Mission Control v1 – laeuft stabil

.gitignore

@@ -0,0 +1,3 @@
+.venv/
+__pycache__/
+*.pyc

README.md

@@ -0,0 +1,85 @@
+# Mission Control
+
+Eine schlanke Steuerzentrale für deinen lokalen `llama-swap`-Stack auf dem Bosgame M5.
+Ein FastAPI-Backend + ein HTML-Dashboard. Kein Build-Schritt, keine Datenbank.
+
+## Was sie kann
+
+- **Modelle & Ports** sehen — liest `llama-swap` (`/running`, `/v1/models`) + deine `config.yaml`
+- **Modell holen** — lädt eine GGUF-Datei von HuggingFace (`hf download`) als Hintergrund-Job mit Live-Log
+- **Einpflegen** — schreibt das Modell automatisch in deine `config.yaml`; `llama-swap` lädt mit `-watch-config` neu
+- **Wartung** — Container/Toolbox aktualisieren, Modelle aus dem Speicher werfen
+- **Schnelltest** — Chat-Box, um ein Modell zu wecken und zu prüfen
+
+Was sie **bewusst nicht** macht: Chat-Logs, Inferenz-Monitoring im Detail — dafür hat `llama-swap`
+schon `/ui` und `/log`. Mission Control ergänzt nur, was fehlt.
+
+## Voraussetzungen
+
+- Python 3.11+
+- `hf` CLI installiert (`pip install -U "huggingface_hub[cli]"`)
+- ein laufendes `llama-swap` — gestartet **mit `-watch-config`**, sonst greift das Auto-Einpflegen nicht
+
+## Installation
+
+```bash
+sudo mkdir -p /opt/mission-control && sudo chown $USER /opt/mission-control
+cp -r app.py static /opt/mission-control/
+cd /opt/mission-control
+python3 -m venv .venv && . .venv/bin/activate
+pip install -r requirements.txt
+uvicorn app:app --host 0.0.0.0 --port 9000
+```
+
+Dann im Browser im LAN: `http://<bosgame-ip>:9000`
+
+### Als Dienst (Autostart)
+
+Pfade in `mission-control.service` anpassen, dann:
+
+```bash
+sudo cp mission-control.service /etc/systemd/system/
+sudo systemctl daemon-reload
+sudo systemctl enable --now mission-control
+```
+
+## Konfiguration (Umgebungsvariablen)
+
+| Variable | Default | Zweck |
+|---|---|---|
+| `MC_LLAMA_SWAP_URL` | `http://127.0.0.1:8080` | wo `llama-swap` lauscht |
+| `MC_CONFIG_PATH` | `/etc/llama-swap/config.yaml` | die `llama-swap`-Config |
+| `MC_MODELS_DIR` | `/srv/models` | wohin GGUFs geladen werden |
+| `MC_CMD_TEMPLATE` | siehe unten | Startbefehl pro Modell |
+| `MC_UPDATE_CMD` | _(leer)_ | Befehl für „Container aktualisieren" |
+| `MC_DEFAULT_TTL` | `300` | Sekunden bis Auto-Unload |
+| `MC_TOKEN` | _(leer)_ | optionales Zugriffs-Token |
+
+### Wichtig: `MC_CMD_TEMPLATE` an deinen Start anpassen
+
+Das ist der Befehl, der pro Modell in die `config.yaml` geschrieben wird. `{model}` und `{ctx}`
+werden von Mission Control ersetzt, **`${PORT}` bleibt stehen** (das ersetzt `llama-swap` selbst).
+
+Direkt auf dem Host (llama-server im PATH):
+```
+llama-server -m {model} --host 127.0.0.1 --port ${PORT} -c {ctx} -ngl 999 -fa 1 --no-mmap
+```
+
+Über den kyuz0-Container (distrobox) — Beispiel, an deinen Toolbox-Namen anpassen:
+```
+distrobox enter llama-vulkan-radv -- llama-server -m {model} --host 127.0.0.1 --port ${PORT} -c {ctx} -ngl 999 -fa 1 --no-mmap
+```
+
+> `-fa 1` (Flash Attention) und `--no-mmap` sind auf Strix Halo Pflicht, sonst drohen Crashes/Slowdowns.
