v0.8.1:0 - delete model weights from disk via card trash icon

Each model card now shows whether its weights are present on disk
(with GB size) or not yet downloaded. When present and the model
isn't currently loaded, a trash icon appears; clicking it pops a
confirmation showing exactly how many GB will be freed and on
which Spark(s), then runs rm -rf on the HF cache directory via SSH.

Cluster-mode models are removed from both Sparks; solo-mode from
Spark 1 only. Safety rails: refuses to delete the currently-loaded
model, refuses during an in-flight swap or download, and the
catalog entry stays intact so it can be re-downloaded anytime.

Backend:
  - new image/app/disk.py: probe_disk + delete_from_disk over SSH
  - GET  /api/models/disk-status — parallel probe across all catalog models
  - DELETE /api/models/{key}/disk — guarded rm -rf, logs to connectivity events

Frontend:
  - on-disk / not-downloaded pills on every card
  - trash icon-btn in card-actions row (hidden when not on disk)
  - confirmation dialog showing per-host bytes-to-free
  - disk-status re-checked every 60s

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

README.md

@@ -84,6 +84,24 @@ Other services on your LAN can hit `GET /api/endpoints` to learn where the curre
 
 `base_url` is filled in whenever Configure Sparks has been completed (even if the underlying service isn't currently up). Pair the URL with `ready: true` to safely route traffic.
 
+## Reporting failures from external apps
+
+Spark Control polls every 5 s, so a brief blip in Parakeet/Magpie/vLLM availability can slip between polls and never make it into the connectivity log. To capture short failures, an external app (e.g. Open WebUI) can POST whenever a call fails (or succeeds):
+
+```bash
+curl -X POST http://<dashboard-url>/api/health-event \
+  -H 'content-type: application/json' \
+  -d '{
+    "service": "parakeet",
+    "ok": false,
+    "source": "open-webui",
+    "error": "HTTP 503",
+    "ms": 420
+  }'
+```
+
+Fields: `service` (required), `ok` (required), `source` (optional, free-form), `error` (optional), `ms` (optional latency). Each POST appends a `report` event to the connectivity log alongside the polling-based transition events.
+
 ## Status
 
 **v0.2.3** — installed and verified on a Start9 server. Five bundled LLMs in the catalog (qwen3-vl, gemma4, qwen36, qwen3-235b-fp8, qwen2.5-72b), plus any custom models added through the UI.

image/app/config.py

@@ -42,6 +42,8 @@ class Settings:
     parakeet_port: int
     magpie_port: int
     bind_port: int
+    open_webui_url: str
+    ngc_api_key: str
 
     @classmethod
     def from_env(cls) -> "Settings":
@@ -55,10 +57,10 @@ class Settings:
             spark2_user=spark2_user,
             parakeet_host=_env("PARAKEET_HOST") or spark2_host,
             parakeet_user=_env("PARAKEET_USER") or spark2_user,
-            parakeet_container=_env("PARAKEET_CONTAINER", "parakeet-asr"),
+            parakeet_container=_env("PARAKEET_CONTAINER") or "parakeet-asr",
             magpie_host=_env("MAGPIE_HOST") or spark2_host,
             magpie_user=_env("MAGPIE_USER") or spark2_user,
-            magpie_container=_env("MAGPIE_CONTAINER", "magpie-tts"),
+            magpie_container=_env("MAGPIE_CONTAINER") or "magpie-tts",
             ssh_key_path=_env("SSH_KEY_PATH"),
             ssh_known_hosts=_env("SSH_KNOWN_HOSTS"),
             models_yaml=_resolve_models_yaml(),
@@ -66,6 +68,8 @@ 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", ""),
+            ngc_api_key=_env("NGC_API_KEY", ""),
         )
 
     @property

image/app/connectivity.py

@@ -0,0 +1,190 @@
+"""Track up/down transitions for any subject (Sparks AND services) and cache MACs.
+
+Persisted to /data/connectivity.json. Schema:
+
+    {
+      "macs": { "spark1": "aa:bb:..", "spark2": "11:22:.." },
+      "current": { "spark1": "up", "parakeet": "up", "magpie": "down", ... },
+      "last_change": { ... },
+      "events": [
+        # Active-probe transition (logged when state flips during polling)
+        { "subject": "spark2", "at": "...", "kind": "transition",
+          "transition": "down" },
+        { "subject": "spark2", "at": "...", "kind": "transition",
+          "transition": "up", "down_seconds": 4500 },
+
+        # Passive report (logged whenever an external app POSTs to
+        # /api/health-event regardless of state change)
+        { "subject": "parakeet", "at": "...", "kind": "report",
+          "ok": false, "source": "open-webui",
+          "detail": "Connection refused", "latency_ms": 320 },
+      ]
+    }
+
+Legacy events from v0.5 with `spark` instead of `subject` and no `kind` field
+are read transparently as kind="transition".
+"""
+from __future__ import annotations
+import json
+import os
+import threading
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Optional
+
+
+MAX_EVENTS = 200  # rolling window — plenty for showing recent history
+
+
+def _path() -> str:
+    return os.environ.get("CONNECTIVITY_LOG", "/data/connectivity.json")
+
+
+_lock = threading.Lock()
+
+
+def _read() -> dict:
+    try:
+        with open(_path()) as f:
+            return json.load(f) or {}
+    except (FileNotFoundError, json.JSONDecodeError):
+        return {}
+
+
+def _write(data: dict) -> None:
+    p = _path()
+    Path(p).parent.mkdir(parents=True, exist_ok=True)
+    tmp = p + ".tmp"
+    with open(tmp, "w") as f:
+        json.dump(data, f, indent=2, sort_keys=False)
+    os.replace(tmp, p)
+
+
+def load() -> dict:
+    with _lock:
+        d = _read()
+        d.setdefault("macs", {})
+        d.setdefault("current", {})
+        d.setdefault("last_change", {})
+        d.setdefault("events", [])
+        return d
+
+
+def record_mac(subject: str, mac: Optional[str]) -> None:
+    if not mac:
+        return
+    with _lock:
+        d = _read()
+        d.setdefault("macs", {})
+        if d["macs"].get(subject) != mac:
+            d["macs"][subject] = mac
+            _write(d)
+
+
+def record_state(subject: str, reachable: bool) -> Optional[dict]:
+    """Update current state for `subject`. If it differs from the last seen
+    state, append a transition event. Returns the event dict if a transition
+    was recorded, else None.
+
+    `subject` can be a Spark host key (spark1/spark2) or a service name
+    (parakeet/magpie/vllm).
+    """
+    new_state = "up" if reachable else "down"
+    now = datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+    with _lock:
+        d = _read()
+        d.setdefault("macs", {})
+        d.setdefault("current", {})
+        d.setdefault("last_change", {})
+        d.setdefault("events", [])
+        prev = d["current"].get(subject)
+        if prev == new_state:
+            return None
+        event: dict = {
+            "subject": subject,
+            "at": now,
+            "kind": "transition",
+            "transition": new_state,
+        }
+        # When we have a previous state and timestamp, compute duration
+        last_change = d["last_change"].get(subject)
+        if prev and last_change:
+            try:
+                prev_dt = datetime.fromisoformat(last_change.replace("Z", "+00:00"))
+                duration = (datetime.now(timezone.utc) - prev_dt).total_seconds()
+                if prev == "down" and new_state == "up":
+                    event["down_seconds"] = round(duration)
+                if prev == "up" and new_state == "down":
+                    event["up_seconds"] = round(duration)
+            except ValueError:
+                pass
+        d["current"][subject] = new_state
+        d["last_change"][subject] = now
+        d["events"].append(event)
+        if len(d["events"]) > MAX_EVENTS:
+            d["events"] = d["events"][-MAX_EVENTS:]
+        _write(d)
+        return event
+
+
+def record_report(
+    subject: str,
+    *,
+    ok: bool,
+    source: str = "external",
+    detail: str = "",
+    latency_ms: Optional[int] = None,
+) -> dict:
+    """Record a passive report from an external caller (e.g. Open WebUI got a
+    503 calling Parakeet). Always appended to the events list; does NOT change
+    the active-probe state (which only the polling probe is authoritative on).
+    """
+    now = datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+    with _lock:
+        d = _read()
+        d.setdefault("events", [])
+        event: dict = {
+            "subject": subject,
+            "at": now,
+            "kind": "report",
+            "ok": bool(ok),
+            "source": source or "external",
+        }
+        if detail:
+            event["detail"] = detail
+        if latency_ms is not None:
+            event["latency_ms"] = int(latency_ms)
+        d["events"].append(event)
+        if len(d["events"]) > MAX_EVENTS:
+            d["events"] = d["events"][-MAX_EVENTS:]
+        _write(d)
+        return event
+
+
+def get_mac(subject: str) -> Optional[str]:
+    d = load()
+    return d.get("macs", {}).get(subject)
+
+
+def _normalize_event(e: dict) -> dict:
+    """Promote legacy v0.5 events to the v0.6 shape so the UI sees one schema."""
+    if "subject" in e:
+        e.setdefault("kind", "transition")
+        return e
+    # Legacy: had "spark" + "transition" only
+    if "spark" in e:
+        e["subject"] = e.pop("spark")
+        e.setdefault("kind", "transition")
+    return e
+
+
+def summary() -> dict:
+    """Compact summary for the UI: known MACs, current state, recent events."""
+    d = load()
+    events = [_normalize_event(dict(e)) for e in d.get("events", [])]
+    return {
+        "macs": d.get("macs", {}),
+        "current": d.get("current", {}),
+        "last_change": d.get("last_change", {}),
+        "events": events[-80:],
+    }

image/app/custom_services.py

@@ -0,0 +1,59 @@
+"""User-installed services persist in /data/services-overrides.yaml.
+
+Format:
+    custom:
+      - key: my-riva
+        kind: stt
+        host: <spark-2-ip>
+        user: <spark-user>
+        container: riva-asr
+        port: 8001
+        health_path: /health
+        image: nvcr.io/nim/nvidia/riva-multilingual:latest
+"""
+from __future__ import annotations
+import os
+from pathlib import Path
+import yaml
+
+
+def _path() -> str:
+    return os.environ.get("SERVICES_OVERRIDES", "/data/services-overrides.yaml")
+
+
+def load_custom_services() -> list[dict]:
+    try:
+        with open(_path()) as f:
+            data = yaml.safe_load(f) or {}
+    except FileNotFoundError:
+        return []
+    return data.get("custom") or []
+
+
+def add_custom_service(entry: dict) -> None:
+    p = _path()
+    Path(p).parent.mkdir(parents=True, exist_ok=True)
+    data: dict = {}
+    try:
+        with open(p) as f:
+            data = yaml.safe_load(f) or {}
+    except FileNotFoundError:
+        pass
+    custom = data.get("custom") or []
+    custom = [c for c in custom if c.get("key") != entry["key"]]
+    custom.append(entry)
+    data["custom"] = custom
+    with open(p, "w") as f:
+        yaml.safe_dump(data, f, sort_keys=False)
+
+
+def delete_custom_service(key: str) -> None:
+    p = _path()
+    try:
+        with open(p) as f:
+            data = yaml.safe_load(f) or {}
+    except FileNotFoundError:
+        return
+    data["custom"] = [c for c in (data.get("custom") or []) if c.get("key") != key]
+    with open(p, "w") as f:
+        yaml.safe_dump(data, f, sort_keys=False)

