feat: Resilienz (Fallback-Ketten) und Metriken (#5)

- Fallback-Provider (app/providers/fallback.py) fuer STT/LLM/TTS: Provider-Kette
  der Reihe nach; Config *_FALLBACK; build_orchestrator baut Ketten (dedupliziert)
- LLM-Stream-Fallback nur solange kein Token gesendet wurde
- Metriken (app/metrics.py): In-Memory Counter/Timer, keine externe Dependency
- HTTP-Middleware (Requests/Latenz/Status je Pfad); Pipeline-Stufen-Timing stt/llm/tts;
  Fallback-/Fehlerzaehler; GET /api/metrics (JSON + Prometheus)
- Tests: 58 gruen (+6); Doku aktualisiert (README, BEDIENUNGSANLEITUNG, Architektur, .env.example)

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

.env.example

@@ -36,3 +36,9 @@ DEFAULT_OUTPUT_ENDPOINT=local-default
 LOCAL_LLM_BASE_URL=http://127.0.0.1:11434/v1
 LOCAL_LLM_API_KEY=dummy
 LOCAL_LLM_MODEL=llama3.1
+
+# --- Resilienz: Fallback-Ketten (kommaseparierte Provider-Namen) ------------
+# Faellt der primaere Provider aus, uebernimmt der naechste.
+# STT_FALLBACK=faster-whisper
+# LLM_FALLBACK=local-openai-compatible
+# TTS_FALLBACK=piper

BEDIENUNGSANLEITUNG.md

@@ -292,7 +292,30 @@ senden — die laufende Antwort wird abgebrochen (`interrupted`-Event).
 
 ---
 
-## 11. Fehlerbehebung
+## 11. Resilienz & Metriken (Betrieb)
+
+**Fallback bei Provider-Ausfall:** Pro Modul eine Ersatzliste setzen (in `.env`).
+Fällt der primäre Provider aus, übernimmt der nächste automatisch:
+
+```
+LLM_FALLBACK=local-openai-compatible
+STT_FALLBACK=faster-whisper
+TTS_FALLBACK=piper
+```
+
+**Metriken ansehen:**
+
+```bash
+curl http://localhost:8080/api/metrics                 # JSON
+curl http://localhost:8080/api/metrics?format=prometheus
+```
+
+Enthält Request-Zahlen/-Laufzeiten, Pipeline-Stufen (`stt`/`llm`/`tts`) und
+Fallback-/Fehlerzähler. Die Werte gelten pro laufendem Prozess.
+
+---
+
+## 12. Fehlerbehebung
 
 | Symptom | Ursache | Lösung |
 |---|---|---|
@@ -313,7 +336,7 @@ Logs erscheinen im Terminal, in dem `make run` läuft. Für mehr Details
 