+
+## ⚠️ Sicherheit
+
+Mission Control führt Shell-Befehle aus (Downloads, Updates) und schreibt deine Config.
+**Niemals offen ins Internet hängen.** Betrieb nur im vertrauenswürdigen LAN. Wenn du
+trotzdem etwas Schutz willst, setz `MC_TOKEN` und trag es oben rechts im Dashboard ein.
+Für echten Remote-Zugriff: per SSH-Tunnel oder Tailscale, nicht per Portfreigabe.
+
+## API (falls du es skripten willst)
+
+`GET /api/status` · `POST /api/download` · `POST /api/register` · `POST /api/unload[?model=]`
+· `POST /api/update` · `POST /api/chat` · `GET /api/jobs` · `GET /api/jobs/{id}`

app.py

@@ -0,0 +1,279 @@
+"""
+Mission Control - eine schlanke Steuerzentrale fuer einen lokalen llama-swap Stack.
+
+Was sie macht:
+  - zeigt konfigurierte + laufende Modelle und ihre Ports (liest llama-swap /running + /v1/models)
+  - laedt neue GGUF-Modelle von HuggingFace (hf download, als Hintergrund-Job)
+  - pflegt ein heruntergeladenes Modell automatisch in die llama-swap config.yaml ein
+  - stoesst Updates an (Container / Toolbox refresh)
+  - laedt Modelle aus dem Speicher (unload) und hat einen kleinen Chat-Test
+
+Bewusst KISS: ein File, In-Memory Jobs, keine Datenbank.
+"""
+
+import os
+import shlex
+import subprocess
+import threading
+import time
+import uuid
+from pathlib import Path
+
+import httpx
+from fastapi import Depends, FastAPI, Header, HTTPException
+from fastapi.responses import FileResponse, JSONResponse
+from fastapi.staticfiles import StaticFiles
+from pydantic import BaseModel
+from ruamel.yaml import YAML
+from ruamel.yaml.scalarstring import LiteralScalarString
+
+# ---------------------------------------------------------------------------
+# Konfiguration (alles ueber Umgebungsvariablen ueberschreibbar)
+# ---------------------------------------------------------------------------
+LLAMA_SWAP_URL = os.environ.get("MC_LLAMA_SWAP_URL", "http://127.0.0.1:8080").rstrip("/")
+CONFIG_PATH = Path(os.environ.get("MC_CONFIG_PATH", "/etc/llama-swap/config.yaml"))
+MODELS_DIR = Path(os.environ.get("MC_MODELS_DIR", "/srv/models"))
+# Befehl, der zum Starten eines Modells in die config.yaml geschrieben wird.
+# {model} = Pfad zur GGUF-Datei, {ctx} = Kontextlaenge, ${PORT} bleibt fuer llama-swap stehen.
+# WICHTIG: an deinen Container-/llama-server-Aufruf anpassen (siehe README).
+CMD_TEMPLATE = os.environ.get(
+    "MC_CMD_TEMPLATE",
+    "llama-server -m {model} --host 127.0.0.1 --port ${PORT} "
+    "-c {ctx} -ngl 999 -fa 1 --no-mmap",
+)
+# Befehl fuer "Container/Toolbox aktualisieren". Standard: kyuz0 refresh-Skript.
+UPDATE_CMD = os.environ.get("MC_UPDATE_CMD", "")
+DEFAULT_TTL = int(os.environ.get("MC_DEFAULT_TTL", "300"))
+TOKEN = os.environ.get("MC_TOKEN", "")  # leer = keine Auth (nur LAN!)
