>
+      </div>
+      <div class="hw-metric">
+        <span class="label">GPU mem${gpuMemNote}</span>
+        ${bar(gpuMemPct, gpuMemPct > 90)}
+        <span class="val">${fmtMiB(gpuMemUsedMiB) || '—'} / ${fmtMiB(gpuMemTotalMiB) || '?'} GB</span>
+      </div>
+      <div class="hw-metric">
+        <span class="label">GPU util</span>
+        ${bar(s.gpu_util_pct || 0, (s.gpu_util_pct || 0) > 90)}
+        <span class="val">${s.gpu_util_pct ?? 0}%${gpuExtrasStr}</span>
+      </div>
+      <div class="hw-metric">
+        <span class="label">Disk</span>
+        ${bar(diskPct, diskPct > 85)}
+        <span class="val">${fmtBytes(s.disk_used_bytes)} / ${fmtBytes(s.disk_total_bytes)}</span>
+      </div>
+    `;
+    grid.appendChild(card);
+  }
+}
+
+// ===================== service classification =====================
+
 function classifyService(s) {
   // returns one of: running | unhealthy | missing | unconfigured | starting
   if (!s.host) return 'unconfigured';
@@ -543,6 +648,18 @@ function openDownloadForm() {
   el('#download-form').classList.remove('hidden');
   el('#download-progress').classList.add('hidden');
   el('#dl-repo').focus();
+  updateDlHfLink();
+}
+
+function updateDlHfLink() {
+  const repo = el('#dl-repo').value.trim();
+  const link = el('#dl-hf-link');
+  if (repo.includes('/')) {
+    link.href = `https://huggingface.co/${encodeURIComponent(repo)}`;
+    link.classList.remove('hidden');
+  } else {
+    link.classList.add('hidden');
+  }
 }
 
 function closeDownloadPanel() {
@@ -672,6 +789,47 @@ function handleDownloadDone(d) {
 
 // ===================== Advanced / Add to catalog =====================
 
+function gpuTotalGB(modelMode) {
+  // Solo uses Spark 1's GPU only. Cluster shares across both — but loading is per-Spark.
+  const s1 = state.hardware?.spark1;
+  const s2 = state.hardware?.spark2;
+  const g1 = s1?.gpu_mem_total_mib ? s1.gpu_mem_total_mib / 1024 : null;
+  const g2 = s2?.gpu_mem_total_mib ? s2.gpu_mem_total_mib / 1024 : null;
+  if (modelMode === 'cluster' && g1 && g2) return Math.min(g1, g2);  // bottleneck
+  return g1 || g2 || null;
+}
+
+function knobContextHint(field, value, mode) {
+  if (field === 'gpu_memory_utilization') {
+    const gb = gpuTotalGB(mode);
+    if (!gb) return '';
+    const used = (value * gb).toFixed(0);
+    const free = (gb - value * gb).toFixed(0);
+    return `~${used} GB allocated to model + KV cache · ~${free} GB left for OS, buffers, other GPU workloads.`;
+  }
+  if (field === 'max_model_len') {
+    if (!value) return '';
+    const pages = Math.round(value / 350);  // ~350 tokens per page
+    const kvBytes = (value * 2 * 4 * 32 * 128);  // rough fp16 KV cache size for typical 32-layer model
+    return `~${pages.toLocaleString()} pages of text (very rough). Larger context = more GPU memory reserved for KV cache.`;
+  }
+  if (field === 'fastsafetensors') return value ? 'Faster cold-start weight loading.' : 'Standard safetensors loading.';
+  if (field === 'prefix_caching') return value ? 'Reuses GPU state for repeated prompt prefixes (e.g. long system prompts).' : 'Off — every request re-processes the full prompt.';
+  if (field === 'kv_cache_dtype') return value === 'fp8' ? 'Halves KV cache memory (fits ~2× more context). Quality cost is usually imperceptible.' : 'Default precision.';
+  return '';
+}
+
+function ensureKnobHint(rowEl, id) {
+  let h = rowEl.querySelector(`.knob-hint[data-for="${id}"]`);
+  if (!h) {
+    h = document.createElement('div');
+    h.className = 'knob-hint muted small';
+    h.dataset.for = id;
+    rowEl.appendChild(h);
+  }
+  return h;
+}
+
 function openAdvanced(key) {
   const m = state.models[key];
   if (!m) return;
@@ -684,6 +842,23 @@ function openAdvanced(key) {
   el('#adv-fst').checked = !!k.fastsafetensors;
   el('#adv-pcache').checked = !!k.prefix_caching;
   el('#adv-fp8').checked = k.kv_cache_dtype === 'fp8';
+
+  // Wire up live knob hints
+  const updateHints = () => {
+    const mml = parseInt(el('#adv-mml').value, 10);
+    const gmu = parseFloat(el('#adv-gmu').value);
+    ensureKnobHint(el('#adv-mml').parentElement, 'mml').textContent = knobContextHint('max_model_len', mml, m.mode);
+    ensureKnobHint(el('#adv-gmu').parentElement, 'gmu').textContent = knobContextHint('gpu_memory_utilization', gmu, m.mode);
+    ensureKnobHint(el('#adv-fst').parentElement, 'fst').textContent = knobContextHint('fastsafetensors', el('#adv-fst').checked, m.mode);
+    ensureKnobHint(el('#adv-pcache').parentElement, 'pcache').textContent = knobContextHint('prefix_caching', el('#adv-pcache').checked, m.mode);
+    ensureKnobHint(el('#adv-fp8').parentElement, 'fp8').textContent = knobContextHint('kv_cache_dtype', el('#adv-fp8').checked ? 'fp8' : 'auto', m.mode);
+  };
+  updateHints();
+  el('#adv-mml').oninput = updateHints;
+  el('#adv-gmu').oninput = (e) => { el('#adv-gmu-out').value = parseFloat(e.target.value).toFixed(2); updateHints(); };
+  el('#adv-fst').onchange = updateHints;
+  el('#adv-pcache').onchange = updateHints;
+  el('#adv-fp8').onchange = updateHints;
   const del = el('#adv-delete');
   del.classList.toggle('hidden', !m.custom);
   del.onclick = async () => {
@@ -778,6 +953,56 @@ function setupAdvancedDialog() {
   el('#adv-gmu').addEventListener('input', (e) => { el('#adv-gmu-out').value = parseFloat(e.target.value).toFixed(2); });
 }
 