image/app/deep_health.py

@@ -0,0 +1,363 @@
+"""Deep health probes for each service.
+
+Why this exists: Triton's /health endpoint returns 200 as long as the HTTP
+layer is alive and the model is registered. It does NOT verify that the CUDA
+context inside the worker process is healthy. We've observed Parakeet getting
+its CUDA context wedged after an OOM, where /health stays green but every
+real transcription returns 500 cudaErrorUnknown.
+
+So this module sends *real* but tiny synthetic inference requests:
+  - Parakeet: 1 second of digital silence (16 kHz mono PCM, in-memory WAV)
+  - Magpie:   short text-to-speech, response audio discarded
+  - vLLM:     1-token chat completion against whatever model is loaded
+
+All synthetic payloads are generated on demand into BytesIO, sent over HTTP,
+and never touched the filesystem (on either spark-control's side or the
+target service's side beyond normal Triton/Riva working memory).
+
+When a probe fails with a signal that looks like a CUDA wedge, we
+automatically issue `docker restart <container>`. Rate-limited to 3 restarts
+per service per 30 minutes to avoid restart loops.
+"""
+from __future__ import annotations
+import asyncio
+import io
+import time
+import wave
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Optional
+
+import httpx
+
+from .config import Settings
+from .connectivity import record_report
+from .services import ServiceDef, run_action, services_from_settings
+
+
+# Default 5-minute interval, controllable via env. Sub-minute is silly for a
+# heavy synthetic probe; we just want to catch wedges within a reasonable
+# window — much faster than the user noticing on their next real call.
+DEFAULT_INTERVAL_SEC = 300.0
+PROBE_TIMEOUT_SEC = 20.0
+RESTART_RATE_LIMIT = 3            # max auto-restarts per service
+RESTART_RATE_WINDOW_SEC = 1800.0  # within a 30-min window
+RESTART_COOLDOWN_SEC = 120.0      # don't restart again within this many seconds of the last one
+STARTUP_GRACE_SEC = 60.0          # don't auto-restart for the first minute after this app boots
+
+
+def _silence_wav(seconds: float = 1.0, sample_rate: int = 16000) -> io.BytesIO:
+    """Return an in-memory WAV file containing `seconds` of digital silence."""
+    n_frames = int(seconds * sample_rate)
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as w:
+        w.setnchannels(1)
+        w.setsampwidth(2)           # int16
+        w.setframerate(sample_rate)
+        w.writeframes(b"\x00\x00" * n_frames)
+    buf.seek(0)
+    return buf
+
+
+def _looks_like_wedge(error: str) -> bool:
+    """Heuristic: does this error string look like a stuck CUDA context that
+    a container restart would clear? We want to be conservative — only act
+    on signals we're confident about, otherwise leave the user in charge."""
+    err = (error or "").lower()
+    needles = [
+        "cudaerrorunknown",
+        "cuda error: unknown",
+        "cuda kernel errors",
+        "internal server error",
+        "engine core initialization failed",
+        "503",   # service unavailable from a dependency
+        "500",   # generic 5xx with a body that may not parse
+    ]
+    return any(n in err for n in needles)
+
+
+@dataclass
+class ProbeResult:
+    ok: bool
+    at: str
+    latency_ms: Optional[int] = None
+    error: str = ""
+    note: str = ""
+
+
+@dataclass
+class ServiceState:
+    last: Optional[ProbeResult] = None
+    last_ok_at: Optional[str] = None
+    restarts: list[float] = field(default_factory=list)
+
+
+class DeepHealth:
+    def __init__(self, settings: Settings, interval_sec: float = DEFAULT_INTERVAL_SEC) -> None:
+        self.settings = settings
+        self.interval_sec = interval_sec
+        self.state: dict[str, ServiceState] = {
+            "parakeet": ServiceState(),
+            "magpie": ServiceState(),
+            "vllm": ServiceState(),
+        }
+        self._stop = asyncio.Event()
+        self._boot_at = time.monotonic()
+
+    # ---- probes ---------------------------------------------------------
+
+    async def probe_parakeet(self) -> ProbeResult:
+        s = self.settings
+        now_iso = datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+        if not s.parakeet_host:
+            return ProbeResult(ok=False, at=now_iso, error="not configured")
+        url = f"http://{s.parakeet_host}:{s.parakeet_port}/v1/audio/transcriptions"
+        wav = _silence_wav(1.0)
+        t0 = time.monotonic()
+        try:
+            async with httpx.AsyncClient(timeout=PROBE_TIMEOUT_SEC) as c:
+                r = await c.post(
+                    url,
+                    files={"file": ("probe.wav", wav, "audio/wav")},
+                    data={"model": "parakeet-tdt-0.6b-v3"},
+                )
+            latency = round((time.monotonic() - t0) * 1000)
+            if 200 <= r.status_code < 300:
+                return ProbeResult(ok=True, at=now_iso, latency_ms=latency)
+            return ProbeResult(
+                ok=False,
+                at=now_iso,
+                latency_ms=latency,
+                error=f"HTTP {r.status_code}: {r.text[:240]}",
+            )
+        except Exception as e:
+            return ProbeResult(ok=False, at=now_iso, error=f"{type(e).__name__}: {e}")
+
+    async def probe_magpie(self) -> ProbeResult:
+        s = self.settings
+        now_iso = datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+        if not s.magpie_host:
+            return ProbeResult(ok=False, at=now_iso, error="not configured")
+        # Magpie /v1/audio/synthesize expects multipart form-data, not JSON.
+        # The (None, value) tuple in httpx's `files=` produces a non-file form field.
+        url = f"http://{s.magpie_host}:{s.magpie_port}/v1/audio/synthesize"
+        form: dict = {"text": (None, "hi"), "language": (None, "en-US")}
+        t0 = time.monotonic()
+        try:
+            async with httpx.AsyncClient(timeout=PROBE_TIMEOUT_SEC) as c:
+                r = await c.post(url, files=form)
+            latency = round((time.monotonic() - t0) * 1000)
+            if 200 <= r.status_code < 300:
+                return ProbeResult(ok=True, at=now_iso, latency_ms=latency)
+            # 4xx that aren't 5xx mean server is alive but our payload is off —
+            # don't classify as wedge.
+            if 400 <= r.status_code < 500:
+                return ProbeResult(
+                    ok=True,
+                    at=now_iso,
+                    latency_ms=latency,
+                    note=f"{r.status_code} — server alive (probe payload may need a voice name)",
+                )
+            return ProbeResult(
+                ok=False,
+                at=now_iso,
+                latency_ms=latency,
+                error=f"HTTP {r.status_code}: {r.text[:240]}",
+            )
+        except Exception as e:
+            return ProbeResult(ok=False, at=now_iso, error=f"{type(e).__name__}: {e}")
+
+    async def probe_vllm(self) -> ProbeResult:
+        s = self.settings
+        now_iso = datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+        if not s.spark1_host:
+            return ProbeResult(ok=False, at=now_iso, error="not configured")
+        base = f"http://{s.spark1_host}:{s.vllm_port}"
+        # Step 1: is there a model loaded?
+        try:
+            async with httpx.AsyncClient(timeout=5.0) as c:
+                r = await c.get(f"{base}/v1/models")
+            if 200 <= r.status_code < 300:
+                models = r.json().get("data") or []
+            else:
+                # 5xx on /v1/models suggests something wedged after a model loaded
+                return ProbeResult(
+                    ok=False,
+                    at=now_iso,
+                    error=f"list_models HTTP {r.status_code}: {r.text[:240]}",
+                )
+        except Exception:
+            # Connection refused / timeout: usually means no vLLM process listening
+            # (the vllm_node container is alive but no `vllm serve` is running yet).
+            # That's an idle state, not a wedge — don't trigger auto-restart.
+            return ProbeResult(
+                ok=True,
+                at=now_iso,
+                note="no model currently loaded (idle)",
+            )
+
+        if not models:
+            return ProbeResult(
+                ok=True,
+                at=now_iso,
+                note="no model currently loaded (idle)",
+            )
+
+        model_id = models[0]["id"]
+        # Step 2: model is loaded; verify it can actually complete a 1-token request.
+        t0 = time.monotonic()
+        try:
+            async with httpx.AsyncClient(timeout=PROBE_TIMEOUT_SEC) as c:
+                r = await c.post(
+                    f"{base}/v1/chat/completions",
+                    json={
+                        "model": model_id,
+                        "messages": [{"role": "user", "content": "hi"}],
+                        "max_tokens": 1,
+                        "temperature": 0,
+                    },
+                )
+            latency = round((time.monotonic() - t0) * 1000)
+            if 200 <= r.status_code < 300:
+                return ProbeResult(ok=True, at=now_iso, latency_ms=latency, note=f"model={model_id}")
+            return ProbeResult(
+                ok=False,
+                at=now_iso,
+                latency_ms=latency,
+                error=f"HTTP {r.status_code}: {r.text[:240]}",
+            )
+        except Exception as e:
+            return ProbeResult(ok=False, at=now_iso, error=f"{type(e).__name__}: {e}")
+
+    # ---- orchestration --------------------------------------------------
+
+    PROBES = {
+        "parakeet": "probe_parakeet",
+        "magpie": "probe_magpie",
+        "vllm": "probe_vllm",
+    }
+
+    async def run_one(self, service: str) -> ProbeResult:
+        fn = getattr(self, self.PROBES[service])
+        result: ProbeResult = await fn()
+        st = self.state[service]
+        prev_ok = st.last.ok if st.last else None
+        st.last = result
+        if result.ok:
+            st.last_ok_at = result.at
+
+        # Log to connectivity history: every failure, plus the first success
+        # after a failure (recovery), plus the first probe ever — but skip
+        # the "still ok" steady-state to keep the log readable.
+        if not result.ok:
+            record_report(
+                service,
+                ok=False,
+                source="deep-health",
+                detail=result.error[:240],
+                latency_ms=result.latency_ms,
+            )
+        elif prev_ok is False:
+            record_report(
+                service,
+                ok=True,
+                source="deep-health",
+                detail="recovered" + (f" — {result.note}" if result.note else ""),
+                latency_ms=result.latency_ms,
+            )
+        elif prev_ok is None:
+            record_report(
+                service,
+                ok=True,
+                source="deep-health",
+                detail="first probe ok" + (f" — {result.note}" if result.note else ""),
+                latency_ms=result.latency_ms,
+            )
+
+        # Maybe auto-restart
+        if not result.ok and _looks_like_wedge(result.error):
+            await self._maybe_restart(service, result.error)
+        return result
+
+    async def _maybe_restart(self, service: str, error: str) -> None:
+        # No restarts during the boot grace period.
+        if time.monotonic() - self._boot_at < STARTUP_GRACE_SEC:
+            return
+        st = self.state[service]
+        now = time.monotonic()
+        st.restarts = [t for t in st.restarts if now - t < RESTART_RATE_WINDOW_SEC]
+        if st.restarts and now - st.restarts[-1] < RESTART_COOLDOWN_SEC:
+            return  # already restarted recently, give it time
+        if len(st.restarts) >= RESTART_RATE_LIMIT:
+            record_report(
+                service,
+                ok=False,
+                source="deep-health",
+                detail=f"rate-limited; not auto-restarting (would be #{len(st.restarts)+1} in 30 min)",
+            )
+            return
+        services = services_from_settings(self.settings)
+        if service not in services:
+            return
+        svc = services[service]
+        if not svc.host or not svc.user:
+            return
+        result = await run_action(self.settings, svc, "restart")
+        st.restarts.append(now)
+        ok = result.get("ok", False)
+        record_report(
+            service,
+            ok=False,
+            source="deep-health",
+            detail=f"auto-restart triggered (wedge: {error[:120]}); restart {'OK' if ok else 'FAILED'}",
+        )
+
+    async def run_all(self) -> dict[str, ProbeResult]:
+        results = {}
+        for name in self.PROBES:
+            results[name] = await self.run_one(name)
+        return results
+
+    async def run_periodic(self) -> None:
+        """Long-running loop. Cancel via .stop()."""
+        # Brief initial wait to let app finish startup
+        try:
+            await asyncio.wait_for(self._stop.wait(), timeout=10.0)
+            return
+        except asyncio.TimeoutError:
+            pass
+        while not self._stop.is_set():
+            try:
+                await self.run_all()
+            except Exception:
+                # Never let the loop die; the periodic check is best-effort
+                pass
+            try:
+                await asyncio.wait_for(self._stop.wait(), timeout=self.interval_sec)
+                return
+            except asyncio.TimeoutError:
+                continue
+
+    def stop(self) -> None:
+        self._stop.set()
+
+    def summary(self) -> dict:
+        out = {}
+        for name, st in self.state.items():
+            last = st.last
+            out[name] = {
+                "last_ok_at": st.last_ok_at,
+                "last": (
+                    {
+                        "ok": last.ok,
+                        "at": last.at,
+                        "latency_ms": last.latency_ms,
+                        "error": last.error,
+                        "note": last.note,
+                    }
+                    if last
+                    else None
+                ),
+                "auto_restarts_window": len(st.restarts),
+            }
+        return out

image/app/disk.py

@@ -0,0 +1,130 @@
+"""On-disk presence + deletion for Hugging Face model caches on the Sparks.
+
+The HF cache layout for a repo `org/name` is:
+
+    ~/.cache/huggingface/hub/models--org--name/
+
+We use `du -sb` to measure size (bytes) and `rm -rf` to free it. All operations
+are gated by the server endpoints, which refuse to delete a currently-loaded
+model or one tied to an in-flight swap/download.
+"""
+from __future__ import annotations
+import asyncio
+import shlex
+from dataclasses import dataclass
+from typing import Optional
+
+from .config import Settings
+from .ssh import ssh_run
+
+
+def repo_to_cache_dirname(repo: str) -> str:
+    """Convert 'org/name' to 'models--org--name' (the HF hub cache directory)."""
+    if "/" not in repo:
+        raise ValueError(f"repo must be in 'org/name' form: {repo!r}")
+    return "models--" + repo.replace("/", "--")
+
+
+def _cache_path(repo: str) -> str:
+    """Full remote path to the model's cache directory."""
+    # Use $HOME so it resolves correctly regardless of the SSH user's home.
+    return f"$HOME/.cache/huggingface/hub/{repo_to_cache_dirname(repo)}"
+
+
+@dataclass
+class HostDiskResult:
+    host: str
+    on_disk: bool
+    size_bytes: int = 0
+    error: Optional[str] = None
+
+
+@dataclass
+class DiskStatus:
+    repo: str
+    on_disk: bool                     # True if present on AT LEAST one host
+    total_bytes: int                  # sum across hosts
+    per_host: list[HostDiskResult]
+
+
+async def probe_host(host: str, user: str, repo: str, settings: Settings) -> HostDiskResult:
+    """Return whether the model's cache dir exists on this host and its size."""
+    if not host or not user:
+        return HostDiskResult(host=host or "?", on_disk=False, error="host not configured")
+    path = _cache_path(repo)
+    # `du -sb` prints bytes; if the dir doesn't exist, `du` returns non-zero.
+    # We test existence explicitly first so we can report on_disk=False cleanly.
+    cmd = (
+        f"if [ -d {shlex.quote(path)} ]; then "
+        f"du -sb {shlex.quote(path)} 2>/dev/null | awk '{{print $1}}'; "
+        f"else echo MISSING; fi"
+    )
+    rc, out, err = await ssh_run(host, user, cmd, settings, timeout=20.0)
+    if rc != 0:
+        return HostDiskResult(host=host, on_disk=False, error=(err or out).strip() or f"rc={rc}")
+    raw = out.strip()
+    if raw == "MISSING" or raw == "":
+        return HostDiskResult(host=host, on_disk=False)
+    try:
+        size = int(raw.splitlines()[-1])
+    except ValueError:
+        return HostDiskResult(host=host, on_disk=False, error=f"unparsable du output: {raw!r}")
+    return HostDiskResult(host=host, on_disk=True, size_bytes=size)
+
+
+async def probe_disk(repo: str, mode: str, settings: Settings) -> DiskStatus:
+    """Probe one model across the relevant Sparks based on its mode (solo|cluster)."""
+    hosts: list[tuple[str, str]] = [(settings.spark1_host, settings.spark1_user)]
+    if mode == "cluster" and settings.spark2_host:
+        hosts.append((settings.spark2_host, settings.spark2_user))
+
+    results = await asyncio.gather(*(probe_host(h, u, repo, settings) for h, u in hosts))
+    on_disk = any(r.on_disk for r in results)
+    total = sum(r.size_bytes for r in results)
+    return DiskStatus(repo=repo, on_disk=on_disk, total_bytes=total, per_host=list(results))
+
+
+async def delete_host(host: str, user: str, repo: str, settings: Settings) -> HostDiskResult:
+    """Probe + rm -rf on one host. Returns bytes freed (0 if the dir wasn't there)."""
+    if not host or not user:
+        return HostDiskResult(host=host or "?", on_disk=False, error="host not configured")
+    path = _cache_path(repo)
+    # Safety: hard-code the prefix in the command so a bad `repo` can never escape.
+    # Compute size first, then remove. If absent, still return success (idempotent).
+    cmd = (
+        f"set -e; "
+        f"P={shlex.quote(path)}; "
+        f"if [ -d \"$P\" ]; then "
+        f"  SIZE=$(du -sb \"$P\" 2>/dev/null | awk '{{print $1}}'); "
+        f"  rm -rf -- \"$P\"; "
+        f"  echo FREED $SIZE; "
+        f"else "
+        f"  echo FREED 0; "
+        f"fi"
+    )
+    rc, out, err = await ssh_run(host, user, cmd, settings, timeout=120.0)
+    if rc != 0:
+        return HostDiskResult(host=host, on_disk=False, error=(err or out).strip() or f"rc={rc}")
+    # Parse the "FREED N" line
+    freed = 0
+    for line in out.splitlines():
+        parts = line.strip().split()
+        if len(parts) == 2 and parts[0] == "FREED":
+            try:
+                freed = int(parts[1])
+            except ValueError:
+                pass
+            break
+    return HostDiskResult(host=host, on_disk=False, size_bytes=freed)
+
+
+async def delete_from_disk(repo: str, mode: str, settings: Settings) -> DiskStatus:
+    """rm -rf the model's cache dir on the relevant Sparks. Idempotent."""
+    hosts: list[tuple[str, str]] = [(settings.spark1_host, settings.spark1_user)]
+    if mode == "cluster" and settings.spark2_host:
+        hosts.append((settings.spark2_host, settings.spark2_user))
+
+    results = await asyncio.gather(*(delete_host(h, u, repo, settings) for h, u in hosts))
+    total_freed = sum(r.size_bytes for r in results)
+    # After deletion, on_disk should be False on all hosts.
+    return DiskStatus(repo=repo, on_disk=False, total_bytes=total_freed, per_host=list(results))