+
+yaml = YAML()
+yaml.preserve_quotes = True
+
+app = FastAPI(title="Mission Control")
+
+# ---------------------------------------------------------------------------
+# Mini Job-System (Hintergrund-Prozesse mit Live-Log)
+# ---------------------------------------------------------------------------
+JOBS: dict[str, dict] = {}
+_LOG_CAP = 400
+
+
+def _run_job(job_id: str, args: list[str], env: dict | None = None):
+    job = JOBS[job_id]
+    job["state"] = "running"
+    try:
+        proc = subprocess.Popen(
+            args,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            text=True,
+            bufsize=1,
+            env={**os.environ, **(env or {})},
+        )
+        for line in proc.stdout:  # type: ignore[union-attr]
+            job["log"].append(line.rstrip("\n"))
+            if len(job["log"]) > _LOG_CAP:
+                del job["log"][0]
+        proc.wait()
+        job["returncode"] = proc.returncode
+        job["state"] = "done" if proc.returncode == 0 else "failed"
+    except Exception as exc:  # noqa: BLE001
+        job["log"].append(f"[mission-control] Fehler: {exc}")
+        job["state"] = "failed"
+        job["returncode"] = -1
+    job["finished_at"] = time.time()
+
+
+def start_job(args: list[str], label: str, env: dict | None = None) -> str:
+    job_id = uuid.uuid4().hex[:12]
+    JOBS[job_id] = {
+        "id": job_id,
+        "label": label,
+        "state": "queued",
+        "log": [f"$ {' '.join(shlex.quote(a) for a in args)}"],
+        "returncode": None,
+        "started_at": time.time(),
+        "finished_at": None,
+    }
+    threading.Thread(target=_run_job, args=(job_id, args, env), daemon=True).start()
+    return job_id
+
+
+# ---------------------------------------------------------------------------
+# Auth (optional)
+# ---------------------------------------------------------------------------
+def auth(x_mc_token: str = Header(default="")):
+    if TOKEN and x_mc_token != TOKEN:
+        raise HTTPException(status_code=401, detail="Falsches oder fehlendes Token.")
+
+
+# ---------------------------------------------------------------------------
+# llama-swap Helfer
+# ---------------------------------------------------------------------------
+def _swap_get(path: str):
+    with httpx.Client(timeout=5.0) as c:
+        r = c.get(f"{LLAMA_SWAP_URL}{path}")
+        r.raise_for_status()
+        return r.json()
+
+
+def read_config() -> dict:
+    if not CONFIG_PATH.exists():
+        return {"models": {}}
+    with CONFIG_PATH.open("r", encoding="utf-8") as f:
+        data = yaml.load(f) or {}
+    if "models" not in data or data["models"] is None:
+        data["models"] = {}
+    return data
+
+
+# ---------------------------------------------------------------------------
+# Request-Modelle
+# ---------------------------------------------------------------------------
+class DownloadReq(BaseModel):
+    repo: str
+    file: str
+    subdir: str | None = None
+
+
+class RegisterReq(BaseModel):
+    alias: str
+    model_path: str
+    ctx: int = 8192
+    ttl: int | None = None
+
+
+class ChatReq(BaseModel):
+    model: str
+    message: str
+
+
+# ---------------------------------------------------------------------------
+# API
+# ---------------------------------------------------------------------------
+@app.get("/api/status", dependencies=[Depends(auth)])
+def status():
+    cfg = read_config()
+    configured = {}
+    for name, spec in (cfg.get("models") or {}).items():
+        spec = spec or {}
+        configured[name] = {
+            "name": name,
+            "ttl": spec.get("ttl", cfg.get("globalTTL", 0)),
+            "cmd": str(spec.get("cmd", "")).strip(),
+            "state": "idle",
+            "port": None,
+        }
+    swap_ok = True
+    try:
+        running = _swap_get("/running")
+        items = running.get("running", running) if isinstance(running, dict) else running
+        for item in items or []:
+            mid = item.get("model") or item.get("id") or item.get("name")
+            if mid in configured:
+                configured[mid]["state"] = item.get("state", "running")
+                configured[mid]["port"] = item.get("port")
+            elif mid:
+                configured[mid] = {
+                    "name": mid, "ttl": None, "cmd": "",
+                    "state": item.get("state", "running"), "port": item.get("port"),