+// ===================== Explain context (LLM commit summary) =====================
+
+let explainEventSource = null;
+
+async function explainContext() {
+  if (explainEventSource) { explainEventSource.close(); explainEventSource = null; }
+  const section = el('#ub-explain-section');
+  const content = el('#ub-explain-content');
+  section.classList.remove('hidden');
+  section.open = true;
+  content.innerHTML = '<span class="muted">Asking the loaded model…</span>';
+  let text = '';
+  const es = new EventSource('/api/explain-updates');
+  explainEventSource = es;
+  let firstChunk = true;
+  es.onmessage = (ev) => {
+    try {
+      const d = JSON.parse(ev.data);
+      if (d.error) {
+        content.innerHTML = `<span class="muted">Couldn't get explanation: ${escapeHtml(d.error)}</span>`;
+        return;
+      }
+      if (firstChunk) { content.innerHTML = ''; firstChunk = false; }
+      if (d.content) {
+        text += d.content;
+        content.textContent = text;
+        content.scrollTop = content.scrollHeight;
+      } else if (d.reasoning) {
+        // Show reasoning tokens but de-emphasized
+        let r = content.querySelector('.reasoning-current');
+        if (!r) {
+          r = document.createElement('div');
+          r.className = 'reasoning reasoning-current';
+          r.textContent = '';
+          content.appendChild(r);
+        }
+        r.textContent += d.reasoning;
+      }
+    } catch {}
+  };
+  es.addEventListener('done', () => {
+    es.close();
+    explainEventSource = null;
+    // strip the reasoning-current marker
+    const r = content.querySelector('.reasoning-current');
+    if (r) r.classList.remove('reasoning-current');
+  });
+  es.onerror = () => { es.close(); explainEventSource = null; };
+}
+
 // ===================== updates (spark-vllm-docker) =====================
 
 const updState = {
@@ -817,19 +1042,23 @@ function renderUpdateBanner() {
   banner.classList.toggle('up-to-date', behind === 0 && !dirty);
   banner.classList.toggle('warn', !!dirty);
 