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:
-            raise RuntimeError("spark1 not configured")
+        # Pick the SSH target and hf-download flags from the mode.
+        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 +1,137 @@
+"""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 .connectivity import record_mac, record_state
+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+=$1} END {print s+0}')
+DEFIF=$(ip route show default 2>/dev/null | awk '{print $5; exit}')
+echo MAC=$(cat /sys/class/net/$DEFIF/address 2>/dev/null)
+""".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"])
+    # MAC address on the default-route interface (for Wake-on-LAN)
+    if info.get("mac"):
+        parsed["mac"] = info["mac"].lower()
+    return parsed
+
+
+class HardwareProbe:
+    """Caches results briefly to avoid hammering the Sparks."""
+
+    def __init__(self, settings: Settings, ttl_sec: float = 4.0, fail_ttl_sec: float = 25.0) -> None:
+        self.settings = settings
+        self.ttl_sec = ttl_sec
+        self.fail_ttl_sec = fail_ttl_sec
+        self._cache: dict[str, tuple[float, dict]] = {}
+        self._locks: dict[str, asyncio.Lock] = {}
+
+    def _ttl_for(self, value: dict) -> float:
+        return self.ttl_sec if value.get("reachable") else self.fail_ttl_sec
+
+    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:
+        s1, s2 = await asyncio.gather(
+            self._one("spark1", self.settings.spark1_host, self.settings.spark1_user),
+            self._one("spark2", self.settings.spark2_host, self.settings.spark2_user),
+        )
+        return {"spark1": s1, "spark2": s2}
+
+    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_for(cached[1])):
+                return cached[1]
+            # Use a shorter timeout for the connect phase; if a previous probe
+            # marked this host unreachable, return the cached failure immediately.
+            rc, out, err = await ssh_run(host, user, _PROBE, self.settings, timeout=6)
+            if rc != 0:
+                result = {"reachable": False, "configured": True, "host": host, "error": err.strip() or out.strip() or f"rc={rc}"}
+                self._cache[key] = (now, result)
+                record_state(key, False)
+                return result
+            parsed = _parse(out)
+            result = {"reachable": True, "configured": True, "host": host, **parsed}
+            self._cache[key] = (now, result)
+            record_state(key, True)
+            if parsed.get("mac"):
+                record_mac(key, parsed["mac"])
+            return result

image/app/nim.py

@@ -0,0 +1,202 @@
+"""NVIDIA NIM container install / lifecycle.
+
+Two pieces:
+  * A small curated catalog of NIM images (so users don't have to copy/paste
+    huge nvcr.io URLs).
+  * An installer that SSHes into the target Spark, runs `docker pull` then
+    `docker run -d --gpus all -p PORT:PORT -v VOLUME:/opt/nim/.cache
+    -e NGC_API_KEY=... IMAGE` and streams output.
+
+Custom services also persist via `overrides.add_custom_service()` so the
+Services panel can show them.
+"""
+from __future__ import annotations
+import asyncio
+import uuid
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Optional
+
+from .config import Settings
+from .ssh import ssh_stream, StreamHandle
+
+
+# Curated list. These are the most useful NIM containers for a dual-Spark
+# audio-and-LLM setup. Browse the full catalog at
+# https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia
+CATALOG_URL = "https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers"
+
+
+SUGGESTED_NIMS: list[dict] = [
+    {
+        "key": "parakeet-tdt-0.6b-v3",
+        "name": "Parakeet TDT 0.6B v3",
+        "image": "nvcr.io/nim/nvidia/parakeet-tdt-0-6b-v3:latest",
+        "default_container": "parakeet-asr",
+        "default_port": 8000,
+        "kind": "stt",
+        "description": "Streaming speech-to-text (English). Used by Open WebUI for voice input. ~1 GB.",
+        "homepage": "https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/parakeet-tdt-0-6b-v3",
+    },
+    {
+        "key": "magpie-tts-multilingual",
+        "name": "Magpie TTS Multilingual",
+        "image": "nvcr.io/nim/nvidia/magpie-tts-multilingual:latest",
+        "default_container": "magpie-tts",
+        "default_port": 9000,
+        "kind": "tts",
+        "description": "Multilingual text-to-speech. Counterpart to Parakeet for 'read aloud'. ~3 GB.",
+        "homepage": "https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/magpie-tts-multilingual",
+    },
+    {
+        "key": "riva-multilingual",
+        "name": "Riva Multilingual ASR",
+        "image": "nvcr.io/nim/nvidia/riva-multilingual:latest",
+        "default_container": "riva-asr",
+        "default_port": 8001,
+        "kind": "stt",
+        "description": "NVIDIA Riva speech-recognition multi-language model. Larger and more accurate than Parakeet.",
+        "homepage": "https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia",
+    },
+]
+
+
+@dataclass
+class NimInstallJob:
+    id: str
+    image: str
+    container: str
+    port: int
+    host: str
+    user: str
+    volume: Optional[str]
+    started_at: str
+    state: str = "starting"   # starting | pulling | running | done | failed
+    phase: str = "Starting…"
+    lines: list[str] = field(default_factory=list)
+    returncode: Optional[int] = None
+    finished_at: Optional[str] = None
+
+    def append(self, line: str) -> None:
+        self.lines.append(line)
+        if len(self.lines) > 1000:
+            del self.lines[: len(self.lines) - 1000]
+
+
+class NimManager:
+    def __init__(self, settings: Settings) -> None:
+        self.settings = settings
+        self.lock = asyncio.Lock()
+        self.jobs: dict[str, NimInstallJob] = {}
+        self.current_job_id: Optional[str] = None
+
+    def get(self, job_id: str) -> NimInstallJob | None:
+        return self.jobs.get(job_id)
+
+    async def trigger(
+        self,
+        *,
+        image: str,
+        container: str,
+        port: int,
+        host: str,
+        user: str,
+        volume: str | None = None,
+        extra_env: dict[str, str] | None = None,
+    ) -> NimInstallJob:
+        if self.lock.locked():
+            raise RuntimeError("Another NIM install is already in progress")
+        if not host or not user:
+            raise RuntimeError("target host not configured")
+        if not self.settings.ngc_api_key:
+            raise RuntimeError(
+                "NGC_API_KEY is not set. Open Configure Sparks in StartOS and paste your NGC personal API key (free at https://ngc.nvidia.com/setup/personal-key)."
+            )
+
+        job = NimInstallJob(
+            id=uuid.uuid4().hex[:8],
+            image=image,
+            container=container,
+            port=port,
+            host=host,
+            user=user,
+            volume=volume or f"{container}-cache",
+            started_at=datetime.now(timezone.utc).isoformat(),
+        )
+        self.jobs[job.id] = job
+        self.current_job_id = job.id
+        asyncio.create_task(self._run(job, extra_env or {}))
+        return job
+
+    async def _run(self, job: NimInstallJob, extra_env: dict[str, str]) -> None:
+        async with self.lock:
+            try:
+                await self._do(job, extra_env)
+                if job.state != "failed":
+                    job.state = "done"
+                    job.returncode = 0
+                    job.phase = "Done"
+            except Exception as e:
+                job.append(f"[error] {type(e).__name__}: {e}")
+                job.state = "failed"
+                if job.returncode is None:
+                    job.returncode = 1
+            finally:
+                job.finished_at = datetime.now(timezone.utc).isoformat()
+                if self.current_job_id == job.id:
+                    self.current_job_id = None
+
+    async def _do(self, job: NimInstallJob, extra_env: dict[str, str]) -> None:
+        # Build the bash one-liner. We use docker login non-interactively with the NGC API key.
+        env_parts = [f'-e NGC_API_KEY=$NGC_API_KEY']
+        for k, v in extra_env.items():
+            env_parts.append(f"-e {k}={v}")
+        env_str = " ".join(env_parts)
+        cmd = (
+            f"set -e; "
+            f"export NGC_API_KEY='{self.settings.ngc_api_key}'; "
+            f"echo '=== docker login nvcr.io ==='; "
+            f"echo \"$NGC_API_KEY\" | docker login nvcr.io -u '$oauthtoken' --password-stdin; "
+            f"echo '=== docker pull {job.image} (this can be 1-10 GB) ==='; "
+            f"docker pull {job.image}; "
+            f"echo '=== remove any prior container with the same name ==='; "
+            f"docker rm -f {job.container} 2>/dev/null || true; "
+            f"echo '=== docker run -d --gpus all -p {job.port}:{job.port} -v {job.volume}:/opt/nim/.cache {env_str} --name {job.container} --restart unless-stopped {job.image} ==='; "
+            f"docker run -d --gpus all "
+            f"-p {job.port}:{job.port} "
+            f"-v {job.volume}:/opt/nim/.cache "
+            f"{env_str} "
+            f"--name {job.container} "
+            f"--restart unless-stopped "
+            f"{job.image}; "
+            f"echo '=== ensuring cache volume is writable by uid 1000 (riva-server) ==='; "
+            f"docker run --rm -v {job.volume}:/cache alpine chown -R 1000:1000 /cache && "
+            f"docker restart {job.container}; "
+            f"echo '=== install complete; container is starting up and will download its model on first boot ==='"
+        )
+        job.append(f"$ <install command for {job.image} on {job.host}>")
+        job.state = "pulling"
+        job.phase = "Pulling image from nvcr.io (this can take a few minutes)…"
+
+        handle = StreamHandle()
+        async for line in ssh_stream(job.host, job.user, cmd, self.settings, handle=handle):
+            # Don't log lines containing the api key
+            if self.settings.ngc_api_key and self.settings.ngc_api_key in line:
+                continue
+            job.append(line)
+            if "docker pull" in line:
+                job.phase = "Pulling image from nvcr.io…"
+            elif "Login Succeeded" in line:
+                job.phase = "Logged in to NGC; pulling image…"
+            elif "Pull complete" in line:
+                job.phase = "Pulling layers…"
+            elif "Status: Downloaded newer image" in line or "Image is up to date" in line:
+                job.phase = "Image ready; starting container…"
+            elif "docker run -d" in line:
+                job.state = "running"
+                job.phase = "Container starting; downloading model on first boot…"
+
+        rc = handle.returncode or 0
+        if rc != 0:
+            job.state = "failed"
+            job.returncode = rc

image/app/server.py

@@ -10,14 +10,22 @@ from pydantic import BaseModel
 from typing import Literal
 
 from .config import Settings
+from .connectivity import get_mac, record_report, record_state, summary as connectivity_summary
+from .custom_services import add_custom_service, delete_custom_service
+from .deep_health import DeepHealth
+from .disk import delete_from_disk, probe_disk
 from .download import DownloadManager
+from .hardware import HardwareProbe
 from .health import check_magpie, check_parakeet, check_vllm
 from .models import load_catalog
+from .nim import SUGGESTED_NIMS, CATALOG_URL, NimManager
 from .overrides import add_custom, delete_custom, extract_knobs_from_args, load_overrides, set_knobs
 from .services import docker_state, run_action, services_from_settings
 from .ssh import ssh_run
 from .swap import SwapManager
 from .updates import UpdateManager, get_update_status
+from .validate import validate_launch
+from .wol import send_local_broadcast, send_via_peer
 
 
 settings = Settings.from_env()
@@ -25,9 +33,24 @@ catalog = load_catalog(settings.models_yaml)
 swap_manager = SwapManager(settings, catalog)
 download_manager = DownloadManager(settings)
 update_manager = UpdateManager(settings)
+hardware_probe = HardwareProbe(settings)
+nim_manager = NimManager(settings)
+deep_health = DeepHealth(settings)
 
 app = FastAPI(title="spark-control", version="0.1.0")
 
+
+@app.on_event("startup")
+async def _start_deep_health() -> None:
+    # Fire-and-forget; the loop catches its own exceptions.
+    asyncio.create_task(deep_health.run_periodic())
+
+
+@app.on_event("shutdown")
+async def _stop_deep_health() -> None:
+    deep_health.stop()
+
+
 _STATIC_DIR = Path(__file__).resolve().parent / "static"
 app.mount("/static", StaticFiles(directory=_STATIC_DIR), name="static")
 
@@ -44,6 +67,7 @@ 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 +140,191 @@ async def del_model(key: str) -> dict:
     return {"ok": True, "key": key}
 