+                }
+    except Exception:  # noqa: BLE001
+        swap_ok = False
+    return {
+        "swap_ok": swap_ok,
+        "swap_url": LLAMA_SWAP_URL,
+        "config_path": str(CONFIG_PATH),
+        "models_dir": str(MODELS_DIR),
+        "models": list(configured.values()),
+    }
+
+
+@app.post("/api/download", dependencies=[Depends(auth)])
+def download(req: DownloadReq):
+    sub = req.subdir or req.repo.split("/")[-1]
+    target = MODELS_DIR / sub
+    target.mkdir(parents=True, exist_ok=True)
+    args = ["hf", "download", req.repo, req.file, "--local-dir", str(target)]
+    job_id = start_job(args, f"download {req.repo}/{req.file}",
+                       env={"HF_XET_HIGH_PERFORMANCE": "1"})
+    JOBS[job_id]["result_path"] = str(target / req.file)
+    return {"job_id": job_id, "expected_path": str(target / req.file)}
+
+
+@app.post("/api/register", dependencies=[Depends(auth)])
+def register(req: RegisterReq):
+    if not Path(req.model_path).exists():
+        raise HTTPException(404, f"Datei nicht gefunden: {req.model_path}")
+    cfg = read_config()
+    cmd = CMD_TEMPLATE.replace("{model}", req.model_path).replace("{ctx}", str(req.ctx))
+    cfg["models"][req.alias] = {
+        "cmd": LiteralScalarString(cmd + "\n"),
+        "ttl": req.ttl if req.ttl is not None else DEFAULT_TTL,
+    }
+    CONFIG_PATH.parent.mkdir(parents=True, exist_ok=True)
+    with CONFIG_PATH.open("w", encoding="utf-8") as f:
+        yaml.dump(cfg, f)
+    return {"ok": True, "alias": req.alias,
+            "note": "In config.yaml geschrieben. llama-swap mit -watch-config laedt automatisch neu."}
+
+
+@app.post("/api/unload", dependencies=[Depends(auth)])
+def unload(model: str | None = None):
+    path = f"/api/models/unload/{model}" if model else "/api/models/unload"
+    try:
+        with httpx.Client(timeout=10.0) as c:
+            r = c.post(f"{LLAMA_SWAP_URL}{path}")
+        return {"ok": r.status_code < 400, "status": r.status_code}
+    except Exception as exc:  # noqa: BLE001
+        raise HTTPException(502, f"llama-swap nicht erreichbar: {exc}")
+
+
+@app.post("/api/update", dependencies=[Depends(auth)])
+def update():
+    if not UPDATE_CMD:
+        raise HTTPException(400, "Kein Update-Befehl gesetzt (MC_UPDATE_CMD).")
+    job_id = start_job(shlex.split(UPDATE_CMD), "update containers")
+    return {"job_id": job_id}
+
+
+@app.post("/api/chat", dependencies=[Depends(auth)])
+def chat(req: ChatReq):
+    payload = {"model": req.model, "messages": [{"role": "user", "content": req.message}]}
+    try:
+        with httpx.Client(timeout=120.0) as c:
+            r = c.post(f"{LLAMA_SWAP_URL}/v1/chat/completions", json=payload)
+            r.raise_for_status()
+            data = r.json()
+        return {"reply": data["choices"][0]["message"]["content"]}
+    except Exception as exc:  # noqa: BLE001
+        raise HTTPException(502, f"Anfrage fehlgeschlagen: {exc}")
+
+
+@app.get("/api/jobs/{job_id}", dependencies=[Depends(auth)])
+def job_status(job_id: str):
+    job = JOBS.get(job_id)
+    if not job:
+        raise HTTPException(404, "Job nicht gefunden.")
+    return job
+
+
+@app.get("/api/jobs", dependencies=[Depends(auth)])
+def jobs_list():
+    return sorted(JOBS.values(), key=lambda j: j["started_at"], reverse=True)[:20]
+
+
+# ---------------------------------------------------------------------------
+# Statisches UI
+# ---------------------------------------------------------------------------
+@app.get("/")
+def index():
+    return FileResponse(Path(__file__).parent / "static" / "index.html")
+
+
+app.mount("/static", StaticFiles(directory=Path(__file__).parent / "static"), name="static")
+
+
+@app.exception_handler(HTTPException)
+def _http_exc(_req, exc: HTTPException):
+    return JSONResponse(status_code=exc.status_code, content={"error": exc.detail})

mission-control.service

@@ -0,0 +1,26 @@
+[Unit]
+Description=Mission Control (lokaler LLM-Stack)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=tobi
+WorkingDirectory=/opt/mission-control
+# Pfade an deinen Stack anpassen:
+Environment=MC_LLAMA_SWAP_URL=http://127.0.0.1:8080
+Environment=MC_CONFIG_PATH=/etc/llama-swap/config.yaml
+Environment=MC_MODELS_DIR=/srv/models
+Environment=MC_DEFAULT_TTL=300
+# Befehl zum Starten eines Modells (in config.yaml geschrieben). ${PORT} bleibt stehen!