+  const explain = el('#ub-explain');
   if (dirty > 0) {
     text.textContent = `${dirty} local change${dirty === 1 ? '' : 's'} in ~/spark-vllm-docker. Resolve before updating.`;
     details.classList.add('hidden');
     apply.classList.add('hidden');
+    explain.classList.add('hidden');
   } else if (behind === 0) {
     text.textContent = `spark-vllm-docker is up to date (${info.current || ''})`;
     details.classList.add('hidden');
     apply.classList.add('hidden');
     list.classList.add('hidden');
+    explain.classList.add('hidden');
   } else {
     text.textContent = `${behind} commit${behind === 1 ? '' : 's'} behind upstream`;
     details.classList.remove('hidden');
     apply.classList.remove('hidden');
+    explain.classList.remove('hidden');
     log.textContent = (info.log || []).join('\n') || '(no log)';
   }
 }
@@ -918,13 +1147,26 @@ async function init() {
     list.open = !list.open;
   });
   el('#ub-apply').addEventListener('click', applyUpdate);
+  el('#ub-explain').addEventListener('click', explainContext);
+  el('#dl-repo').addEventListener('input', updateDlHfLink);
   setupCatalogDialog();
   setupAdvancedDialog();
+  // Open WebUI link from /api/config
+  try {
+    state.config = await fetchJSON('/api/config');
+    if (state.config.open_webui_url) {
+      const a = el('#open-webui-link');
+      a.href = state.config.open_webui_url;
+      a.classList.remove('hidden');
+    }
+  } catch {}
   await loadModels();
   await pollStatus();
   await renderServices();
+  pollHardware();
   pollUpdates();
   setInterval(pollStatus, 5000);
+  setInterval(pollHardware, 8000);    // every 8s
   setInterval(pollUpdates, 300000);  // every 5 min
 }
 

image/app/static/index.html

@@ -16,6 +16,7 @@
     <div class="current" id="current">
       <span class="muted">connecting…</span>
     </div>
+    <a id="open-webui-link" class="topbar-btn hidden" href="#" target="_blank" rel="noopener" title="Open Open WebUI">Open chat ↗</a>
   </header>
 
   <main>
@@ -24,6 +25,11 @@
       <span>Run the <em>Configure Sparks</em> action in StartOS to set hostnames, then run <em>Test Connection</em>.</span>
     </section>
 
+    <section id="hardware-panel" class="hardware-panel hidden">
+      <h2 class="section-title">Spark hardware</h2>
+      <div id="hardware-grid" class="hardware-grid"></div>
+    </section>
+
     <section id="endpoint-panel" class="endpoint-panel hidden">
       <div class="ep-title muted small">OpenAI-compatible endpoint</div>
       <div class="ep-row">
@@ -133,11 +139,20 @@
           <label class="dl-row">
             <span class="dl-label">HuggingFace repo</span>
             <input type="text" id="dl-repo" placeholder="e.g. RedHatAI/Qwen3.6-35B-A3B-NVFP4" autocomplete="off">
+            <a id="dl-hf-link" class="dl-hf-link hidden" href="#" target="_blank" rel="noopener" title="Open on Hugging Face">↗</a>
           </label>
+          <div class="dl-help muted small">
+            <a href="https://huggingface.co/models?other=vllm" target="_blank" rel="noopener">Browse vLLM-compatible models</a>
+            · NVFP4-quantized models (e.g. <code>RedHatAI/...</code>) are best for Blackwell hardware
+          </div>
           <div class="dl-row">
             <span class="dl-label">Where</span>
-            <label class="radio"><input type="radio" name="dl-mode" value="solo" checked> Spark 1 only (solo)</label>
+            <label class="radio"><input type="radio" name="dl-mode" value="spark1" checked> Spark 1 only</label>
-            <label class="radio"><input type="radio" name="dl-mode" value="cluster"> Both Sparks (cluster, copy in parallel)</label>
+            <label class="radio"><input type="radio" name="dl-mode" value="spark2"> Spark 2 only</label>
+            <label class="radio"><input type="radio" name="dl-mode" value="cluster"> Both Sparks (for cluster models)</label>
+          </div>
+          <div class="dl-help muted small">
+            For <strong>solo</strong> models, download to wherever you'll run them. For <strong>cluster</strong> models (-tp 2), both Sparks need the weights — "Both" downloads to one Spark and rsyncs to the other in parallel.
           </div>
           <div class="dl-actions">
             <button id="dl-cancel" class="btn">Cancel</button>
@@ -178,6 +193,7 @@
       <div class="ub-row">
         <span id="ub-text">Checking for updates…</span>
         <span class="spacer"></span>
+        <button id="ub-explain" class="btn small-btn hidden">✨ Explain context</button>
         <button id="ub-details" class="btn small-btn hidden">Show details</button>
         <button id="ub-apply" class="btn small-btn primary hidden">Apply update</button>
       </div>
@@ -185,6 +201,10 @@
         <summary class="muted small">Pending commits</summary>
         <pre id="ub-log" class="snippet"></pre>
       </details>
+      <details id="ub-explain-section" class="hidden">
+        <summary class="muted small">Explained by the loaded LLM</summary>
+        <div id="ub-explain-content" class="explain-content"></div>
+      </details>
       <div id="ub-progress" class="hidden">
         <div class="phase-row">
           <div class="phase" id="ub-phase">Applying update…</div>