 
+@app.get("/api/models/disk-status")
+async def get_models_disk_status() -> dict:
+    """Probe each catalog model's HF cache on the appropriate Spark(s) in parallel.
+
+    Result is keyed by model key: {on_disk, total_bytes, per_host:[{host,on_disk,size_bytes,error?}]}.
+    Designed to be called once on dashboard load; takes ~1–3s depending on Spark count.
+    """
+    if not settings.configured:
+        return {"configured": False, "models": {}}
+    keys = list(catalog.models.keys())
+    statuses = await asyncio.gather(*(
+        probe_disk(catalog.models[k].repo, catalog.models[k].mode, settings) for k in keys
+    ), return_exceptions=True)
+    out: dict[str, dict] = {}
+    for k, s in zip(keys, statuses):
+        if isinstance(s, Exception):
+            out[k] = {"on_disk": False, "total_bytes": 0, "per_host": [], "error": str(s)}
+            continue
+        out[k] = {
+            "on_disk": s.on_disk,
+            "total_bytes": s.total_bytes,
+            "per_host": [
+                {"host": r.host, "on_disk": r.on_disk, "size_bytes": r.size_bytes, **({"error": r.error} if r.error else {})}
+                for r in s.per_host
+            ],
+        }
+    return {"configured": True, "models": out}
+
+
+@app.delete("/api/models/{key}/disk")
+async def del_model_disk(key: str) -> dict:
+    """Delete a model's weights from the Spark filesystem(s). The catalog entry stays.
+
+    Safety rails:
+      - Refuses if the model is currently loaded on vLLM.
+      - Refuses if a swap or download is in flight.
+      - Idempotent: if the cache dir is already gone on a host, that host reports 0 bytes freed.
+    """
+    if key not in catalog.models:
+        raise HTTPException(404, f"unknown model: {key}")
+    m = catalog.models[key]
+
+    # Refuse if currently loaded
+    try:
+        vllm = await check_vllm(settings)
+    except Exception:
+        vllm = {}
+    if vllm.get("ok") and vllm.get("current_model") == m.repo:
+        raise HTTPException(
+            409,
+            f"'{m.display_name}' is the currently loaded model. Switch to a different model first, then try again."
+        )
+
+    # Refuse if a swap is in flight
+    if swap_manager.current_job_id:
+        raise HTTPException(409, "a model swap is in progress; wait for it to finish")
+
+    # Refuse if a download is in flight for this same repo (a different model's download is fine)
+    if download_manager.current_job_id:
+        job = download_manager.get(download_manager.current_job_id)
+        if job and job.repo == m.repo:
+            raise HTTPException(409, "this model is currently downloading; cancel or wait for it to finish")
+
+    status = await delete_from_disk(m.repo, m.mode, settings)
+    # Audit log
+    record_report(
+        f"disk:{key}",
+        ok=True,
+        source="disk-delete",
+        detail=f"freed {status.total_bytes} bytes across {len(status.per_host)} host(s)",
+    )
+    return {
+        "ok": True,
+        "key": key,
+        "repo": m.repo,
+        "bytes_freed": status.total_bytes,
+        "per_host": [
+            {"host": r.host, "size_bytes": r.size_bytes, **({"error": r.error} if r.error else {})}
+            for r in status.per_host
+        ],
+    }
+
+
+@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/connectivity")
+async def get_connectivity() -> dict:
+    """Up/down transition log per Spark + cached MACs."""
+    return connectivity_summary()
+
+
+@app.get("/api/deep-health")
+async def get_deep_health() -> dict:
+    """Last result + auto-restart counters for each service's synthetic probe."""
+    return deep_health.summary()
+
+
+@app.post("/api/deep-health/{service}/run")
+async def run_deep_health(service: str) -> dict:
+    """Manually run a single service's deep-health probe right now."""
+    if service not in deep_health.PROBES:
+        raise HTTPException(404, f"unknown service: {service}")
+    result = await deep_health.run_one(service)
+    return {
+        "ok": result.ok,
+        "at": result.at,
+        "latency_ms": result.latency_ms,
+        "error": result.error,
+        "note": result.note,
+    }
+
+
+class HealthEventBody(BaseModel):
+    service: str                 # e.g. "parakeet", "magpie", "vllm"
+    ok: bool                     # true on success, false on failure
+    source: str | None = None    # what app reported (e.g. "open-webui")
+    error: str | None = None     # optional detail
+    ms: int | None = None        # optional latency
+
+
+@app.post("/api/health-event")
+async def post_health_event(body: HealthEventBody) -> dict:
+    """Passive endpoint: any LAN app can POST here when its call to one of our
+    services succeeds or (more usefully) fails. We log the report into the
+    connectivity history so a brief blip that polling misses still surfaces.
+
+    Example:
+        curl -X POST http://<dashboard>/api/health-event \\
+          -H 'content-type: application/json' \\
+          -d '{"service":"parakeet","ok":false,"error":"503","source":"open-webui","ms":420}'
+    """
+    if not body.service.strip():
+        raise HTTPException(400, "service is required")
+    event = record_report(
+        body.service.strip(),
+        ok=body.ok,
+        source=(body.source or "external").strip(),
+        detail=(body.error or "").strip(),
+        latency_ms=body.ms,
+    )
+    return {"ok": True, "recorded": event}
+
+
+@app.post("/api/spark/{name}/wake")
+async def wake_spark(name: str) -> dict:
+    """Send a Wake-on-LAN magic packet for the named Spark.
+
+    Tries the OTHER Spark (if reachable) first because the packet has to
+    originate on the target's LAN segment to be reliable. Falls back to a
+    direct UDP broadcast from this container.
+    """
+    if name not in ("spark1", "spark2"):
+        raise HTTPException(404, f"unknown spark: {name}")
+    mac = get_mac(name)
+    if not mac:
+        raise HTTPException(400, f"MAC for {name} not yet known; bring it up once so we can probe it, then this will work next time it sleeps")
+
+    # Find the peer's connectivity to decide the path.
+    other = "spark2" if name == "spark1" else "spark1"
+    other_host = settings.spark1_host if other == "spark1" else settings.spark2_host
+    other_user = settings.spark1_user if other == "spark1" else settings.spark2_user
+
+    delivered_via = None
+    via_peer_ok = False
+    via_peer_err = ""
+    if other_host and other_user:
+        via_peer_ok, via_peer_err = await send_via_peer(other_host, other_user, mac, settings)
+        if via_peer_ok:
+            delivered_via = other
+
+    if not via_peer_ok:
+        # Fall back to direct from this container
+        try:
+            send_local_broadcast(mac)
+            delivered_via = "container"
+        except Exception as e:
+            raise HTTPException(500, f"WoL failed: peer={via_peer_err!r} container={e!r}")
+
+    return {"ok": True, "spark": name, "mac": mac, "delivered_via": delivered_via}
+
+
 @app.get("/api/services")
 async def get_services() -> dict:
     """Lifecycle state of always-on support services (Parakeet, Magpie, …).
@@ -158,9 +367,113 @@ async def get_services() -> dict:
     results = await asyncio.gather(*[one(n) for n in services.keys()])
     for name, info in results:
         out[name] = info
+        # Feed http reachability into the connectivity log (transition-only)
+        record_state(name, bool(info.get("http_ready")))
     return out
 
 
+@app.get("/api/nim/catalog")
+async def get_nim_catalog() -> dict:
+    return {
+        "catalog_url": CATALOG_URL,
+        "ngc_key_configured": bool(settings.ngc_api_key),
+        "suggested": SUGGESTED_NIMS,
+    }
+
+
+class NimInstallBody(BaseModel):
+    image: str
+    container: str
+    port: int
+    host: Literal["spark1", "spark2"] = "spark2"
+    kind: str = ""
+    register: bool = True   # write to custom services overrides after install
+
+
+@app.post("/api/nim/install")
+async def post_nim_install(body: NimInstallBody) -> dict:
+    target_host = settings.spark1_host if body.host == "spark1" else settings.spark2_host
+    target_user = settings.spark1_user if body.host == "spark1" else settings.spark2_user
+    try:
+        job = await nim_manager.trigger(
+            image=body.image,
+            container=body.container,
+            port=body.port,
+            host=target_host,
+            user=target_user,
+        )
+    except RuntimeError as e:
+        raise HTTPException(409 if "in progress" in str(e) else 400, str(e))
+
+    if body.register:
+        # Persist in custom services so the panel shows it after install.
+        add_custom_service({
+            "key": body.container,
+            "kind": body.kind or "nim",
+            "host": target_host,
+            "user": target_user,
+            "container": body.container,
+            "port": body.port,
+            "image": body.image,
+        })
+    return {"job_id": job.id, "image": job.image, "container": job.container, "state": job.state}
+
+
+@app.get("/api/nim/install/{job_id}")
+async def get_nim_install(job_id: str) -> dict:
+    job = nim_manager.get(job_id)
+    if job is None:
+        raise HTTPException(404, "no such job")
+    return {
+        "id": job.id,
+        "image": job.image,
+        "container": job.container,
+        "port": job.port,
+        "host": job.host,
+        "state": job.state,
+        "phase": job.phase,
+        "started_at": job.started_at,
+        "finished_at": job.finished_at,
+        "returncode": job.returncode,
+        "lines": job.lines,
+    }
+
+
+@app.get("/api/nim/install/{job_id}/stream")
+async def stream_nim_install(job_id: str):
+    job = nim_manager.get(job_id)
+    if job is None:
+        raise HTTPException(404, "no such job")
+
+    async def gen():
+        sent = 0
+        last_phase = None
+        while True:
+            n = len(job.lines)
+            if n > sent:
+                for line in job.lines[sent:n]:
+                    yield f"data: {json.dumps({'line': line})}\n\n"
+                sent = n
+            if job.phase != last_phase:
+                yield f"event: phase\ndata: {json.dumps({'state': job.state, 'phase': job.phase})}\n\n"
+                last_phase = job.phase
+            if job.returncode is not None and sent >= len(job.lines):
+                yield f"event: done\ndata: {json.dumps({'state': job.state, 'returncode': job.returncode})}\n\n"
+                return
+            await asyncio.sleep(0.5)
+
+    return StreamingResponse(gen(), media_type="text/event-stream")
+
+
+@app.delete("/api/services/{name}")
+async def del_service(name: str) -> dict:
+    # Only allow deleting custom services (not the bundled parakeet/magpie keys)
+    if name in ("parakeet", "magpie"):
+        raise HTTPException(400, "built-in service; cannot delete (use Configure Sparks to point at a different host)")
+    delete_custom_service(name)
+    return {"ok": True, "name": name}
+
+
 @app.post("/api/services/{name}/{action}")
 async def service_action(name: str, action: str) -> dict:
     services = services_from_settings(settings)
@@ -212,6 +525,10 @@ async def get_status() -> dict:
         check_parakeet(settings),
         check_magpie(settings),
     )
+    # Feed health into the connectivity log (deduped — only logs on transition)
+    record_state("vllm", bool(vllm.get("ok")))
+    record_state("parakeet", bool(parakeet.get("ok")))
+    record_state("magpie", bool(magpie.get("ok")))
     current_key = _identify_current_model(vllm.get("current_model"))
     return {
         "configured": settings.configured,
@@ -237,6 +554,15 @@ class SwapRequest(BaseModel):
     dry_run: bool = False
 
 
+@app.post("/api/swap/{key}/validate")
+async def validate_swap(key: str) -> dict:
+    """Pre-flight check: run vLLM's argparse layer against the proposed launch
+    command WITHOUT starting an engine. Cheap (~5 s) and doesn't disturb the
+    currently-loaded model.
+    """
+    return await validate_launch(key, catalog, settings)
+
+
 @app.post("/api/swap")
 async def post_swap(req: SwapRequest) -> dict:
     if not settings.configured and not req.dry_run:
@@ -297,7 +623,7 @@ 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 +702,81 @@ 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/services.py

@@ -5,6 +5,7 @@ machinery. We just run `docker start|stop|restart <container>` via SSH on the
 appropriate host.
 """
 from __future__ import annotations
+import time
 from dataclasses import dataclass
 from typing import Literal, Optional
 
@@ -12,6 +13,25 @@ from .config import Settings
 from .ssh import ssh_run
 
 
+# Cache the "unreachable" verdict per (host, user) for a short period so that a
+# repeated docker_state call doesn't re-pay the 6 s SSH connect timeout each time.
+_UNREACHABLE_TTL = 25.0
+_unreachable_cache: dict[tuple[str, str], float] = {}
+
+
+def _is_recently_unreachable(host: str, user: str) -> bool:
+    ts = _unreachable_cache.get((host, user))
+    return bool(ts and time.monotonic() - ts < _UNREACHABLE_TTL)
+
+
+def _mark_unreachable(host: str, user: str) -> None:
+    _unreachable_cache[(host, user)] = time.monotonic()
+
+
+def _clear_unreachable(host: str, user: str) -> None:
+    _unreachable_cache.pop((host, user), None)
+
+
 ServiceName = Literal["parakeet", "magpie"]
 ServiceAction = Literal["start", "stop", "restart"]
 
@@ -27,7 +47,8 @@ class ServiceDef:
 
 
 def services_from_settings(s: Settings) -> dict[str, ServiceDef]:
-    return {
+    from .custom_services import load_custom_services
+    out: dict[str, ServiceDef] = {
         "parakeet": ServiceDef(
             name="parakeet",
             kind="stt",
@@ -45,19 +66,38 @@ def services_from_settings(s: Settings) -> dict[str, ServiceDef]:
             port=s.magpie_port,
         ),
     }
+    for entry in load_custom_services():
+        key = entry.get("key")
+        if not key or key in out:
+            continue
+        out[key] = ServiceDef(
+            name=key,
+            kind=entry.get("kind", ""),
+            host=entry.get("host", ""),
+            user=entry.get("user", ""),
+            container=entry.get("container", key),
+            port=int(entry.get("port", 0)),
+        )
+    return out
 
 
 async def docker_state(settings: Settings, svc: ServiceDef) -> dict:
     """Get docker state (running, exited, restarting, etc.) + restart count."""
     if not svc.host or not svc.user:
         return {"state": "unconfigured", "restart_count": None, "uptime": None}
+    if _is_recently_unreachable(svc.host, svc.user):
+        return {"state": "unreachable", "host_unreachable": True, "restart_count": None, "uptime": None}
     cmd = (
         f"docker inspect {svc.container} "
         f"--format '{{{{.State.Status}}}}|{{{{.State.StartedAt}}}}|{{{{.RestartCount}}}}|{{{{.State.ExitCode}}}}|{{{{.State.Error}}}}' "
         f"2>&1 || echo 'NOT_FOUND'"
     )
-    rc, out, _ = await ssh_run(svc.host, svc.user, cmd, settings, timeout=10)
+    rc, out, _ = await ssh_run(svc.host, svc.user, cmd, settings, timeout=6)
     out = out.strip()
+    if rc == 124 or "timeout after" in out.lower():
+        _mark_unreachable(svc.host, svc.user)
+        return {"state": "unreachable", "host_unreachable": True, "restart_count": None, "uptime": None}
+    _clear_unreachable(svc.host, svc.user)
     if rc != 0 or out.startswith("NOT_FOUND") or "Error" in out and "no such object" in out.lower():
         return {"state": "missing", "restart_count": None, "uptime": None, "raw": out}
     parts = out.split("|")

image/app/static/app.js

@@ -13,8 +13,13 @@ const state = {
   swap_progress: 0,        // 0–1
   services: {},
   service_action_in_flight: null,  // e.g. "parakeet:restart"
+  hardware: {},
+  config: {},
   configured: true,
   timer_handle: null,
+  deep_health: {},
+  disk_status: {},         // keyed by model key: { on_disk, total_bytes, per_host }
+  disk_status_loaded: false,
 };
 