+Environment=MC_CMD_TEMPLATE=llama-server -m {model} --host 127.0.0.1 --port ${PORT} -c {ctx} -ngl 999 -fa 1 --no-mmap
+# Update-Befehl (z.B. kyuz0 refresh oder podman pull):
+Environment=MC_UPDATE_CMD=/opt/amd-strix-halo-toolboxes/refresh-toolboxes.sh all
+# Optionales Token fuer minimale Absicherung im LAN:
+# Environment=MC_TOKEN=einlangesgeheimnis
+ExecStart=/opt/mission-control/.venv/bin/uvicorn app:app --host 0.0.0.0 --port 9000
+Restart=on-failure
+RestartSec=3
+
+[Install]
+WantedBy=multi-user.target

requirements.txt

@@ -0,0 +1,4 @@
+fastapi>=0.110
+uvicorn[standard]>=0.29
+httpx>=0.27
+ruamel.yaml>=0.18

static/index.html

@@ -0,0 +1,248 @@
+<!DOCTYPE html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Mission Control</title>
+<style>
+  :root{
+    --bg:#0d1117; --panel:#151b23; --panel2:#1b222c; --line:rgba(255,255,255,.08);
+    --line2:rgba(255,255,255,.14); --tx:#d7dee7; --mut:#8b97a5;
+    --on:#46c06a; --warn:#e0a32e; --err:#e5534b; --act:#4493e0;
+    --mono:ui-monospace,"SF Mono",Menlo,Consolas,monospace;
+    --sans:system-ui,-apple-system,Segoe UI,Roboto,sans-serif;
+  }
+  *{box-sizing:border-box}
+  body{margin:0;background:var(--bg);color:var(--tx);font-family:var(--sans);font-size:15px;line-height:1.5}
+  a{color:var(--act)}
+  .wrap{max-width:1040px;margin:0 auto;padding:0 20px 64px}
+  header{display:flex;align-items:center;gap:16px;padding:20px 0 18px;border-bottom:1px solid var(--line);
+    position:sticky;top:0;background:var(--bg);z-index:5;flex-wrap:wrap}
+  .brand{font-weight:600;letter-spacing:.2px;font-size:18px}
+  .brand b{color:var(--on)}
+  .pill{display:inline-flex;align-items:center;gap:8px;font-family:var(--mono);font-size:12.5px;
+    padding:5px 11px;border:1px solid var(--line);border-radius:999px;color:var(--mut);background:var(--panel)}
+  .dot{width:8px;height:8px;border-radius:50%;background:var(--mut)}
+  .dot.on{background:var(--on);box-shadow:0 0 0 0 rgba(70,192,106,.5);animation:pulse 2.2s infinite}
+  .dot.off{background:var(--err)}
+  @keyframes pulse{0%{box-shadow:0 0 0 0 rgba(70,192,106,.45)}70%{box-shadow:0 0 0 7px rgba(70,192,106,0)}100%{box-shadow:0 0 0 0 rgba(70,192,106,0)}}
+  .spacer{flex:1}
+  .tokin{font-family:var(--mono);font-size:12.5px;background:var(--panel);border:1px solid var(--line);
+    color:var(--tx);border-radius:8px;padding:6px 9px;width:130px}
+  h2{font-size:14px;font-weight:600;letter-spacing:.4px;text-transform:uppercase;color:var(--mut);margin:0 0 12px}
+  .grid{display:grid;grid-template-columns:1fr 1fr;gap:18px;margin-top:22px}
+  @media(max-width:780px){.grid{grid-template-columns:1fr}}
+  .card{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:18px}
+  .card.full{grid-column:1/-1}
+  table{width:100%;border-collapse:collapse;font-size:14px}
+  th{text-align:left;font-weight:500;color:var(--mut);font-size:12px;text-transform:uppercase;
+    letter-spacing:.4px;padding:0 10px 9px}
+  td{padding:11px 10px;border-top:1px solid var(--line)}
+  .mid{font-family:var(--mono);font-size:13px;color:#e8eef5}
+  .badge{font-family:var(--mono);font-size:11.5px;padding:3px 9px;border-radius:6px;display:inline-block}
+  .b-run{background:rgba(70,192,106,.14);color:var(--on)}
+  .b-idle{background:rgba(139,151,165,.14);color:var(--mut)}
+  .b-load{background:rgba(224,163,46,.16);color:var(--warn)}
+  .port{font-family:var(--mono);color:var(--mut);font-size:13px}