image/app/static/style.css

@@ -45,6 +45,17 @@ body {
 .logo-dot { width: 10px; height: 10px; border-radius: 50%; background: var(--accent); box-shadow: 0 0 12px var(--accent); }
 .current { flex: 1; text-align: right; font-size: 14px; }
 .current strong { color: var(--accent); }
+.topbar-btn {
+  background: var(--surface-2);
+  border: 1px solid var(--border);
+  color: var(--text);
+  padding: 5px 10px;
+  border-radius: 6px;
+  font-size: 12px;
+  text-decoration: none;
+  transition: border-color 0.15s, background 0.15s;
+}
+.topbar-btn:hover { background: #24242c; border-color: var(--accent); color: var(--accent); }
 
 main {
   max-width: 880px;
@@ -268,6 +279,14 @@ main {
   font: 13px ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
 }
 .modal-row textarea { font-family: inherit; resize: vertical; }
+.modal-row .knob-hint {
+  color: var(--muted);
+  font-size: 11px;
+  line-height: 1.5;
+  margin-top: 2px;
+  padding-left: 2px;
+}
+.modal-row.inline .knob-hint { width: 100%; margin-left: 22px; margin-top: 0; }
 .modal-row input:focus, .modal-row textarea:focus, .modal-row select:focus { outline: 1px solid var(--info); border-color: var(--info); }
 .modal-row input[type='range'] { padding: 0; flex: 1; }
 .modal-fieldset {
@@ -295,6 +314,31 @@ main {
 .ub-context a { color: var(--info); text-decoration: none; }
 .ub-context a:hover { text-decoration: underline; }
 .ub-context em { font-style: normal; color: var(--text); font-weight: 500; }
+
+#ub-explain-section { margin-top: 8px; }
+#ub-explain-section summary { cursor: pointer; padding: 4px 0; }
+.explain-content {
+  background: #08080b;
+  border: 1px solid var(--border);
+  border-radius: 6px;
+  padding: 12px 14px;
+  margin-top: 8px;
+  font-size: 13px;
+  line-height: 1.6;
+  color: #c7c7d1;
+  white-space: pre-wrap;
+  word-break: break-word;
+  max-height: 320px;
+  overflow: auto;
+}
+.explain-content .reasoning {
+  color: var(--muted);
+  font-style: italic;
+  font-size: 11px;
+  border-left: 2px solid var(--border);
+  padding-left: 10px;
+  margin: 4px 0;
+}
 .update-banner.up-to-date {
   border-color: var(--border);
   color: var(--muted);
@@ -306,6 +350,51 @@ main {
 #ub-list summary { cursor: pointer; padding: 4px 0; }
 #ub-progress { margin-top: 10px; }
 
+/* ===== Hardware dashboard ===== */
+
+.hardware-grid {
+  display: grid;
+  gap: 14px;
+  grid-template-columns: repeat(auto-fill, minmax(320px, 1fr));
+}
+.hw-card {
+  background: var(--surface);
+  border: 1px solid var(--border);
+  border-radius: var(--radius);
+  padding: 14px 16px;
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+.hw-card .head {
+  display: flex;
+  align-items: baseline;
+  gap: 8px;
+  margin-bottom: 4px;
+}
+.hw-card .head .name { font-weight: 600; font-size: 15px; }
+.hw-card .head .meta { color: var(--muted); font-size: 12px; margin-left: auto; }
+.hw-card.unreachable { border-color: rgba(239, 68, 68, 0.4); }
+.hw-card.unreachable .name { color: var(--error); }
+.hw-metric { display: flex; align-items: center; gap: 10px; font-size: 12px; }
+.hw-metric .label { color: var(--muted); width: 56px; flex-shrink: 0; text-transform: uppercase; letter-spacing: 0.05em; font-size: 11px; }
+.hw-metric .bar { flex: 1; height: 8px; background: var(--surface-2); border-radius: 4px; overflow: hidden; position: relative; }
+.hw-metric .bar > span {
+  display: block;
+  height: 100%;
+  background: linear-gradient(90deg, var(--info), var(--accent));
+  border-radius: 4px;
+  transition: width 0.4s ease-out;
+}
+.hw-metric .bar.warn > span { background: linear-gradient(90deg, var(--warn), var(--error)); }
+.hw-metric .val {
+  font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+  font-size: 12px;
+  color: var(--text);
+  min-width: 110px;
+  text-align: right;
+}
+
 /* ===== Section header (title + action button) ===== */
 