 const el = (sel) => document.querySelector(sel);
@@ -54,23 +59,52 @@ function renderCards() {
       ? `<div class="desc">${escapeHtml(m.description)}</div>`
       : '';
     const customPill = m.custom ? `<span class="tag custom-pill">custom</span>` : '';
+    // Disk-presence pill + trash button. Until /api/models/disk-status comes back,
+    // we don't know — render a neutral placeholder.
+    const disk = state.disk_status[key];
+    let diskPill = '';
+    if (state.disk_status_loaded) {
+      if (disk && disk.on_disk) {
+        const gb = (disk.total_bytes / 1e9);
+        diskPill = `<span class="tag on-disk" title="Weights present on disk">on disk · ${gb.toFixed(1)} GB</span>`;
+      } else {
+        diskPill = `<span class="tag not-on-disk" title="Weights not downloaded">not downloaded</span>`;
+      }
+    }
+    // Trash button — hidden if not on disk; disabled (with tooltip) if currently loaded.
+    let trashBtn = '';
+    if (state.disk_status_loaded && disk && disk.on_disk) {
+      const disabled = isActive || isSwapping;
+      const tip = isActive
+        ? 'Currently loaded — switch to another model first'
+        : isSwapping
+        ? 'A swap is in progress'
+        : 'Delete weights from disk';
+      trashBtn = `<button class="icon-btn danger" data-disk-del-key="${key}" title="${escapeHtml(tip)}" aria-label="Delete from disk" ${disabled ? 'disabled' : ''}>${trashIcon}</button>`;
+    }
     card.innerHTML = `
       <div class="name">${escapeHtml(m.display_name)}</div>
       <div class="meta">
         <span class="tag mode-${m.mode}">${m.mode}</span>
         <span class="tag">${m.size_gb} GB</span>
         ${customPill}
+        ${diskPill}
         ${(m.capabilities || []).map(c => `<span class="tag cap">${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' : ''}>
           ${isActive ? 'Current' : 'Switch to this'}
         </button>
+        <button class="btn test-btn" data-test-key="${key}" title="Pre-flight check the launch command without starting the engine">Test</button>
         <button class="btn adv-btn" data-adv-key="${key}" title="Advanced settings">Advanced</button>
+        ${trashBtn}
       </div>
+      <div class="test-result hidden" data-test-result-for="${key}"></div>
     `;
     root.appendChild(card);
   }
@@ -80,6 +114,42 @@ function renderCards() {
   for (const btn of root.querySelectorAll('[data-adv-key]')) {
     btn.addEventListener('click', () => openAdvanced(btn.dataset.advKey));
   }
+  for (const btn of root.querySelectorAll('[data-test-key]')) {
+    btn.addEventListener('click', () => testLaunch(btn.dataset.testKey, btn));
+  }
+  for (const btn of root.querySelectorAll('[data-disk-del-key]')) {
+    btn.addEventListener('click', () => openDiskDeleteDialog(btn.dataset.diskDelKey));
+  }
+}
+
+const trashIcon = '<svg viewBox="0 0 24 24" width="14" height="14" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><polyline points="3 6 5 6 21 6"></polyline><path d="M19 6l-1 14a2 2 0 0 1-2 2H8a2 2 0 0 1-2-2L5 6"></path><path d="M10 11v6"></path><path d="M14 11v6"></path><path d="M9 6V4a2 2 0 0 1 2-2h2a2 2 0 0 1 2 2v2"></path></svg>';
+
+async function testLaunch(key, btn) {
+  const resultEl = document.querySelector(`[data-test-result-for="${key}"]`);
+  if (!resultEl) return;
+  const originalText = btn.textContent;
+  btn.disabled = true;
+  btn.textContent = 'Testing…';
+  resultEl.classList.remove('hidden', 'ok', 'fail');
+  resultEl.innerHTML = '<span class="muted small">Checking launch args against vLLM\'s parser…</span>';
+  try {
+    const r = await fetchJSON(`/api/swap/${encodeURIComponent(key)}/validate`, { method: 'POST' });
+    if (r.ok) {
+      resultEl.classList.add('ok');
+      resultEl.innerHTML = `<span class="ok-mark">✓</span> Launch args parse OK. <span class="muted small">(Doesn't guarantee runtime success — only catches argparse-level issues.)</span>`;
+    } else {
+      resultEl.classList.add('fail');
+      const err = escapeHtml(r.error || 'unknown error');
+      const stage = r.stage ? ` <span class="muted small">(${escapeHtml(r.stage)})</span>` : '';
+      resultEl.innerHTML = `<span class="fail-mark">✗</span> Would fail: ${err}${stage}`;
+    }
+  } catch (e) {
+    resultEl.classList.add('fail');
+    resultEl.innerHTML = `<span class="fail-mark">✗</span> Test failed: ${escapeHtml(e.message)}`;
+  } finally {
+    btn.disabled = false;
+    btn.textContent = originalText;
+  }
 }
 
 function renderCurrent(status) {
@@ -93,6 +163,225 @@ 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');
+    try { state.connectivity = await fetchJSON('/api/connectivity'); } catch {}
+    renderHardware();
+  } catch (e) { console.warn('hardware poll failed', e); }
+}
+
+function fmtDuration(sec) {
+  if (sec == null) return '';
+  if (sec < 60) return `${Math.round(sec)}s`;
+  if (sec < 3600) return `${Math.round(sec / 60)}m`;
+  if (sec < 86400) {
+    const h = Math.floor(sec / 3600);
+    const m = Math.round((sec % 3600) / 60);
+    return m ? `${h}h ${m}m` : `${h}h`;
+  }
+  const d = Math.floor(sec / 86400);
+  const h = Math.round((sec % 86400) / 3600);
+  return h ? `${d}d ${h}h` : `${d}d`;
+}
+
+function openConnectivityDialog() {
+  const dlg = el('#connectivity-dialog');
+  const content = el('#connectivity-content');
+  const c = state.connectivity || {};
+  const events = c.events || [];
+  if (events.length === 0) {
+    content.innerHTML = '<div class="muted small">No events recorded yet. Once a Spark or service goes down and back up (or an external app reports a failure), entries appear here.</div>';
+    dlg.showModal();
+    return;
+  }
+  const bySubject = {};
+  for (const e of events) {
+    const subj = e.subject || e.spark || 'unknown';  // legacy fallback
+    (bySubject[subj] = bySubject[subj] || []).push(e);
+  }
+  // Sort subjects: hosts first, then services, alphabetical
+  const hostOrder = ['spark1', 'spark2'];
+  const subjects = Object.keys(bySubject).sort((a, b) => {
+    const ia = hostOrder.indexOf(a);
+    const ib = hostOrder.indexOf(b);
+    if (ia >= 0 && ib >= 0) return ia - ib;
+    if (ia >= 0) return -1;
+    if (ib >= 0) return 1;
+    return a.localeCompare(b);
+  });
+
+  const html = subjects.map((subj) => {
+    const evs = bySubject[subj];
+    const transitions = evs.filter(e => (e.kind || 'transition') === 'transition');
+    const reports = evs.filter(e => e.kind === 'report');
+    const downs = transitions.filter(e => e.transition === 'down').length;
+    const failedReports = reports.filter(e => !e.ok).length;
+    const mac = c.macs?.[subj];
+    const summaryParts = [];
+    if (transitions.length) summaryParts.push(`${transitions.length} probe transition${transitions.length===1?'':'s'} (${downs} down)`);
+    if (reports.length) summaryParts.push(`${reports.length} app report${reports.length===1?'':'s'} (${failedReports} failed)`);
+    const isHost = hostOrder.includes(subj);
+    return `
+      <div class="conn-spark">
+        <h4>${escapeHtml(subj)}${isHost ? ' <span class="muted small">[host]</span>' : ' <span class="muted small">[service]</span>'}${mac ? ` <span class="muted small">${escapeHtml(mac)}</span>` : ''}</h4>
+        <div class="conn-summary">${summaryParts.join(' · ') || 'no events'}</div>
+        ${evs.slice(-30).reverse().map(e => renderConnEvent(e)).join('')}
+      </div>
+    `;
+  }).join('');
+  content.innerHTML = html;
+  dlg.showModal();
+}
+
+function renderConnEvent(e) {
+  const when = escapeHtml((e.at || '').replace('T', ' ').replace('Z', ''));
+  const kind = e.kind || 'transition';
+  if (kind === 'report') {
+    const ok = !!e.ok;
+    const source = escapeHtml(e.source || 'external');
+    const detail = e.detail ? ` — ${escapeHtml(e.detail)}` : '';
+    const latency = e.latency_ms != null ? ` (${e.latency_ms} ms)` : '';
+    return `
+      <div class="conn-event ${ok ? 'up' : 'down'} report">
+        <span class="when">${when}</span>
+        <span class="what">${ok ? '◷ report: ok' : '◷ report: failed'} <span class="muted">from</span> ${source}${detail}</span>
+        <span class="dur">${latency}</span>
+      </div>
+    `;
+  }
+  const down = e.down_seconds != null ? `was down ${fmtDuration(e.down_seconds)}` : '';
+  const up = e.up_seconds != null ? `was up ${fmtDuration(e.up_seconds)}` : '';
+  return `
+    <div class="conn-event ${e.transition}">
+      <span class="when">${when}</span>
+      <span class="what">${e.transition === 'up' ? '↑ came back online' : '↓ dropped offline'}</span>
+      <span class="dur">${down}${up}</span>
+    </div>
+  `;
+}
+
+async function wakeSpark(name) {
+  try {
+    const r = await fetchJSON(`/api/spark/${name}/wake`, { method: 'POST' });
+    alert(`Wake-on-LAN sent to ${name} (MAC ${r.mac}, via ${r.delivered_via}). Give it ~30 seconds to wake; the card will go green when it comes back.`);
+  } catch (e) {
+    alert(`Wake failed: ${e.message}`);
+  }
+}
+
+function renderHardware() {
+  const panel = el('#hardware-panel');
+  const grid = el('#hardware-grid');
+  const hw = state.hardware || {};
+  const keys = Object.keys(hw).filter(k => hw[k] && (hw[k].configured !== false));
+  if (keys.length === 0) { panel.classList.add('hidden'); return; }
+  panel.classList.remove('hidden');
+  grid.innerHTML = '';
+  for (const key of keys) {
+    const s = hw[key];
+    const card = document.createElement('div');
+    if (!s.reachable) {
+      card.className = 'hw-card unreachable';
+      const mac = state.connectivity?.macs?.[key];
+      const wolRow = mac
+        ? `<div class="wol-row">
+             <span class="mac-display">${escapeHtml(mac)}</span>
+             <span class="spacer"></span>
+             <button class="btn" data-wake="${escapeHtml(key)}">Wake (WoL)</button>
+           </div>`
+        : `<div class="muted small">MAC not yet known — once it's been up once with this dashboard installed, "Wake" will appear here.</div>`;
+      card.innerHTML = `
+        <div class="head">
+          <span class="name">${escapeHtml(key)}</span>
+          <span class="meta">unreachable</span>
+        </div>
+        <div class="muted small">${escapeHtml(s.host || '')} — ${escapeHtml(s.error || 'no response')}</div>
+        ${wolRow}
+        <div class="muted small" style="line-height:1.5">
+          If Wake-on-LAN doesn't bring it back, manual steps:
+          <ol style="margin: 6px 0 0 18px; padding: 0;">
+            <li>Verify it's powered on (check the front LED).</li>
+            <li>Ping it from another LAN device.</li>
+            <li>Power-cycle it physically.</li>
+            <li>If it boots, this card will go green again automatically.</li>
+          </ol>
+        </div>
+      `;
+      grid.appendChild(card);
+      continue;
+    }
+    const ramPct = s.ram_used_bytes && s.ram_total_bytes ? (s.ram_used_bytes / s.ram_total_bytes) * 100 : 0;
+    const diskPct = s.disk_used_bytes && s.disk_total_bytes ? (s.disk_used_bytes / s.disk_total_bytes) * 100 : 0;
+    const loadPct = (s.load && s.cores) ? Math.min(100, (s.load[0] / s.cores) * 100) : 0;
+    // GPU memory: on unified-memory systems (DGX Spark) total is N/A, so use system RAM as the pool.
+    const gpuMemTotalMiB = s.gpu_mem_total_mib || (s.gpu_unified_memory ? (s.ram_total_bytes / (1024 * 1024)) : null);
+    const gpuMemUsedMiB = s.gpu_mem_used_mib ?? null;
+    const gpuMemPct = (gpuMemTotalMiB && gpuMemUsedMiB != null) ? (gpuMemUsedMiB / gpuMemTotalMiB) * 100 : 0;
+    const gpuMemNote = s.gpu_unified_memory ? ' <span class="muted">(unified)</span>' : '';
+    const gpuExtras = [];
+    if (s.gpu_temp_c != null) gpuExtras.push(`${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)} / ${fmtBytes(s.disk_total_bytes)}</span>
+      </div>
+    `;
+    grid.appendChild(card);
+  }
+}
+
+// ===================== service classification =====================
+
 function classifyService(s) {
   // returns one of: running | unhealthy | missing | unconfigured | starting
   if (!s.host) return 'unconfigured';
@@ -143,15 +432,49 @@ async function renderServices() {
       if (action === 'stop' && cls !== 'running' && cls !== 'starting' && cls !== 'unhealthy') return true;
       return false;
     };
+    const copyIcon = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>`;
+    const hostStr = s.host ? `${s.host}:${s.port}` : '';
     const hostRow = s.host
-      ? `<div class="row"><span class="k">Host</span><span class="v">${escapeHtml(s.host)}:${s.port}</span></div>`
+      ? `<div class="row"><span class="k">Host</span><span class="v copyable" data-copy-self title="Click to copy">${escapeHtml(hostStr)}</span><button class="icon-btn" data-copy-text="${escapeHtml(hostStr)}" title="Copy host" aria-label="Copy">${copyIcon}</button></div>`
       : `<div class="row"><span class="k">Host</span><span class="v muted-v">not configured</span></div>`;
+    const urlRow = s.base_url
+      ? `<div class="row"><span class="k">URL</span><span class="v copyable" data-copy-self title="Click to copy">${escapeHtml(s.base_url)}</span><button class="icon-btn" data-copy-text="${escapeHtml(s.base_url)}" title="Copy URL" aria-label="Copy">${copyIcon}</button></div>`
+      : '';
     const modelRow = s.model
-      ? `<div class="row"><span class="k">Model</span><span class="v">${escapeHtml(s.model)}</span></div>`
+      ? `<div class="row"><span class="k">Model</span><span class="v copyable" data-copy-self title="Click to copy">${escapeHtml(s.model)}</span><button class="icon-btn" data-copy-text="${escapeHtml(s.model)}" title="Copy model" aria-label="Copy">${copyIcon}</button></div>`
       : '';
     const restartsRow = s.restart_count != null && s.restart_count > 1
       ? `<div class="row"><span class="k">Restarts</span><span class="v">${s.restart_count}</span></div>`
       : '';
+    const dh = state.deep_health?.[name];
+    let deepRow = '';
+    if (dh && dh.last) {
+      const last = dh.last;
+      const when = (last.at || '').slice(11, 19);  // HH:MM:SS
+      const verdict = last.ok
+        ? `<span class="dh-ok">deep check ok</span>`
+        : `<span class="dh-fail">deep check FAILED</span>`;
+      const lat = last.latency_ms != null ? ` <span class="muted">${last.latency_ms} ms</span>` : '';
+      const restarts = dh.auto_restarts_window > 0
+        ? ` <span class="muted">· ${dh.auto_restarts_window} auto-restart${dh.auto_restarts_window === 1 ? '' : 's'} in 30 min</span>`
+        : '';
+      deepRow = `
+        <div class="row deep-row">
+          <span class="k">Deep</span>
+          <span class="v deep-v">${verdict} <span class="muted small">${escapeHtml(when)}</span>${lat}${restarts}</span>
+          <button class="icon-btn dh-run-btn" data-dh-run="${escapeHtml(name)}" title="Run deep check now">↻</button>
+        </div>
+        ${last.ok ? '' : `<div class="deep-error muted small">${escapeHtml((last.error || last.note || '').slice(0, 200))}</div>`}
+      `;
+    } else if (dh) {
+      deepRow = `
+        <div class="row deep-row">
+          <span class="k">Deep</span>
+          <span class="v muted-v">no probe yet</span>
+          <button class="icon-btn dh-run-btn" data-dh-run="${escapeHtml(name)}" title="Run deep check now">↻</button>
+        </div>
+      `;
+    }
     card.innerHTML = `
       <div class="head">
         <span class="name">${escapeHtml(name)}</span>
@@ -159,8 +482,10 @@ async function renderServices() {
         <span class="status">${statusLabel(cls)}</span>
       </div>
       ${hostRow}
+      ${urlRow}
       ${modelRow}
       ${restartsRow}
+      ${deepRow}
       <div class="service-actions">
         <button class="btn" data-svc-action="${name}:start" ${disable('start') ? 'disabled' : ''}>Start</button>
         <button class="btn" data-svc-action="${name}:restart" ${disable('restart') ? 'disabled' : ''}>Restart</button>
@@ -172,6 +497,25 @@ async function renderServices() {
   for (const btn of grid.querySelectorAll('.btn[data-svc-action]')) {
     btn.addEventListener('click', () => onServiceAction(btn.dataset.svcAction));
   }
+  for (const btn of grid.querySelectorAll('[data-dh-run]')) {
+    btn.addEventListener('click', () => onDeepHealthRun(btn.dataset.dhRun, btn));
+  }
+}
+
+async function onDeepHealthRun(name, btn) {
+  btn.disabled = true;
+  const orig = btn.textContent;
+  btn.textContent = '…';
+  try {
+    await fetchJSON(`/api/deep-health/${encodeURIComponent(name)}/run`, { method: 'POST' });
+  } catch (e) {
+    console.warn('deep-health run failed', e);
+  } finally {
+    try { state.deep_health = await fetchJSON('/api/deep-health'); } catch {}
+    btn.textContent = orig;
+    btn.disabled = false;
+    renderServices();
+  }
 }
 
 async function onServiceAction(key) {
@@ -212,31 +556,50 @@ function renderEndpoint(status) {
   el('#ep-curl-snippet').textContent = snippet;
 }
 