 ---
 
-## 12. Tests ausführen
+## 13. Tests ausführen
 
 ```bash
 make test

Docs/voice-assistant-architecture.md

@@ -171,6 +171,7 @@ ein No-op; `LoopbackOutput` sammelt die Chunks (testbar ohne Hardware).
 | `POST /api/admin/users`              | Nutzer anlegen (Admin-Key) → Token einmalig |
 | `GET  /api/me` · `PUT /api/me/prefs` | aktueller Nutzer + dauerhafte Präferenzen |
 | `GET/POST/DELETE /api/me/memories`   | Langzeit-Erinnerungen des Nutzers |
+| `GET  /api/metrics`                  | Metriken (JSON / Prometheus) |
 | `WS   /ws/chat`                       | Echtzeit-Chat (Text rein, Streaming-Events) |
 | `WS   /ws/voice`                      | Echtzeit-Sprache (Audio rein → STT → Antwort) |
 
@@ -206,7 +207,7 @@ Reihenfolge der Weiterentwicklung:
 2. **(erledigt)** Cloud-Fundament: Bearer-Token-Auth, Mehrbenutzer, persistenter SQLite-Store, Mandanten-Trennung, dauerhafte Nutzer-Präferenzen. Offen: Skalierung auf gemeinsamen Store (Postgres/Redis) für mehrere Instanzen.
 3. **(erledigt)** Konversationsgedächtnis: Kurzzeit-Gesprächsverlauf pro Session + Langzeit-Erinnerungen pro Nutzer (manuell gepflegt, als LLM-Kontext). Offen: **automatische** Extraktion/Zusammenfassung von Erinnerungen aus Gesprächen.
 4. **(weitgehend erledigt)** Echtzeit: WebSocket-Streaming-Chat (`/ws/chat`), **Token-Level-LLM-Streaming (SSE, `stream:true`)**, **Audio-Streaming (chunked TTS satzweise, `audio_stream:true`)**, **Audio-Eingang (`/ws/voice`)**, **Barge-in/Turn-Manager (`interrupt` bricht laufende Antwort ab)** und **VAD-Aeusserungserkennung (energie-basiert, opt-in)** sind umgesetzt. Offen: **echte partielle Live-Transkripte (Streaming-STT-Dienst, wortweise)** und **WebRTC (aiortc)** — beide brauchen schwere Abhaengigkeiten/Dienste. Heute laeuft STT pro Aeusserung.
-5. **Resilienz:** Fallback-Policy (remote KI fällt aus → lokaler/alternativer Provider), Metriken/Tracing.
+5. **(weitgehend erledigt)** Resilienz: Fallback-Ketten je Modul (`*_FALLBACK`, Provider faellt aus → naechster) und In-Memory-Metriken (`/api/metrics`: Request/Latenz, Pipeline-Stufen, Fallback/Fehler; JSON + Prometheus). Offen: verteiltes Tracing, Alerting.
 6. **Betrieb:** Kosten-/Quota-Kontrolle pro Nutzer; Notfall-/Eskalationskonzept (Senioren-Kontext).
 7. **TransportRouter** als eigene lokal/remote-Achse aktivieren.
 
@@ -224,13 +225,14 @@ voice-assistant-scaffold/
 │   ├── dependencies.py    # Registries, ResolvedRoute, resolve_route, Store-/Router-Singleton
 │   ├── store.py           # Persistenz: Store-Interface + SQLiteStore (Nutzer/Sessions/Verlauf)
 │   ├── auth.py            # Bearer-Token-Auth (require_user) + Admin-Schutz
+│   ├── metrics.py         # In-Memory-Metriken (Counter/Timer, JSON + Prometheus)
 │   ├── errors.py          # RoutingError -> HTTP 422
 │   ├── schemas.py         # Pydantic-Modelle
 │   ├── api/               # health, chat, speak, transcribe, devices, sessions, config, admin, me, ws
 │   ├── core/              # orchestrator
 │   ├── audio/             # router, transport_router, vad, endpoints/input|output/*
 │   ├── pipeline/          # input_cleaner, spoken_response_adapter, tts_normalizer, sentence_chunker
-│   └── providers/         # stt/ llm/ tts/  (openrouter + lokale Stubs)
+│   └── providers/         # stt/ llm/ tts/ (openrouter + lokale Stubs) + fallback.py
 ├── config/                # voice-assistant.example.toml (+ lokale .toml, gitignored)
 ├── data/                  # SQLite-DB (gitignored)
 ├── deploy/                # systemd unit + env-Beispiel

README.md

@@ -16,6 +16,7 @@ Praktische Bedienung: [`BEDIENUNGSANLEITUNG.md`](BEDIENUNGSANLEITUNG.md).
 - **Geschichtete Konfiguration** mit Profilen (`local-dev` / `hybrid` / `cloud`)
 - **Routing auf jeder Ebene:** Default → Profil → Nutzer → Session → Request
 - **Authentifizierung** (Bearer-Token) + persistente Nutzer/Sessions (SQLite)
+- **Resilienz:** Fallback-Ketten je Modul (Provider fällt aus → nächster) + Metriken
 - **Gesprächsgedächtnis pro Session:** Verlauf wird gespeichert und fließt ins LLM
 - **Langzeit-Erinnerungen pro Nutzer:** dauerhafte Fakten/Vorlieben als LLM-Kontext
 - **WebSocket-Streaming-Chat** (`/ws/chat`) als Echtzeit-Transport
@@ -90,6 +91,7 @@ Aktive Konfiguration prüfen: `curl http://localhost:8080/api/config`.
 | `GET/POST/DELETE /api/me/memories` | Langzeit-Erinnerungen des Nutzers verwalten |
 | `WS   /ws/chat`                 | Echtzeit-Chat über WebSocket (Text rein, Streaming-Events) |