 .section-header {
@@ -358,6 +447,24 @@ main {
   min-width: 200px;
 }
 .dl-row input[type='text']:focus { outline: 1px solid var(--info); border-color: var(--info); }
+.dl-hf-link {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  background: var(--surface-2);
+  border: 1px solid var(--border);
+  color: var(--info);
+  padding: 7px 10px;
+  border-radius: 6px;
+  text-decoration: none;
+  font-size: 14px;
+  flex-shrink: 0;
+}
+.dl-hf-link:hover { background: rgba(96, 165, 250, 0.08); border-color: var(--info); }
+.dl-help { padding-left: 122px; line-height: 1.6; }
+.dl-help a { color: var(--info); text-decoration: none; }
+.dl-help a:hover { text-decoration: underline; }
+.dl-help code { background: var(--surface-2); padding: 1px 5px; border-radius: 3px; font-size: 11px; }
 .radio { display: inline-flex; align-items: center; gap: 6px; font-size: 13px; color: var(--text); cursor: pointer; }
 .radio input { accent-color: var(--accent); }
 .dl-actions { display: flex; gap: 8px; justify-content: flex-end; margin-top: 10px; }
@@ -489,6 +596,9 @@ main {
   font-size: 11px;
   color: #5c5c66;
 }
+.card .repo a { color: inherit; text-decoration: none; }
+.card .repo a:hover { color: var(--info); text-decoration: underline; }
+.card .repo .hf-icon { font-size: 13px; opacity: 0.7; }
 .tag {
   background: var(--surface-2);
   border: 1px solid var(--border);

package/startos/actions/configureSparks.ts

@@ -76,6 +76,15 @@ const inputSpec = InputSpec.of({
     placeholder: 'magpie-tts',
     masked: false,
   }),
+  open_webui_url: Value.text({
+    name: 'Open WebUI URL (optional)',
+    description:
+      'If you also run Open WebUI on your LAN, paste its URL here. Spark Control will then show a one-click "Open chat" button next to the current model so you can jump straight to it.',
+    required: false,
+    default: null,
+    placeholder: 'e.g. https://open-webui.yourserver.local',
+    masked: false,
+  }),
 })
 
 export const configureSparks = sdk.Action.withInput(

package/startos/fileModels/sparkConfig.yaml.ts

@@ -14,6 +14,8 @@ export const sparkConfigSchema = z.object({
   magpie_host: z.string().catch(''),
   magpie_user: z.string().catch(''),
   magpie_container: z.string().catch(''),
+  // Optional Open WebUI deep-link
+  open_webui_url: z.string().catch(''),
 })
 
 export type SparkConfig = z.infer<typeof sparkConfigSchema>

package/startos/main.ts

@@ -19,6 +19,7 @@ export const main = sdk.setupMain(async ({ effects }) => {
     magpie_host: '',
     magpie_user: '',
     magpie_container: '',
+    open_webui_url: '',
   }
 
   return sdk.Daemons.of(effects).addDaemon('primary', {
@@ -47,6 +48,7 @@ export const main = sdk.setupMain(async ({ effects }) => {
         MAGPIE_USER: cfg.magpie_user,
         MAGPIE_CONTAINER: cfg.magpie_container,
         MODELS_OVERRIDES: '/data/models-overrides.yaml',
+        OPEN_WEBUI_URL: cfg.open_webui_url,
         BIND_PORT: String(uiPort),
       },
     },

package/startos/versions/v0_1_0.ts

@@ -1,10 +1,10 @@
 import { VersionInfo, IMPOSSIBLE } from '@start9labs/start-sdk'
 
 export const v0_1_0 = VersionInfo.of({
-  version: '0.2.4:0',
+  version: '0.3.0:0',
   releaseNotes: {
     en_US:
-      'Hotfix + UX polish: fixes parakeet/magpie status showing as "Unknown" (empty container-name env var no longer overrode the default). Copy buttons are now compact icons, and the values themselves are clickable to copy. Service cards show host + URL + model as separate copyable rows. The update banner now spells out what is being updated (spark-vllm-docker, the upstream LLM-cluster project).',
+      'v0.3: Spark hardware dashboard (RAM, disk, GPU memory + utilization, CPU load, uptime per Spark). Per-model Advanced settings now show plain-English hints tied to your actual GPU memory (e.g. "0.85 GPU util leaves ~18 GB free"). "Explain context" button on the update banner asks the loaded LLM to summarize pending commits in plain English. Optional Open WebUI URL in Configure Sparks shows a one-click "Open chat" button in the top bar. Downloads can now target Spark 1, Spark 2, or both. Each model card links out to its Hugging Face page.',
   },
   migrations: {
     up: async ({ effects }) => {},