-function setupCopyButtons() {
-  document.body.addEventListener('click', async (e) => {
-    const btn = e.target.closest('.copy-btn');
-    if (!btn) return;
-    const targetSel = btn.dataset.copy;
-    if (!targetSel) return;
-    const target = el(targetSel);
-    if (!target) return;
-    const text = target.textContent;
+async function copyText(text, indicatorEl) {
   try {
     await navigator.clipboard.writeText(text);
-      const original = btn.textContent;
-      btn.classList.add('copied');
-      btn.textContent = 'Copied';
-      setTimeout(() => {
-        btn.classList.remove('copied');
-        btn.textContent = original;
-      }, 1400);
+    if (indicatorEl) {
+      indicatorEl.classList.add('copied');
+      setTimeout(() => indicatorEl.classList.remove('copied'), 1200);
+    }
+    return true;
   } catch {
-      // Clipboard API may fail over plain HTTP; fall back to selection
+    // Plain HTTP fallback: select the text so the user can ⌘C
+    if (indicatorEl) {
       const range = document.createRange();
-      range.selectNode(target);
+      range.selectNode(indicatorEl);
       window.getSelection().removeAllRanges();
       window.getSelection().addRange(range);
     }
+    return false;
+  }
+}
+
+function setupCopyButtons() {
+  document.body.addEventListener('click', async (e) => {
+    // Inline icon copy with literal text (used for dynamically-rendered service rows)
+    const litBtn = e.target.closest('[data-copy-text]');
+    if (litBtn) {
+      await copyText(litBtn.dataset.copyText, litBtn);
+      return;
+    }
+    // Copy buttons (with svg icon) referenced by data-copy="selector"
+    const btn = e.target.closest('[data-copy]');
+    if (btn) {
+      const target = el(btn.dataset.copy);
+      if (target) {
+        await copyText(target.textContent, btn);
+        target.classList.add('copied');
+        setTimeout(() => target.classList.remove('copied'), 1200);
+      }
+      return;
+    }
+    // Self-copy: clicking the text itself
+    const selfCopy = e.target.closest('[data-copy-self]');
+    if (selfCopy) {
+      await copyText(selfCopy.textContent, selfCopy);
+    }
   });
 }
 
@@ -380,9 +743,14 @@ async function pollStatus() {
     renderCurrent(status);
     renderEndpoint(status);
     renderHealth(status);
+    // If models hasn't loaded yet (init may have hit a transient proxy timeout), retry.
+    if (!state.models || Object.keys(state.models).length === 0) {
+      try { await loadModels(); } catch {}
+    }
     // Refresh services state lazily — every 5s poll triggers this too.
     try {
       state.services = await fetchJSON('/api/services');
+      try { state.deep_health = await fetchJSON('/api/deep-health'); } catch {}
       renderServices();
     } catch {}
     if (status.current_swap_job && status.current_swap_job !== state.swap_job_id) {
@@ -403,6 +771,78 @@ async function loadModels() {
   state.models = data.models || {};
 }
 
+async function loadDiskStatus() {
+  // Probes each catalog model's HF cache over SSH; takes a beat. Best-effort.
+  try {
+    const r = await fetchJSON('/api/models/disk-status');
+    if (r && r.models) {
+      state.disk_status = r.models;
+      state.disk_status_loaded = true;
+      renderCards();
+    }
+  } catch (e) {
+    // Silent — pills just won't render. Don't block dashboard.
+    console.warn('disk-status probe failed:', e.message);
+  }
+}
+
+function fmtBytesShort(n) {
+  if (!Number.isFinite(n) || n <= 0) return '0 B';
+  if (n >= 1e9) return `${(n / 1e9).toFixed(1)} GB`;
+  if (n >= 1e6) return `${(n / 1e6).toFixed(1)} MB`;
+  if (n >= 1e3) return `${(n / 1e3).toFixed(1)} KB`;
+  return `${n} B`;
+}
+
+function openDiskDeleteDialog(key) {
+  const m = state.models[key];
+  const disk = state.disk_status[key];
+  if (!m || !disk || !disk.on_disk) return;
+  const dlg = el('#disk-delete-dialog');
+  el('#dd-summary').innerHTML = `Free <strong>${fmtBytesShort(disk.total_bytes)}</strong> by removing <strong>${escapeHtml(m.display_name)}</strong> (<code>${escapeHtml(m.repo)}</code>) from disk.`;
+  const hostsEl = el('#dd-hosts');
+  hostsEl.innerHTML = '';
+  for (const h of (disk.per_host || [])) {
+    if (!h.on_disk) continue;
+    const li = document.createElement('li');
+    li.innerHTML = `<code>${escapeHtml(h.host)}</code> — ${fmtBytesShort(h.size_bytes)}`;
+    hostsEl.appendChild(li);
+  }
+  const errEl = el('#dd-error');
+  errEl.classList.add('hidden');
+  errEl.textContent = '';
+
+  const confirm = el('#dd-confirm');
+  const cancel = el('#dd-cancel');
+  const onCancel = () => dlg.close();
+  const onConfirm = async () => {
+    confirm.disabled = true;
+    cancel.disabled = true;
+    confirm.textContent = 'Deleting…';
+    try {
+      const r = await fetchJSON(`/api/models/${encodeURIComponent(key)}/disk`, { method: 'DELETE' });
+      dlg.close();
+      // Optimistically clear local disk state for this key, then refresh.
+      delete state.disk_status[key];
+      renderCards();
+      // Eagerly re-probe so size is accurate (and shows "not downloaded" pill).
+      loadDiskStatus();
+      const freed = r && typeof r.bytes_freed === 'number' ? fmtBytesShort(r.bytes_freed) : '';
+      console.log(`Deleted ${m.display_name} from disk${freed ? ` — freed ${freed}` : ''}.`);
+    } catch (e) {
+      errEl.textContent = e.message || 'Delete failed';
+      errEl.classList.remove('hidden');
+    } finally {
+      confirm.disabled = false;
+      cancel.disabled = false;
+      confirm.textContent = 'Delete from disk';
+    }
+  };
+  cancel.onclick = onCancel;
+  confirm.onclick = onConfirm;
+  dlg.showModal();
+}
+
 async function triggerSwap(modelKey) {
   if (state.swap_job_id) return;
   try {
@@ -518,6 +958,18 @@ function openDownloadForm() {
   el('#download-form').classList.remove('hidden');
   el('#download-progress').classList.add('hidden');
   el('#dl-repo').focus();
+  updateDlHfLink();
+}
+
+function updateDlHfLink() {
+  const repo = el('#dl-repo').value.trim();
+  const link = el('#dl-hf-link');
+  if (repo.includes('/')) {
+    link.href = `https://huggingface.co/${encodeURIComponent(repo)}`;
+    link.classList.remove('hidden');
+  } else {
+    link.classList.add('hidden');
+  }
 }
 
 function closeDownloadPanel() {
@@ -647,6 +1099,47 @@ 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;
@@ -659,6 +1152,23 @@ 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 () => {
@@ -753,6 +1263,197 @@ function setupAdvancedDialog() {
   el('#adv-gmu').addEventListener('input', (e) => { el('#adv-gmu-out').value = parseFloat(e.target.value).toFixed(2); });
 }
 
+// ===================== NIM installer =====================
+
+const nimState = {
+  catalog: null,
+  job_id: null,
+  eventsource: null,
+  timer: null,
+  started_at: null,
+};
+
+async function loadNimCatalog() {
+  try {
+    nimState.catalog = await fetchJSON('/api/nim/catalog');
+    el('#nim-catalog-link').href = nimState.catalog.catalog_url;
+    const warn = el('#nim-key-warn');
+    if (!nimState.catalog.ngc_key_configured) {
+      warn.classList.add('nim-key-warn');
+      warn.innerHTML = '⚠️ NGC API key not set. Open <strong>Configure Sparks</strong> in StartOS and paste your NGC personal API key, otherwise installs will fail. <a href="https://ngc.nvidia.com/setup/personal-key" target="_blank" rel="noopener">Get a key</a>';
+    } else {
+      warn.classList.remove('nim-key-warn');
+      warn.textContent = '';
+    }
+    const grid = el('#nim-suggested');
+    grid.innerHTML = '';
+    for (const s of nimState.catalog.suggested || []) {
+      const card = document.createElement('div');
+      card.className = 'nim-card';
+      card.innerHTML = `
+        <div class="info">
+          <div class="name">${escapeHtml(s.name)} <span class="muted small">· ${escapeHtml(s.kind || 'nim')}</span></div>
+          <div class="desc">${escapeHtml(s.description || '')}</div>
+          <div class="img">${escapeHtml(s.image)}</div>
+          <div class="links">${s.homepage ? `<a href="${escapeHtml(s.homepage)}" target="_blank" rel="noopener">View on NGC ↗</a>` : ''}</div>
+        </div>
+        <button type="button" class="btn primary nim-pick" data-image="${escapeHtml(s.image)}" data-container="${escapeHtml(s.default_container)}" data-port="${s.default_port}" data-kind="${escapeHtml(s.kind)}">Pick</button>
+      `;
+      grid.appendChild(card);
+    }
+    grid.querySelectorAll('.nim-pick').forEach(btn => {
+      btn.addEventListener('click', () => {
+        el('#nim-image').value = btn.dataset.image;
+        el('#nim-container').value = btn.dataset.container;
+        el('#nim-port').value = btn.dataset.port;
+        el('#nim-kind').value = btn.dataset.kind || 'nim';
+      });
+    });
+  } catch (e) { console.warn('nim catalog failed', e); }
+}
+
+function openNimDialog() {
+  loadNimCatalog();
+  el('#nim-dialog').showModal();
+}
+
+async function submitNim(e) {
+  e.preventDefault();
+  const body = {
+    image: el('#nim-image').value.trim(),
+    container: el('#nim-container').value.trim(),
+    port: parseInt(el('#nim-port').value, 10),
+    host: el('#nim-host').value,
+    kind: el('#nim-kind').value,
+  };
+  if (!body.image || !body.container || !body.port) {
+    alert('Image, container name, and port are required.');
+    return;
+  }
+  try {
+    const r = await fetchJSON('/api/nim/install', {
+      method: 'POST',
+      headers: { 'content-type': 'application/json' },
+      body: JSON.stringify(body),
+    });
+    el('#nim-dialog').close();
+    attachNimProgress(r.job_id);
+  } catch (e) {
+    alert('Install failed: ' + e.message);
+  }
+}
+
+function nimTimerStart(at) {
+  nimState.started_at = at;
+  if (nimState.timer) clearInterval(nimState.timer);
+  const tick = () => {
+    if (!nimState.started_at) return;
+    const sec = Math.max(0, Math.floor((Date.now() - nimState.started_at) / 1000));
+    const m = Math.floor(sec / 60);
+    const s = sec % 60;
+    el('#nim-prog-elapsed').textContent = `${m}:${s.toString().padStart(2, '0')}`;
+  };
+  tick();
+  nimState.timer = setInterval(tick, 500);
+}
+
+async function attachNimProgress(jobId) {
+  nimState.job_id = jobId;
+  el('#nim-prog-log').textContent = '';
+  el('#nim-prog-title').textContent = 'Installing…';
+  el('#nim-progress-dialog').showModal();
+  try {
+    const snap = await fetchJSON(`/api/nim/install/${jobId}`);
+    nimTimerStart(Date.parse(snap.started_at));
+    el('#nim-prog-phase').textContent = snap.phase || 'Working…';
+    el('#nim-prog-log').textContent = (snap.lines || []).join('\n');
+    if (snap.returncode !== null) { onNimDone(snap); return; }
+  } catch { nimTimerStart(Date.now()); }
+  const es = new EventSource(`/api/nim/install/${jobId}/stream`);
+  nimState.eventsource = es;
+  es.onmessage = ev => {
+    try {
+      const d = JSON.parse(ev.data);
+      if (d.line !== undefined) {
+        const log = el('#nim-prog-log');
+        log.textContent += d.line + '\n';
+        log.scrollTop = log.scrollHeight;
+      }
+    } catch {}
+  };
+  es.addEventListener('phase', ev => {
+    try { el('#nim-prog-phase').textContent = JSON.parse(ev.data).phase; } catch {}
+  });
+  es.addEventListener('done', ev => {
+    let d = {}; try { d = JSON.parse(ev.data); } catch {}
+    onNimDone(d);
+  });
+  es.onerror = () => { es.close(); nimState.eventsource = null; };
+}
+
+function onNimDone(d) {
+  if (nimState.eventsource) { nimState.eventsource.close(); nimState.eventsource = null; }
+  if (nimState.timer) { clearInterval(nimState.timer); nimState.timer = null; }
+  if (d.state === 'failed') {
+    el('#nim-prog-title').textContent = `Failed (rc=${d.returncode})`;
+    el('#nim-prog-phase').textContent = 'Failed';
+  } else {
+    el('#nim-prog-title').textContent = 'Installed';
+    el('#nim-prog-phase').textContent = 'Done ✓ — service will appear when the container reports healthy.';
+  }
+  pollStatus();
+}
+
+// ===================== 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 = {
@@ -792,19 +1493,23 @@ 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)';
   }
 }
@@ -893,14 +1598,41 @@ 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);
+  el('#open-nim').addEventListener('click', openNimDialog);
+  el('#nim-cancel').addEventListener('click', () => el('#nim-dialog').close());
+  el('#nim-form').addEventListener('submit', submitNim);
+  el('#nim-prog-close').addEventListener('click', () => el('#nim-progress-dialog').close());
+  el('#open-connectivity').addEventListener('click', openConnectivityDialog);
+  el('#connectivity-close').addEventListener('click', () => el('#connectivity-dialog').close());
+  // Wake-on-LAN buttons live on unreachable hardware cards; delegate.
+  el('#hardware-grid').addEventListener('click', (e) => {
+    const btn = e.target.closest('[data-wake]');
+    if (btn) wakeSpark(btn.dataset.wake);
+  });
   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();
+  // Disk-status probe runs after first paint — slow over SSH and not blocking.
+  loadDiskStatus();
   setInterval(pollStatus, 5000);
+  setInterval(pollHardware, 8000);    // every 8s
   setInterval(pollUpdates, 300000);  // every 5 min
+  setInterval(loadDiskStatus, 60000); // every 60s — disk state changes rarely
 }
 
 init();