 | `WS   /ws/voice`                | Echtzeit-Sprache (Audio rein → Transkript → Antwort) |
+| `GET  /api/metrics`             | Metriken (JSON, oder `?format=prometheus`) |
 
 Beispiel (Sprachausgabe an den Test-Loopback, lokaler TTS-Stub):
 
@@ -162,6 +164,30 @@ eine neue Eingabe) abbrechen — der Server stoppt das Streaming und meldet
 > Sprechens) und **WebRTC** sind als nächste Increments vorgesehen (siehe
 > Architektur-Dokument). Heute läuft STT pro Äußerung.
 
+## Resilienz & Metriken
+
+**Fallback-Ketten:** Pro Modul lässt sich eine Ersatz-Provider-Liste setzen. Fällt
+der primäre Provider aus (Timeout/Fehler), übernimmt transparent der nächste:
+
+```bash
+# z. B. Cloud-LLM mit lokalem Fallback
+LLM_FALLBACK=local-openai-compatible
+STT_FALLBACK=faster-whisper
+TTS_FALLBACK=piper
+```
+
+Die Kette ist `Route-Provider` + `*_FALLBACK` (dedupliziert). Erfolgreiche Fallbacks
+und Provider-Fehler werden gezählt.
+
+**Metriken** (`GET /api/metrics`): Request-Counts/-Latenzen pro Pfad, Pipeline-Stufen
+(`stt`/`llm`/`tts`), Fallback-/Fehlerzähler — als JSON oder Prometheus-Text
+(`?format=prometheus`). In-Memory pro Prozess (keine externe Dependency).
+
+```bash
+curl http://localhost:8080/api/metrics
+curl http://localhost:8080/api/metrics?format=prometheus
+```
+
 ## Authentifizierung
 
 Standardmäßig (`AUTH_ENABLED=true`) sind `chat`/`speak`/`transcribe`/`sessions`/`me`

app/api/metrics.py

@@ -0,0 +1,13 @@
+from fastapi import APIRouter, Query
+from fastapi.responses import PlainTextResponse
+
+from app.metrics import metrics
+
+router = APIRouter()
+
+
+@router.get("/metrics")
+async def get_metrics(format: str = Query(default="json", description="json | prometheus")):
+    if format == "prometheus":
+        return PlainTextResponse(metrics.prometheus(), media_type="text/plain; version=0.0.4")
+    return metrics.snapshot()

app/config.py

@@ -124,6 +124,9 @@ class Settings(BaseSettings):
     admin_api_key: str = ""
     auth_enabled: bool = True
     history_max_messages: int = 10
+    stt_fallback: str = ""  # kommaseparierte Provider-Namen (Fallback-Kette)
+    llm_fallback: str = ""
+    tts_fallback: str = ""
     model_config = SettingsConfigDict(
         env_file=ENV_FILE, case_sensitive=False, extra="ignore"
     )

app/core/orchestrator.py

@@ -1,5 +1,10 @@
 from app.schemas import AudioChunk, PipelineTrace
 from app.pipeline.sentence_chunker import SentenceChunker
+from app.metrics import timer, metrics
+
+
+def _stage(name: str):
+    return timer("stage_duration_seconds", {"stage": name})
 
 # Festes Ausgabeformat der TTS-Stufe (s16le PCM, 24 kHz, mono).
 TTS_AUDIO_FORMAT = "pcm"
@@ -48,6 +53,7 @@ class Orchestrator:
         # input dient hier nur der Validierung/Metadaten; das Audio kommt per Upload.
         if input is not None:
             await input.capabilities()
+        with _stage("stt"):
             trace.raw_transcript = await self.stt.transcribe(
                 audio_bytes,
                 fmt=fmt,
@@ -67,6 +73,7 @@ class Orchestrator:
     ):
         spoken = await self.spoken_adapter.run(text, language=language)
         normalized = await self.tts_normalizer.run(spoken, language=language)
+        with _stage("tts"):
             audio = await self.tts.synthesize(normalized, voice=voice)
         await self._emit_to_output(audio, output)
         return audio
@@ -84,6 +91,7 @@ class Orchestrator:
         trace.raw_transcript = text
         trace.cleaned_transcript = await self.input_cleaner.run(text or "")
 
+        with _stage("llm"):
             trace.semantic_response = await self.llm.complete(
                 trace.cleaned_transcript or "",
                 history=history,
@@ -100,6 +108,7 @@ class Orchestrator:
             language=language,
         )
 
+        with _stage("tts"):
             audio = await self.tts.synthesize(
                 trace.tts_ready_text,
                 voice=voice,

app/dependencies.py

@@ -19,6 +19,11 @@ from app.providers.llm.openrouter import OpenRouterLLMProvider
 from app.providers.tts.openrouter import OpenRouterTTSProvider
 from app.providers.tts.chatterbox import ChatterboxTTSProvider
 from app.providers.tts.piper import PiperTTSProvider
+from app.providers.fallback import (
+    FallbackSTTProvider,
+    FallbackLLMProvider,
+    FallbackTTSProvider,
+)
 from app.pipeline.input_cleaner import InputCleaner
 from app.pipeline.spoken_response_adapter import SpokenResponseAdapter
 from app.pipeline.tts_normalizer import TTSNormalizer