+  .empty{color:var(--mut);font-size:14px;padding:14px 4px}
+  label{display:block;font-size:12.5px;color:var(--mut);margin:0 0 5px}
+  input,textarea,select{width:100%;background:var(--panel2);border:1px solid var(--line);color:var(--tx);
+    border-radius:8px;padding:9px 11px;font-family:var(--mono);font-size:13px;margin-bottom:12px}
+  textarea{resize:vertical;min-height:64px;font-family:var(--sans)}
+  .row{display:flex;gap:10px}.row>div{flex:1}
+  button{font-family:var(--sans);font-size:13.5px;font-weight:500;border:1px solid var(--line2);
+    background:var(--panel2);color:var(--tx);border-radius:8px;padding:9px 15px;cursor:pointer}
+  button:hover{border-color:var(--act)}
+  button.primary{background:var(--act);border-color:var(--act);color:#fff}
+  button.primary:hover{filter:brightness(1.08)}
+  button.ghost{padding:5px 11px;font-size:12.5px}
+  button:disabled{opacity:.5;cursor:not-allowed}
+  .reply{margin-top:12px;background:var(--panel2);border:1px solid var(--line);border-radius:8px;
+    padding:12px;white-space:pre-wrap;font-size:14px;min-height:20px;color:var(--tx)}
+  .log{font-family:var(--mono);font-size:12px;line-height:1.65;background:#0a0e13;border:1px solid var(--line);
+    border-radius:8px;padding:12px;max-height:240px;overflow:auto;white-space:pre-wrap;color:#aeb9c4}
+  .job{border:1px solid var(--line);border-radius:8px;margin-bottom:8px;overflow:hidden}
+  .job-h{display:flex;align-items:center;gap:10px;padding:10px 12px;cursor:pointer}
+  .job-h .mid{flex:1}
+  .toast{position:fixed;bottom:20px;left:50%;transform:translateX(-50%);background:#1b222c;
+    border:1px solid var(--line2);border-radius:10px;padding:11px 16px;font-size:13.5px;
+    opacity:0;transition:.25s;pointer-events:none;max-width:90vw}
+  .toast.show{opacity:1}
+  .toast.err{border-color:var(--err);color:#ffb4ae}
+  .hint{font-size:12px;color:var(--mut);margin:-4px 0 12px}
+  .mono-sm{font-family:var(--mono);font-size:11.5px;color:var(--mut)}
+</style>
+</head>
+<body>
+<div class="wrap">
+  <header>
+    <span class="brand">Mission <b>Control</b></span>
+    <span class="pill"><span id="hdot" class="dot"></span><span id="hlabel">verbinde…</span></span>
+    <span class="spacer"></span>
+    <input id="token" class="tokin" placeholder="Token (optional)" autocomplete="off">
+  </header>
+
+  <div class="grid">
+    <!-- MODELLE -->
+    <div class="card full">
+      <h2>Modelle &amp; Ports</h2>
+      <table>
+        <thead><tr><th>Modell</th><th>Status</th><th>Port</th><th style="text-align:right">Aktion</th></tr></thead>
+        <tbody id="models"></tbody>
+      </table>
+      <div id="models-empty" class="empty" style="display:none">Noch keine Modelle konfiguriert — zieh dir unten eins rein. 👇</div>
+    </div>
+
+    <!-- DOWNLOAD -->
+    <div class="card">
+      <h2>Modell holen</h2>
+      <label>HuggingFace-Repo</label>
+      <input id="dl-repo" placeholder="unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF">
+      <label>Datei (GGUF)</label>
+      <input id="dl-file" placeholder="Q4_K_M/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf">
+      <button class="primary" onclick="pull()">Modell herunterladen</button>
+      <div id="register-box" style="display:none;margin-top:16px;border-top:1px solid var(--line);padding-top:14px">
+        <h2>Einpflegen</h2>