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>
@@ -24,22 +25,47 @@
       <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">
+      <div class="section-header">
+        <h2 class="section-title">Spark hardware</h2>
+        <button id="open-connectivity" class="btn small-btn">Connectivity log</button>
+      </div>
+      <div id="hardware-grid" class="hardware-grid"></div>
+
+      <dialog id="connectivity-dialog" class="modal">
+        <form method="dialog" class="modal-form">
+          <h3>Spark connectivity history</h3>
+          <p class="muted small">Most recent up/down transitions per Spark. Tracked since this dashboard was installed.</p>
+          <div id="connectivity-content" class="connectivity-content"></div>
+          <div class="modal-actions">
+            <button type="button" id="connectivity-close" class="btn">Close</button>
+          </div>
+        </form>
+      </dialog>
+    </section>
+
     <section id="endpoint-panel" class="endpoint-panel hidden">
       <div class="ep-title muted small">OpenAI-compatible endpoint</div>
       <div class="ep-row">
         <span class="ep-label">Base URL</span>
-        <code class="ep-value" id="ep-url">—</code>
-        <button class="copy-btn" data-copy="#ep-url" title="Copy base URL">Copy</button>
+        <code class="ep-value copyable" id="ep-url" data-copy-self title="Click to copy">—</code>
+        <button class="icon-btn" data-copy="#ep-url" title="Copy base URL" aria-label="Copy">
+          <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>
+        </button>
       </div>
       <div class="ep-row">
         <span class="ep-label">Model ID</span>
-        <code class="ep-value" id="ep-model">—</code>
-        <button class="copy-btn" data-copy="#ep-model" title="Copy model ID">Copy</button>
+        <code class="ep-value copyable" id="ep-model" data-copy-self title="Click to copy">—</code>
+        <button class="icon-btn" data-copy="#ep-model" title="Copy model ID" aria-label="Copy">
+          <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>
+        </button>
       </div>
       <details class="ep-curl">
         <summary class="muted small">curl example</summary>
-        <pre id="ep-curl-snippet" class="snippet"></pre>
-        <button class="copy-btn small" data-copy="#ep-curl-snippet">Copy snippet</button>
+        <pre id="ep-curl-snippet" class="snippet copyable" data-copy-self title="Click to copy"></pre>
+        <button class="icon-btn" data-copy="#ep-curl-snippet" title="Copy snippet" aria-label="Copy">
+          <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>
+        </button>
       </details>
     </section>
 
@@ -64,8 +90,66 @@
     </section>
 
     <section id="services-panel" class="services hidden">
+      <div class="section-header">
         <h2 class="section-title">Always-on services</h2>
+        <button id="open-nim" class="btn small-btn">+ Install NIM</button>
+      </div>
       <div id="services-grid" class="services-grid"></div>
+
+      <dialog id="nim-dialog" class="modal">
+        <form method="dialog" class="modal-form" id="nim-form">
+          <h3>Install a NVIDIA NIM container</h3>
+          <p class="muted small" id="nim-key-warn"></p>
+          <p class="muted small">Pick a curated container below or paste any image from <a href="#" id="nim-catalog-link" target="_blank" rel="noopener">the NGC NIM catalog</a>. Spark Control will <code>docker pull</code> and <code>docker run</code> it on the target Spark.</p>
+
+          <div id="nim-suggested" class="nim-grid"></div>
+
+          <fieldset class="modal-fieldset">
+            <legend>Custom image</legend>
+            <label class="modal-row"><span>Image (nvcr.io/...)</span><input type="text" id="nim-image" placeholder="nvcr.io/nim/nvidia/<name>:latest"></label>
+            <label class="modal-row"><span>Container name</span><input type="text" id="nim-container" placeholder="my-service"></label>
+            <label class="modal-row"><span>Port</span><input type="number" id="nim-port" min="1" max="65535"></label>
+            <label class="modal-row"><span>Kind</span>
+              <select id="nim-kind">
+                <option value="nim">NIM (other)</option>
+                <option value="stt">STT (speech-to-text)</option>
+                <option value="tts">TTS (text-to-speech)</option>
+                <option value="vision">Vision</option>
+                <option value="embedding">Embedding</option>
+              </select>
+            </label>
+            <label class="modal-row"><span>Target Spark</span>
+              <select id="nim-host">
+                <option value="spark2">Spark 2 (default for support services)</option>
+                <option value="spark1">Spark 1 (head node)</option>
+              </select>
+            </label>
+          </fieldset>
+
+          <div class="modal-actions">
+            <button type="button" id="nim-cancel" class="btn">Cancel</button>
+            <button type="submit" class="btn primary" id="nim-start">Install</button>
+          </div>
+        </form>
+      </dialog>
+
+      <dialog id="nim-progress-dialog" class="modal">
+        <form method="dialog" class="modal-form">
+          <h3 id="nim-prog-title">Installing…</h3>
+          <div class="phase-row">
+            <div class="phase" id="nim-prog-phase">Starting…</div>
+            <span class="spacer"></span>
+            <span class="timer" id="nim-prog-elapsed">0:00</span>
+          </div>
+          <details open>
+            <summary class="muted small">Log</summary>
+            <pre id="nim-prog-log" class="log"></pre>
+          </details>
+          <div class="modal-actions">
+            <button type="button" id="nim-prog-close" class="btn">Close</button>
+          </div>
+        </form>
+      </dialog>
     </section>
 
     <section id="models-section">
@@ -104,6 +188,20 @@
         </form>
       </dialog>
 
+      <dialog id="disk-delete-dialog" class="modal">
+        <form method="dialog" class="modal-form">
+          <h3>Delete model weights from disk?</h3>
+          <p id="dd-summary" class="muted small"></p>
+          <ul class="muted small dd-hosts" id="dd-hosts"></ul>
+          <p class="muted small">This is reversible — you can re-download from the catalog at any time. The catalog entry stays intact.</p>
+          <p id="dd-error" class="muted small dd-error hidden"></p>
+          <div class="modal-actions">
+            <button type="button" id="dd-cancel" class="btn">Cancel</button>
+            <button type="button" id="dd-confirm" class="btn danger">Delete from disk</button>
+          </div>
+        </form>
+      </dialog>
+
       <dialog id="advanced-dialog" class="modal">
         <form method="dialog" class="modal-form" id="advanced-form">
           <h3 id="adv-title">Advanced settings</h3>
@@ -127,11 +225,20 @@
           <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="cluster"> Both Sparks (cluster, copy in parallel)</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="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>
@@ -165,9 +272,14 @@
     </section>
 
     <section id="update-banner" class="update-banner hidden">
+      <div class="ub-context muted small">
+        Updates to <strong><a href="https://github.com/eugr/spark-vllm-docker" target="_blank" rel="noopener">eugr/spark-vllm-docker</a></strong>
+        — the upstream project that orchestrates vLLM on your Sparks (launch-cluster.sh, recipes, mods). These are <em>not</em> firmware, OS, or model updates.
+      </div>
       <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>
@@ -175,6 +287,10 @@
         <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;
@@ -97,7 +108,8 @@ main {
   overflow-x: auto;
   white-space: nowrap;
 }
-.copy-btn {
+.copy-btn,
+.icon-btn {
   appearance: none;
   background: var(--surface-2);
   border: 1px solid var(--border);
@@ -108,15 +120,27 @@ main {
   cursor: pointer;
   transition: color 0.15s, border-color 0.15s, background 0.15s;
   flex-shrink: 0;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
 }
-.copy-btn:hover { color: var(--text); border-color: #34343c; }
-.copy-btn.copied {
+.icon-btn { padding: 5px 7px; }
+.icon-btn svg { width: 14px; height: 14px; display: block; }
+.copy-btn:hover,
+.icon-btn:hover { color: var(--text); border-color: #34343c; }
+.copy-btn.copied,
+.icon-btn.copied {
   color: var(--accent);
   border-color: rgba(74, 222, 128, 0.4);
   background: rgba(74, 222, 128, 0.08);
 }
+.icon-btn.copied svg { color: var(--accent); }
 .copy-btn.small { padding: 3px 8px; font-size: 11px; }
 
+.copyable { cursor: pointer; }
+.copyable:hover { outline: 1px solid rgba(96, 165, 250, 0.5); }
+.copyable.copied { outline: 1px solid var(--accent); background: rgba(74, 222, 128, 0.05); }
+
 .ep-curl { margin-top: 8px; }
 .ep-curl summary { cursor: pointer; padding: 4px 0; }
 .ep-curl[open] summary { margin-bottom: 6px; }
@@ -255,6 +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 {
@@ -274,10 +306,39 @@ main {
   background: var(--surface);
   border: 1px solid rgba(96, 165, 250, 0.4);
   border-radius: var(--radius);
-  padding: 10px 14px;
+  padding: 12px 14px;
   margin-top: 18px;
   font-size: 13px;
 }
+.ub-context { margin-bottom: 8px; line-height: 1.5; }
+.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);
@@ -289,6 +350,90 @@ 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-card.unreachable ol { color: var(--muted); }
+.hw-card .wol-row {
+  margin-top: 8px;
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  font-size: 12px;
+  color: var(--muted);
+}
+.hw-card .wol-row .btn { padding: 5px 10px; font-size: 12px; }
+.hw-card .mac-display { font-family: ui-monospace, SFMono-Regular, Menlo, monospace; }
+
+.connectivity-content {
+  max-height: 360px;
+  overflow-y: auto;
+  border: 1px solid var(--border);
+  border-radius: 6px;
+  padding: 10px;
+  background: var(--surface-2);
+}
+.conn-spark { margin-bottom: 16px; }
+.conn-spark h4 { font-size: 13px; margin: 0 0 8px; color: var(--text); }
+.conn-event {
+  font-size: 12px;
+  display: flex;
+  gap: 10px;
+  padding: 4px 0;
+  border-bottom: 1px solid rgba(255,255,255,0.04);
+  font-family: ui-monospace, SFMono-Regular, Menlo, monospace;
+}
+.conn-event:last-child { border-bottom: 0; }
+.conn-event .when { color: var(--muted); flex-shrink: 0; }
+.conn-event .what { flex: 1; }
+.conn-event.up .what { color: var(--accent); }
+.conn-event.down .what { color: var(--error); }
+.conn-event.report .what { font-style: italic; }
+.conn-event .muted { color: var(--muted); font-style: normal; }
+.conn-event .dur { color: var(--muted); }
+.conn-summary { color: var(--muted); font-size: 11px; padding: 4px 0 10px; }
+.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 {
@@ -341,6 +486,24 @@ 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; }
@@ -353,6 +516,37 @@ main {
 #dl-log-details { margin-top: 12px; }
 #dl-log-details summary { cursor: pointer; padding: 4px 0; }
 
+/* ===== NIM install dialog ===== */
+
+.modal#nim-dialog,
+.modal#nim-progress-dialog { max-width: 640px; }
+.nim-grid {
+  display: grid;
+  gap: 8px;
+  grid-template-columns: 1fr;
+  max-height: 240px;
+  overflow-y: auto;
+  margin-bottom: 4px;
+}
+.nim-card {
+  background: var(--surface-2);
+  border: 1px solid var(--border);
+  border-radius: 6px;
+  padding: 10px 12px;
+  display: flex;
+  gap: 10px;
+  align-items: flex-start;
+}
+.nim-card .info { flex: 1; }
+.nim-card .name { font-weight: 600; font-size: 13px; }
+.nim-card .desc { color: var(--muted); font-size: 12px; margin-top: 4px; }
+.nim-card .img { font-family: ui-monospace, SFMono-Regular, Menlo, monospace; color: #6b6b75; font-size: 11px; margin-top: 4px; word-break: break-all; }
+.nim-card .btn { padding: 6px 12px; font-size: 12px; flex-shrink: 0; }
+.nim-card .links { font-size: 11px; margin-top: 4px; }
+.nim-card .links a { color: var(--info); text-decoration: none; }
+.nim-card .links a:hover { text-decoration: underline; }
+.nim-key-warn { color: var(--warn); }
+
 /* ===== Section titles ===== */
 
 .section-title {
@@ -409,13 +603,38 @@ main {
 
 .service-card .row {
   display: flex;
+  align-items: center;
   font-size: 12px;
   color: var(--muted);
   gap: 6px;
 }
 .service-card .row .k { width: 60px; flex-shrink: 0; }
-.service-card .row .v { color: var(--text); font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace; word-break: break-all; }
+.service-card .row .v {
+  color: var(--text);
+  font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+  word-break: break-all;
+  flex: 1;
+  padding: 2px 4px;
+  border-radius: 4px;
+}
 .service-card .row .v.muted-v { color: var(--muted); font-family: inherit; }
+.service-card .row .v.copyable:hover { outline: 1px solid rgba(96, 165, 250, 0.5); }
+.service-card .row .v.copyable.copied { outline: 1px solid var(--accent); background: rgba(74, 222, 128, 0.05); }
+.service-card .row .icon-btn { padding: 3px 6px; }
+.service-card .row .icon-btn svg { width: 12px; height: 12px; }
+.service-card .deep-row .deep-v { display: flex; align-items: center; gap: 6px; font-family: inherit; flex-wrap: wrap; }
+.service-card .dh-ok { color: var(--accent); }
+.service-card .dh-fail { color: var(--error); font-weight: 500; }
+.service-card .dh-run-btn { font-family: inherit; }
+.service-card .deep-error {
+  padding: 4px 8px;
+  background: rgba(239, 68, 68, 0.06);
+  border-left: 2px solid var(--error);
+  border-radius: 4px;
+  font-family: ui-monospace, SFMono-Regular, Menlo, monospace;
+  font-size: 11px;
+  word-break: break-word;
+}
 
 .service-actions {
   display: flex;
@@ -460,6 +679,9 @@ 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);
@@ -492,8 +714,31 @@ main {
 .card.active .btn { background: rgba(74, 222, 128, 0.12); color: var(--accent); border-color: rgba(74, 222, 128, 0.4); }
 .card-actions { display: flex; gap: 6px; }
 .card-actions .btn.primary { flex: 1; }
-.card .adv-btn { padding: 8px 12px; font-size: 12px; }
+.card .adv-btn,
+.card .test-btn { padding: 8px 12px; font-size: 12px; }
 .card .custom-pill { color: var(--info); border-color: rgba(96, 165, 250, 0.4); }
+.tag.on-disk { color: var(--accent); border-color: rgba(74, 222, 128, 0.4); }
+.tag.not-on-disk { color: var(--muted); border-color: var(--border); opacity: 0.7; }
+.card-actions .icon-btn.danger { color: var(--error); border-color: rgba(239, 68, 68, 0.3); margin-left: auto; }
+.card-actions .icon-btn.danger:hover:not(:disabled) { background: rgba(239, 68, 68, 0.08); border-color: var(--error); color: var(--error); }
+.card-actions .icon-btn.danger:disabled { opacity: 0.35; cursor: not-allowed; }
+.dd-hosts { padding-left: 18px; margin: 4px 0 8px; }
+.dd-hosts code { background: var(--surface-2); padding: 1px 5px; border-radius: 4px; }
+.dd-error { color: var(--error); }
+
+.test-result {
+  font-size: 12px;
+  line-height: 1.45;
+  padding: 8px 10px;
+  border-radius: 5px;
+  margin-top: 4px;
+  border: 1px solid var(--border);
+  background: var(--surface-2);
+}
+.test-result.ok { border-color: rgba(74, 222, 128, 0.4); background: rgba(74, 222, 128, 0.04); }
+.test-result.fail { border-color: rgba(239, 68, 68, 0.45); background: rgba(239, 68, 68, 0.06); word-break: break-word; }
+.test-result .ok-mark { color: var(--accent); font-weight: 600; }
+.test-result .fail-mark { color: var(--error); font-weight: 600; }
 
