feat(admin): LLM-/GPU-Status-Karte (read-only) — Plan-Schritt 2

- app/admin_llm.py: read-only Statusabfragen ohne sudo (docker ps, ollama ps,
  nvidia-smi, systemctl --user) mit sicheren Defaults bei fehlenden Tools.
- GET /api/admin/llm/status (require_admin): Backend, Modell, Backend-Status,
  geladene Ollama-Modelle, GPU-Auslastung, Gateway-Dienst-Status.
- Admin Status-Tab: LLM-Backend-Karte mit GPU-Balken.
- Tests: Auth-Gate + Antwortschema (160 grün).
- Doku §7.5: Status-Tab um LLM/GPU-Karte ergänzt.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/admin_llm.py

@@ -0,0 +1,115 @@
+"""Read-only Statusabfragen rund um das lokale LLM-Backend (fuer das Admin-Panel).
+
+Alles nur lesend, ohne sudo: `docker ps`, `ollama ps`, `nvidia-smi`, `systemctl --user`.
+Fehlende Tools oder Fehler fuehren zu sicheren Defaults (None/[]), nie zu Exceptions.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import shutil
+
+from app.config import settings
+
+
+async def _run(cmd: list[str], timeout: float = 6.0) -> str | None:
+    """Fuehrt ein Kommando aus (shell=False) und liefert stdout, oder None bei Fehler."""
+    if not shutil.which(cmd[0]):
+        return None
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.DEVNULL,
+            env={**os.environ, "XDG_RUNTIME_DIR": os.environ.get(
+                "XDG_RUNTIME_DIR", f"/run/user/{os.getuid()}")},
+        )
+        out, _ = await asyncio.wait_for(proc.communicate(), timeout=timeout)
+        if proc.returncode != 0:
+            return None
+        return out.decode("utf-8", "replace")
+    except (asyncio.TimeoutError, OSError):
+        return None
+
+
+def _detect_backend(base_url: str) -> str:
+    if ":11434" in base_url:
+        return "ollama"
+    if ":8001" in base_url:
+        return "llamacpp"
+    return "unknown"
+
+
+async def _llamacpp_running(container: str = "va_llm") -> bool:
+    out = await _run(["docker", "ps", "--filter", f"name=^{container}$", "--format", "{{.Names}}"])
+    return bool(out and container in out)
+
+
+async def _ollama_loaded() -> tuple[bool, list[dict]]:
+    """(dienst_erreichbar, [geladene Modelle])."""
+    out = await _run(["ollama", "ps"])
+    if out is None:
+        return False, []
+    models: list[dict] = []
+    lines = [ln for ln in out.splitlines() if ln.strip()]
+    for ln in lines[1:]:  # Kopfzeile ueberspringen
+        # Spalten sind durch 2+ Leerzeichen getrennt: NAME ID SIZE PROCESSOR CONTEXT UNTIL
+        import re
+        cols = re.split(r"\s{2,}", ln.strip())
+        if cols:
+            models.append({
+                "name": cols[0],
+                "size": cols[2] if len(cols) > 2 else "",
+                "processor": cols[3] if len(cols) > 3 else "",
+            })
+    return True, models
+
+
+async def _gpus() -> list[dict]:
+    out = await _run([
+        "nvidia-smi",
+        "--query-gpu=index,memory.used,memory.total",
+        "--format=csv,noheader,nounits",
+    ])
+    if out is None:
+        return []
+    gpus: list[dict] = []
+    for ln in out.splitlines():
+        parts = [p.strip() for p in ln.split(",")]
+        if len(parts) == 3 and parts[0].isdigit():
+            used, total = int(parts[1]), int(parts[2])
+            gpus.append({
+                "index": int(parts[0]),
+                "used_mib": used,
+                "total_mib": total,
+                "percent": round(used / total * 100) if total else 0,
+            })
+    return gpus
+
+
+async def _gateway_service_active() -> bool:
+    out = await _run(["systemctl", "--user", "is-active", "voice-assistant.service"], timeout=4.0)
+    return bool(out and out.strip() == "active")
+
+
+async def llm_status() -> dict:
+    """Aggregierter, read-only LLM-/System-Status fuer das Admin-Panel."""
+    base_url = settings.local_llm_base_url
+    backend = _detect_backend(base_url)
+    llamacpp, (ollama_reachable, ollama_models), gpus, gw = await asyncio.gather(
+        _llamacpp_running(),
+        _ollama_loaded(),
+        _gpus(),
+        _gateway_service_active(),
+    )
+    return {
+        "backend": backend,
+        "model": settings.local_llm_model,
+        "base_url": base_url,
+        "llamacpp_running": llamacpp,
+        "ollama_reachable": ollama_reachable,
+        "ollama_loaded": ollama_models,
+        "gpus": gpus,
+        "gateway_service_active": gw,
+    }

app/api/admin.py

@@ -6,6 +6,7 @@ from fastapi import APIRouter, Depends, HTTPException, Request, WebSocket, WebSo
 from fastapi.responses import FileResponse
 from pydantic import BaseModel
 