@@ -198,11 +203,32 @@ def resolve_route(
     return ResolvedRoute(**resolved)
 
 
+_FALLBACK_CLASS = {
+    "stt": FallbackSTTProvider,
+    "llm": FallbackLLMProvider,
+    "tts": FallbackTTSProvider,
+}
+
+
+def _provider_chain(registry, primary: str, fallback_csv: str, module: str, cfg: Settings):
+    """Baut primaeren Provider + optionale Fallback-Kette (dedupliziert, Reihenfolge erhalten)."""
+    names = [primary] + [n.strip() for n in (fallback_csv or "").split(",") if n.strip()]
+    seen, ordered = set(), []
+    for name in names:
+        if name not in seen:
+            seen.add(name)
+            ordered.append(name)
+    entries = [(name, _from_registry(registry, name, module.upper(), cfg)) for name in ordered]
+    if len(entries) == 1:
+        return entries[0][1]
+    return _FALLBACK_CLASS[module](module, entries)
+
+
 def build_orchestrator(route: ResolvedRoute, cfg: Settings = settings) -> Orchestrator:
     return Orchestrator(
-        stt=get_stt_provider(route.stt_provider, cfg),
-        llm=get_llm_provider(route.llm_provider, cfg),
-        tts=get_tts_provider(route.tts_provider, cfg),
+        stt=_provider_chain(STT_REGISTRY, route.stt_provider, cfg.stt_fallback, "stt", cfg),
+        llm=_provider_chain(LLM_REGISTRY, route.llm_provider, cfg.llm_fallback, "llm", cfg),
+        tts=_provider_chain(TTS_REGISTRY, route.tts_provider, cfg.tts_fallback, "tts", cfg),
         input_cleaner=InputCleaner(),
         spoken_adapter=SpokenResponseAdapter(),
         tts_normalizer=TTSNormalizer(),

app/main.py

@@ -1,4 +1,8 @@
-from fastapi import FastAPI
+import time
+
+from fastapi import FastAPI, Request
+
+from app.metrics import metrics
 from app.api.health import router as health_router
 from app.api.chat import router as chat_router
 from app.api.transcribe import router as transcribe_router
@@ -8,9 +12,27 @@ from app.api.sessions import router as sessions_router
 from app.api.config import router as config_router
 from app.api.admin import router as admin_router
 from app.api.me import router as me_router
+from app.api.metrics import router as metrics_router
 from app.api.ws import router as ws_router
 
 app = FastAPI(title="Voice Assistant Gateway")
+
+
+@app.middleware("http")
+async def record_metrics(request: Request, call_next):
+    start = time.perf_counter()
+    response = await call_next(request)
+    duration = time.perf_counter() - start
+    # Route-Template (z. B. /api/sessions/{session_id}/route) statt konkreter URL,
+    # um die Label-Kardinalitaet niedrig zu halten.
+    route = request.scope.get("route")
+    path = getattr(route, "path", request.url.path)
+    labels = {"method": request.method, "path": path}
+    metrics.inc("http_requests_total", {**labels, "status": response.status_code})
+    metrics.observe("http_request_duration_seconds", duration, labels)
+    return response
+
+
 app.include_router(health_router)
 app.include_router(chat_router, prefix="/api")
 app.include_router(transcribe_router, prefix="/api")
@@ -20,4 +42,5 @@ app.include_router(sessions_router, prefix="/api")
 app.include_router(config_router, prefix="/api")
 app.include_router(admin_router, prefix="/api")
 app.include_router(me_router, prefix="/api")
+app.include_router(metrics_router, prefix="/api")
 app.include_router(ws_router)

app/metrics.py

@@ -0,0 +1,82 @@
+"""Schlanke In-Memory-Metriken (Counter + Timer) fuer einen Prozess.
+
+Bewusst ohne externe Dependency. Fuer mehrere Instanzen/Prozesse spaeter durch
+einen gemeinsamen Backend (z. B. Prometheus-Exporter) ersetzbar.
+"""
+
+import threading
+import time
+from collections import defaultdict
+
+
+class Metrics:
+    def __init__(self):
+        self._lock = threading.Lock()
+        self._counters: dict[str, float] = defaultdict(float)
+        self._timers: dict[str, list] = defaultdict(lambda: [0.0, 0])  # [sum, count]
+
+    @staticmethod
+    def _key(name: str, labels: dict | None) -> str:
+        if not labels:
+            return name
+        rendered = ",".join(f'{k}="{v}"' for k, v in sorted(labels.items()))
+        return f"{name}{{{rendered}}}"
+
+    def inc(self, name: str, labels: dict | None = None, value: float = 1.0) -> None:
+        with self._lock:
+            self._counters[self._key(name, labels)] += value
+
+    def observe(self, name: str, seconds: float, labels: dict | None = None) -> None:
+        with self._lock:
+            agg = self._timers[self._key(name, labels)]