 .footer {
   margin-top: 28px;

image/app/validate.py

@@ -0,0 +1,137 @@
+"""Pre-flight validation of a proposed vLLM launch command.
+
+Runs vLLM's own argparse layer (EngineArgs) inside the vllm_node container WITHOUT
+starting the engine. Catches:
+
+  * unknown flag names (typos)
+  * bad types / values that argparse rejects
+  * deprecated flags removed in the installed vLLM version
+
+Does NOT catch (these surface only during real engine init):
+  * model-architecture-specific constraints (e.g. Qwen3.6 Mamba block_size)
+  * OOM at weight-loading time
+  * Triton / CUDA-kernel compatibility errors
+
+A pre-flight check that returns "ok" is therefore NOT a guarantee — but a
+"failed" verdict is a definitive 'don't bother with the real swap'.
+"""
+from __future__ import annotations
+import json
+import shlex
+from typing import Any
+
+from .config import Settings
+from .models import Catalog, build_launch_command
+from .ssh import ssh_run
+
+
+# Validates the proposed args against the same combined parser vLLM uses for
+# `vllm serve` (engine args + server args + frontend args). Returns one JSON
+# line on stdout: {"ok": true, ...} or {"ok": false, ...}.
+_VALIDATOR_SCRIPT = r"""
+import argparse, json, sys
+
+# Mirror what `vllm serve` does internally: FlexibleArgumentParser (which is
+# more lenient about dashes vs underscores) wrapped with make_arg_parser
+# (which adds engine + server + frontend args).
+parser = None
+try:
+    # Newer vLLM path
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
+except Exception:
+    try:
+        # Older fallback
+        from vllm.engine.arg_utils import FlexibleArgumentParser
+    except Exception:
+        FlexibleArgumentParser = argparse.ArgumentParser  # type: ignore
+
+try:
+    from vllm.entrypoints.openai.cli_args import make_arg_parser
+    parser = make_arg_parser(FlexibleArgumentParser(add_help=False))
+except Exception:
+    pass
+if parser is None:
+    try:
+        from vllm.engine.arg_utils import EngineArgs
+        parser = FlexibleArgumentParser(add_help=False)
+        EngineArgs.add_cli_args(parser)
+    except Exception as e:
+        print(json.dumps({"ok": False, "stage": "import", "error": f"{type(e).__name__}: {e}"}))
+        sys.exit(0)
+
+class _ArgError(Exception):
+    pass
+
+def _err(message):
+    raise _ArgError(message)
+
+parser.error = _err  # capture argparse errors instead of sys.exit(2)
+
+try:
+    raw = sys.stdin.read()
+    arglist = json.loads(raw)
+    ns = parser.parse_args(arglist)
+    print(json.dumps({"ok": True, "model": getattr(ns, "model", None)}))
+except _ArgError as e:
+    print(json.dumps({"ok": False, "stage": "parse", "error": str(e)}))
+except SystemExit as e:
+    print(json.dumps({"ok": False, "stage": "parse", "error": f"argparse exit {e.code}"}))
+except Exception as e:
+    print(json.dumps({"ok": False, "stage": "parse", "error": f"{type(e).__name__}: {e}"}))
+"""
+
+
+def _vllm_arg_list(key: str, model_def, catalog: Catalog) -> list[str]:
+    """Reconstruct the args list passed to `vllm serve` (without the positional model)."""
+    cmd = build_launch_command(key, model_def, catalog.defaults)
+    # build_launch_command yields:
+    #   ./launch-cluster.sh [--solo] -d exec vllm serve <repo> <args...>
+    # We just want the bits after `vllm serve <repo>`.
+    tokens = shlex.split(cmd)
+    if "serve" not in tokens:
+        return []
+    i = tokens.index("serve")
+    after = tokens[i + 1 :]  # repo, then args
+    if not after:
+        return []
+    args = after[1:]  # drop the repo
+    # EngineArgs expects --model=REPO rather than positional, so prepend it.
+    return [f"--model={after[0]}", *args]
+
+
+async def validate_launch(key: str, catalog: Catalog, settings: Settings) -> dict:
+    if key not in catalog.models:
+        return {"ok": False, "stage": "lookup", "error": f"unknown model: {key}"}
+    if not settings.spark1_host or not settings.spark1_user:
+        return {"ok": False, "stage": "config", "error": "spark1 not configured"}
+
+    model = catalog.models[key]
+    arg_list = _vllm_arg_list(key, model, catalog)
+    if not arg_list:
+        return {"ok": False, "stage": "build", "error": "failed to build args list"}
+
+    payload = json.dumps(arg_list).replace("'", "'\\''")
+    # Pipe the JSON args list to a here-doc Python invocation. The validator
+    # reads from stdin to avoid shell-escaping the args themselves.
+    cmd = (
+        f"echo '{payload}' | docker exec -i vllm_node python3 -c "
+        + shlex.quote(_VALIDATOR_SCRIPT)
+    )
+
+    rc, out, err = await ssh_run(settings.spark1_host, settings.spark1_user, cmd, settings, timeout=20)
+    if rc != 0 and not out.strip():
+        return {
+            "ok": False,
+            "stage": "ssh",
+            "error": err.strip() or f"rc={rc}",
+            "cmd_args": arg_list,
+            "launch_cmd": build_launch_command(key, model, catalog.defaults),
+        }
+    last = out.strip().splitlines()[-1] if out.strip() else ""
+    try:
+        result: dict[str, Any] = json.loads(last)
+    except json.JSONDecodeError:
+        result = {"ok": False, "stage": "decode", "error": "validator did not return JSON", "raw": out[-500:]}
+    result["cmd_args"] = arg_list
+    result["launch_cmd"] = build_launch_command(key, model, catalog.defaults)
+    return result

image/app/wol.py

@@ -0,0 +1,69 @@
+"""Wake-on-LAN.
+
+Two delivery paths, tried in order:
+
+  1. SSH into the other Spark and have IT broadcast — most reliable because the
+     packet originates from the same LAN subnet as the sleeping Spark.
+  2. Direct UDP broadcast from this container. May or may not work depending
+     on the StartOS container's network namespace.
+
+The DGX Spark's NIC must have WoL enabled in firmware/OS for either path to
+actually wake the box; this module just delivers the magic packet correctly.
+"""
+from __future__ import annotations
+import asyncio
+import re
+import socket
+
+from .config import Settings
+from .ssh import ssh_run
+
+
+_MAC_RE = re.compile(r"^[0-9a-fA-F]{2}([:-]?[0-9a-fA-F]{2}){5}$")
+
+
+def normalize_mac(mac: str) -> str:
+    mac = mac.strip().lower()
+    if not _MAC_RE.match(mac):
+        raise ValueError(f"invalid MAC address: {mac!r}")
+    return mac.replace("-", ":")
+
+
+def build_magic_packet(mac: str) -> bytes:
+    mac_bytes = bytes.fromhex(normalize_mac(mac).replace(":", ""))
+    return b"\xff" * 6 + mac_bytes * 16
+
+
+def send_local_broadcast(mac: str, broadcast: str = "255.255.255.255", port: int = 9) -> None:
+    """Send from THIS container. May not reach the LAN in some topologies."""
+    pkt = build_magic_packet(mac)
+    s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+    try:
+        s.setsockopt(socket.SOL_SOCKET, socket.SO_BROADCAST, 1)
+        s.sendto(pkt, (broadcast, port))
+        # Also send to port 7 (alternate WoL convention) for safety
+        s.sendto(pkt, (broadcast, 7))
+    finally:
+        s.close()
+
+
+async def send_via_peer(host: str, user: str, mac: str, settings: Settings) -> tuple[bool, str]:
+    """Use a different (reachable) Spark to send the WoL packet to its peer.
+
+    Uses Python 3 (always present on the Sparks for vLLM) to avoid depending on
+    wakeonlan / etherwake being installed.
+    """
+    normalized = normalize_mac(mac)
+    mac_hex = normalized.replace(":", "")
+    py = (
+        "python3 -c \""
+        "import socket; "
+        f"m=bytes.fromhex('{mac_hex}'); "
+        "s=socket.socket(socket.AF_INET, socket.SOCK_DGRAM); "
+        "s.setsockopt(socket.SOL_SOCKET, socket.SO_BROADCAST, 1); "
+        "s.sendto(b'\\xff'*6 + m*16, ('255.255.255.255', 9)); "
+        "s.sendto(b'\\xff'*6 + m*16, ('255.255.255.255', 7)); "
+        "print('sent')\""
+    )
+    rc, out, err = await ssh_run(host, user, py, settings, timeout=8)
+    return rc == 0 and "sent" in out, (err.strip() or out.strip() or f"rc={rc}")

image/models.yaml

@@ -30,6 +30,7 @@ models:
       - -tp=2
       - --distributed-executor-backend=ray
       - --max-model-len=32768
+      - --max-num-batched-tokens=16384
 
   gemma4:
     display_name: "Gemma 4 31B"
@@ -45,6 +46,7 @@ models:
     vllm_args:
       - --gpu-memory-utilization=0.8
       - --max-model-len=32768
+      - --max-num-batched-tokens=16384
       - --reasoning-parser=gemma4
       - --tool-call-parser=gemma4
       - --enable-auto-tool-choice
@@ -66,6 +68,7 @@ models:
     vllm_args:
       - --gpu-memory-utilization=0.85
       - --max-model-len=65536
+      - --max-num-batched-tokens=16384
       - --reasoning-parser=qwen3
       - --moe_backend=flashinfer_cutlass
       - --load-format=fastsafetensors

known-issues.md

@@ -20,6 +20,14 @@ The trick is the `docker run --rm alpine chown` — it runs as root inside the t
 
 This flag is Blackwell-specific. If vLLM in the container reports `unrecognized arguments: --moe_backend` or similar, edit `models.yaml` for `qwen36` and drop that flag. The swap UI does NOT auto-fallback in v0.1 — failure surfaces in the log stream.
 
+## Qwen3.6 Mamba block-size assertion (fixed in v0.6.0:1)
+
+Qwen3.6 uses a Mamba-attention hybrid that requires `--max-num-batched-tokens >= 2096`. vLLM's default is 2048, which trips `AssertionError: In Mamba cache align mode, block_size (2096) must be <= max_num_batched_tokens (2048)`. Fix: bake `--max-num-batched-tokens=16384` into the bundled qwen36 entry — matches the upstream qwen3.5-35b-a3b-fp8 recipe.
+
+## Multimodal token budget for vision models (fixed in v0.8.0:1)
+
+After the eugr/spark-vllm-docker update, vLLM became stricter about multimodal token budgets. Vision-capable models like Gemma 4 31B and Qwen3-VL crash at engine init with `ValueError: Chunked MM input disabled but max_tokens_per_mm_item (2496) is larger than max_num_batched_tokens (2048)`. Fix: bake `--max-num-batched-tokens=16384` into every model that has the `vision` capability. Now applied to qwen3-vl, gemma4, and qwen36 (which was already set for the Mamba issue).
+
 ## Two SSH paths to Spark 1 from the laptop
 
 `ssh <spark-user>@<spark-1-ip>` does NOT work from the laptop because the NVIDIA Sync ssh_config only has a Host entry for `<spark-1-host>.local`. Always use the `.local` hostname or `<spark-2-ip>`-style entries that ARE matched.

package/startos/actions/configureSparks.ts

@@ -76,6 +76,24 @@ 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,
+  }),
+  ngc_api_key: Value.text({
+    name: 'NGC API key (optional)',
+    description:
+      'NVIDIA NGC personal API key — needed to install NIM containers (Parakeet, Magpie, etc.) from nvcr.io. Get one free at https://ngc.nvidia.com/setup/personal-key. Stored only on this Start9 server; passed to docker as the NGC_API_KEY env var when installing NIM services.',
+    required: false,
+    default: null,
+    placeholder: 'starts with "nvapi-..."',
+    masked: true,
+  }),
 })
 
 export const configureSparks = sdk.Action.withInput(

package/startos/fileModels/sparkConfig.yaml.ts

@@ -14,6 +14,10 @@ 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(''),
+  // Optional NGC API key for pulling NIM containers from nvcr.io/nim/...
+  ngc_api_key: z.string().catch(''),
 })
 
 export type SparkConfig = z.infer<typeof sparkConfigSchema>

package/startos/main.ts

@@ -19,6 +19,8 @@ export const main = sdk.setupMain(async ({ effects }) => {
     magpie_host: '',
     magpie_user: '',
     magpie_container: '',
+    open_webui_url: '',
+    ngc_api_key: '',
   }
 
   return sdk.Daemons.of(effects).addDaemon('primary', {
@@ -47,6 +49,10 @@ export const main = sdk.setupMain(async ({ effects }) => {
         MAGPIE_USER: cfg.magpie_user,
         MAGPIE_CONTAINER: cfg.magpie_container,
         MODELS_OVERRIDES: '/data/models-overrides.yaml',
+        SERVICES_OVERRIDES: '/data/services-overrides.yaml',
+        CONNECTIVITY_LOG: '/data/connectivity.json',
+        OPEN_WEBUI_URL: cfg.open_webui_url,
+        NGC_API_KEY: cfg.ngc_api_key,
         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: '0.2.3:0',
+  version: '0.8.1:0',
   releaseNotes: {
     en_US:
-      'Per-model Advanced settings + downloaded-model catalog flow. Each card now has an Advanced button: max context tokens, GPU memory %, and optimization toggles (fastsafetensors, prefix caching, FP8 KV cache). After a download finishes, a dialog appears to add the model to the catalog with those same knobs as launch defaults. Custom models can be deleted. Overrides persist in /data/models-overrides.yaml and survive package updates.',
+      'v0.8.1: model weights can now be deleted from disk directly from the dashboard. Each model card shows whether the weights are present (with on-disk GB size) or not yet downloaded. When present and the model is NOT currently loaded, a small trash icon appears on the card; clicking it pops a confirmation showing how many GB will be freed and on which Spark(s), then runs `rm -rf` on the Hugging Face cache directory via SSH. Cluster-mode models are deleted from both Sparks; solo-mode from Spark 1 only. Safety rails: refuses to delete the currently-loaded model, refuses during an in-flight swap or download, and the catalog entry stays — you can always re-download. Disk status is probed once on dashboard load and re-checked every 60s.',
   },
   migrations: {
     up: async ({ effects }) => {},