+from app.admin_llm import llm_status
 from app.auth import is_admin_user, require_admin, require_admin_or_user
 from app.config import settings
 from app.dependencies import get_store
@@ -187,6 +188,14 @@ async def export_db():
     )
 
 
+# ── LLM-/System-Status (read-only) ──────────────────────────────────────────
+
+@router.get("/admin/llm/status", dependencies=[Depends(require_admin)])
+async def get_llm_status():
+    """Read-only Status: aktives Backend, Modell, GPU-Auslastung, Dienste."""
+    return await llm_status()
+
+
 # ── Aussprache-Lexikon CRUD ─────────────────────────────────────────────────
 
 def _read_pronunciation(lang: str) -> dict:

app/web/app.js

@@ -956,13 +956,56 @@ async function loadEmergencyEvents() {
 // ════════════════════════════════════════
 // TAB: STATUS
 // ════════════════════════════════════════
+// LLM-/GPU-Status-Karte (read-only) für den Status-Tab.
+function renderLlmCard(llm) {
+  if (!llm) return "";
+  const backendLabel = { ollama: "Ollama", llamacpp: "llama.cpp", unknown: "—" }[llm.backend] || llm.backend;
+  const dot = (on) => `<span class="h-2.5 w-2.5 rounded-full ${on ? "bg-emerald-500" : "bg-slate-400"}"></span>`;
+  const row = (label, val) =>
+    `<dt class="text-slate-500 dark:text-slate-400">${label}</dt>
+     <dd class="font-mono text-xs text-slate-700 dark:text-slate-300">${escHtml(val ?? "—")}</dd>`;
+
+  const loaded = (llm.ollama_loaded || []).map((m) =>
+    `${escHtml(m.name)} <span class="text-slate-400">(${escHtml(m.processor || "")})</span>`
+  ).join(", ") || "—";
+
+  const gpuBars = (llm.gpus || []).map((g) => `
+    <div class="flex items-center gap-3">
+      <span class="w-10 shrink-0 text-xs text-slate-500 dark:text-slate-400">GPU ${g.index}</span>
+      <div class="flex-1 h-3 bg-slate-100 dark:bg-slate-700 rounded-full overflow-hidden">
+        <div class="h-full ${g.percent > 85 ? "bg-red-500" : "bg-blue-500 dark:bg-blue-600"} rounded-full" style="width:${g.percent}%"></div>
+      </div>
+      <span class="w-28 text-right text-xs font-mono text-slate-500 dark:text-slate-400">${g.used_mib} / ${g.total_mib} MiB</span>
+    </div>`).join("") || '<p class="text-xs text-slate-400">keine GPU-Daten (nvidia-smi)</p>';
+
+  return `
+    <div class="bg-white dark:bg-slate-800 rounded-xl border border-slate-200 dark:border-slate-700 p-4">
+      <div class="flex items-center gap-2 mb-3">
+        ${dot(true)}<h3 class="font-semibold text-sm">LLM-Backend</h3>
+        <span class="text-xs rounded-full bg-blue-100 dark:bg-blue-900/40 text-blue-700 dark:text-blue-300 px-2 py-0.5 font-medium">${escHtml(backendLabel)}</span>
+      </div>
+      <dl class="grid grid-cols-2 gap-x-8 gap-y-1.5 text-sm mb-3">
+        ${row("Modell", llm.model)}
+        ${row("URL", llm.base_url)}
+        ${row("Ollama erreichbar", llm.ollama_reachable ? "ja" : "nein")}
+        ${row("llama.cpp läuft", llm.llamacpp_running ? "ja" : "nein")}
+        ${row("Geladene Modelle", loaded)}
+        ${row("Gateway-Dienst", llm.gateway_service_active ? "aktiv (systemd)" : "Vordergrund/aus")}
+      </dl>
+      <div class="space-y-2 pt-2 border-t border-slate-100 dark:border-slate-700">
+        ${gpuBars}
+      </div>
+    </div>`;
+}
+
 async function loadStatus() {
   const container = $("#status-content");
   container.innerHTML = '<p class="text-sm text-slate-400">lade …</p>';
   try {
-    const [cfg, met] = await Promise.all([
+    const [cfg, met, llm] = await Promise.all([
       fetch("/api/config").then((r) => r.json()),
       fetch("/api/metrics").then((r) => r.json()),
+      adminFetch("/api/admin/llm/status"),
     ]);
     const route    = cfg.default_route || {};
     const counters = met.counters      || {};
@@ -1000,6 +1043,8 @@ async function loadStatus() {
         </dl>
       </div>
 
+      ${renderLlmCard(llm)}
+
       <!-- Laufzeit-Metriken -->
       <div class="bg-white dark:bg-slate-800 rounded-xl border border-slate-200 dark:border-slate-700 p-4">
         <h3 class="font-semibold text-sm mb-3">Laufzeit-Metriken</h3>

app/web/index.html

@@ -250,6 +250,6 @@
     <div id="status" class="w-full max-w-3xl mx-auto mt-1 min-h-[1rem] text-xs text-slate-500 dark:text-slate-400"></div>
   </footer>
 
-  <script src="/app.js?v=28"></script>
+  <script src="/app.js?v=29"></script>
 </body>
 </html>