+        <div class="row">
+          <div><label>Alias</label><input id="rg-alias"></div>
+          <div><label>Kontext</label><input id="rg-ctx" value="8192"></div>
+        </div>
+        <input id="rg-path" class="mono-sm" readonly>
+        <button class="primary" onclick="register()">In Config eintragen</button>
+      </div>
+    </div>
+
+    <!-- WARTUNG + TEST -->
+    <div class="card">
+      <h2>Wartung</h2>
+      <button onclick="update()">Container aktualisieren</button>
+      <button onclick="unloadAll()">Alles aus dem Speicher</button>
+      <div class="hint" style="margin-top:10px">Update-Befehl wird per <span class="mono-sm">MC_UPDATE_CMD</span> gesetzt.</div>
+      <h2 style="margin-top:18px">Schnelltest</h2>
+      <select id="chat-model"></select>
+      <textarea id="chat-msg" placeholder="Schreib was, um ein Modell zu wecken…"></textarea>
+      <button class="primary" onclick="sendChat()" id="chat-btn">Senden</button>
+      <div id="chat-reply" class="reply" style="display:none"></div>
+    </div>
+
+    <!-- AKTIVITAET -->
+    <div class="card full">
+      <h2>Aktivität</h2>
+      <div id="jobs"></div>
+      <div id="jobs-empty" class="empty">Noch nichts losgemacht.</div>
+    </div>
+  </div>
+</div>
+<div id="toast" class="toast"></div>
+
+<script>
+const $ = s => document.querySelector(s);
+let TOKEN = localStorage.getItem("mc_token") || "";
+$("#token").value = TOKEN;
+$("#token").addEventListener("change", e => { TOKEN = e.target.value.trim(); localStorage.setItem("mc_token", TOKEN); refresh(); });
+
+function hdr(){ return TOKEN ? {"Content-Type":"application/json","X-MC-Token":TOKEN} : {"Content-Type":"application/json"}; }
+async function api(path, opts={}){
+  const r = await fetch(path, {headers: hdr(), ...opts});
+  const data = await r.json().catch(()=>({}));
+  if(!r.ok) throw new Error(data.error || ("HTTP "+r.status));
+  return data;
+}
+let _tt;
+function toast(msg, err=false){
+  const t = $("#toast"); t.textContent = msg; t.className = "toast show" + (err?" err":"");
+  clearTimeout(_tt); _tt = setTimeout(()=>t.className="toast",3200);
+}
+
+function badge(state){
+  if(state==="running" || state==="ready") return '<span class="badge b-run">geladen</span>';
+  if(state==="loading" || state==="starting") return '<span class="badge b-load">lädt…</span>';
+  return '<span class="badge b-idle">bereit</span>';
+}
+
+async function refresh(){
+  let s;
+  try{ s = await api("/api/status"); }
+  catch(e){ $("#hdot").className="dot off"; $("#hlabel").textContent="Backend nicht erreichbar"; return; }
+  const ok = s.swap_ok;
+  $("#hdot").className = "dot " + (ok?"on":"off");
+  $("#hlabel").textContent = (ok?"llama-swap online · ":"llama-swap offline · ") + s.swap_url.replace(/^https?:\/\//,"");
+
+  const tb = $("#models"); tb.innerHTML = "";
+  $("#models-empty").style.display = s.models.length ? "none" : "block";
+  const sel = $("#chat-model"); const cur = sel.value; sel.innerHTML = "";
+  for(const m of s.models){
+    const tr = document.createElement("tr");
+    tr.innerHTML = `<td class="mid">${m.name}</td><td>${badge(m.state)}</td>
+      <td class="port">${m.port ?? "auto"}</td>
+      <td style="text-align:right"><button class="ghost" onclick="unloadOne('${m.name}')">Entladen</button></td>`;
+    tb.appendChild(tr);
+    sel.insertAdjacentHTML("beforeend", `<option>${m.name}</option>`);
+  }
+  if(cur) sel.value = cur;
+}
+
+async function pull(){
+  const repo = $("#dl-repo").value.trim(), file = $("#dl-file").value.trim();
+  if(!repo || !file) return toast("Repo und Datei angeben.", true);
+  try{
+    const r = await api("/api/download", {method:"POST", body: JSON.stringify({repo, file})});
+    toast("Download gestartet.");
+    const stem = file.split("/").pop().replace(/\.gguf$/i,"");
+    $("#rg-alias").value = stem; $("#rg-path").value = r.expected_path;
+    $("#register-box").style.display = "block";
+    trackJob(r.job_id);
+  }catch(e){ toast(e.message, true); }
+}
+
+async function register(){
+  const alias = $("#rg-alias").value.trim(), model_path = $("#rg-path").value, ctx = parseInt($("#rg-ctx").value)||8192;
+  if(!alias) return toast("Alias angeben.", true);
+  try{ await api("/api/register",{method:"POST",body:JSON.stringify({alias,model_path,ctx})});
+    toast("Eingepflegt — llama-swap lädt neu."); refresh();
+  }catch(e){ toast(e.message, true); }
+}
+
+async function unloadOne(m){ try{ await api("/api/unload?model="+encodeURIComponent(m),{method:"POST"}); toast("Entladen: "+m); setTimeout(refresh,600);}catch(e){toast(e.message,true);} }
+async function unloadAll(){ try{ await api("/api/unload",{method:"POST"}); toast("Alle Modelle entladen."); setTimeout(refresh,600);}catch(e){toast(e.message,true);} }
+async function update(){ try{ const r = await api("/api/update",{method:"POST"}); toast("Update läuft."); trackJob(r.job_id);}catch(e){toast(e.message,true);} }
+
+async function sendChat(){
+  const model = $("#chat-model").value, message = $("#chat-msg").value.trim();
+  if(!model) return toast("Kein Modell vorhanden.", true);
+  if(!message) return;
+  const btn = $("#chat-btn"); btn.disabled = true; btn.textContent = "…";
+  const box = $("#chat-reply"); box.style.display="block"; box.textContent="(wecke Modell, kann beim Swap kurz dauern…)";
+  try{ const r = await api("/api/chat",{method:"POST",body:JSON.stringify({model,message})}); box.textContent = r.reply; }
+  catch(e){ box.textContent = "Fehler: "+e.message; }
+  btn.disabled=false; btn.textContent="Senden"; refresh();
+}
+
+// --- Jobs ---
+const tracked = new Set();
+function trackJob(id){ tracked.add(id); renderJobs(); }
+async function renderJobs(){
+  let jobs;
+  try{ jobs = await api("/api/jobs"); }catch(e){ return; }
+  $("#jobs-empty").style.display = jobs.length ? "none" : "block";
+  const c = $("#jobs"); c.innerHTML = "";
+  for(const j of jobs){
+    const open = tracked.has(j.id);
+    const st = j.state==="done" ? '<span class="badge b-run">fertig</span>'
+      : j.state==="failed" ? '<span class="badge" style="background:rgba(229,83,75,.16);color:#e5534b">fehler</span>'
+      : '<span class="badge b-load">läuft…</span>';
+    const div = document.createElement("div"); div.className="job";
+    div.innerHTML = `<div class="job-h" onclick="toggleJob('${j.id}')">
+        <span class="mid">${j.label}</span>${st}</div>
+      ${open ? `<div class="log">${(j.log||[]).join("\n").replace(/</g,"&lt;")}</div>` : ""}`;
+    c.appendChild(div);
+  }
+}
+function toggleJob(id){ tracked.has(id) ? tracked.delete(id) : tracked.add(id); renderJobs(); }
+
+refresh(); renderJobs();
+setInterval(refresh, 3000);
+setInterval(renderJobs, 1500);
+</script>
+</body>
+</html>