+            agg[0] += seconds
+            agg[1] += 1
+
+    def snapshot(self) -> dict:
+        with self._lock:
+            counters = dict(self._counters)
+            timers = {
+                key: {
+                    "sum": agg[0],
+                    "count": agg[1],
+                    "avg": (agg[0] / agg[1] if agg[1] else 0.0),
+                }
+                for key, agg in self._timers.items()
+            }
+        return {"counters": counters, "timers": timers}
+
+    def prometheus(self) -> str:
+        snap = self.snapshot()
+        lines = []
+        for key, value in sorted(snap["counters"].items()):
+            lines.append(f"{key} {value}")
+        for key, agg in sorted(snap["timers"].items()):
+            base, _, labels = key.partition("{")
+            suffix = ("{" + labels) if labels else ""
+            lines.append(f"{base}_sum{suffix} {agg['sum']}")
+            lines.append(f"{base}_count{suffix} {agg['count']}")
+        return "\n".join(lines) + "\n"
+
+    def reset(self) -> None:
+        with self._lock:
+            self._counters.clear()
+            self._timers.clear()
+
+
+metrics = Metrics()
+
+
+class timer:
+    """Context-Manager: misst die Dauer und schreibt sie als Timer-Beobachtung."""
+
+    def __init__(self, name: str, labels: dict | None = None):
+        self.name = name
+        self.labels = labels
+
+    def __enter__(self):
+        self._start = time.perf_counter()
+        return self
+
+    def __exit__(self, *exc):
+        metrics.observe(self.name, time.perf_counter() - self._start, self.labels)
+        return False

app/providers/fallback.py

@@ -0,0 +1,85 @@
+"""Fallback-Ketten: versuchen mehrere Provider der Reihe nach.
+
+Faellt der primaere Provider aus (Timeout/Fehler), wird transparent der naechste
+versucht. Erfolgreicher Fallback und Provider-Fehler werden als Metrik erfasst.
+"""
+
+from collections.abc import AsyncIterator
+
+from app.metrics import metrics
+
+
+class _Chain:
+    def __init__(self, module: str, entries: list[tuple[str, object]]):
+        self.module = module
+        self.entries = entries  # [(provider_name, provider), ...]
+
+    def _on_error(self, name: str) -> None:
+        metrics.inc("provider_error_total", {"module": self.module, "provider": name})
+
+    def _on_fallback(self) -> None:
+        metrics.inc("provider_fallback_total", {"module": self.module})
+
+
+class FallbackSTTProvider(_Chain):
+    async def transcribe(self, audio_bytes, fmt, language=None) -> str:
+        last_exc = None
+        for index, (name, provider) in enumerate(self.entries):
+            try:
+                result = await provider.transcribe(audio_bytes, fmt, language=language)
+                if index > 0:
+                    self._on_fallback()
+                return result
+            except Exception as exc:  # noqa: BLE001 - bewusst breit fuer Resilienz
+                last_exc = exc
+                self._on_error(name)
+        raise last_exc
+
+
+class FallbackLLMProvider(_Chain):
+    async def complete(self, text, history=None, session_id=None) -> str:
+        last_exc = None
+        for index, (name, provider) in enumerate(self.entries):
+            try:
+                result = await provider.complete(text, history=history, session_id=session_id)
+                if index > 0:
+                    self._on_fallback()
+                return result
+            except Exception as exc:  # noqa: BLE001
+                last_exc = exc
+                self._on_error(name)
+        raise last_exc
+
+    async def stream(self, text, history=None, session_id=None) -> AsyncIterator[str]:
+        last_exc = None
+        for index, (name, provider) in enumerate(self.entries):
+            produced = False
+            try:
+                async for delta in provider.stream(text, history=history, session_id=session_id):
+                    produced = True
+                    yield delta
+                if index > 0:
+                    self._on_fallback()
+                return
+            except Exception as exc:  # noqa: BLE001
+                last_exc = exc
+                self._on_error(name)
+                if produced:
+                    # Schon Token gesendet -> kein Fallback mehr moeglich.
+                    raise
+        raise last_exc
+
+
+class FallbackTTSProvider(_Chain):
+    async def synthesize(self, text, voice=None, audio_format="pcm") -> bytes:
+        last_exc = None
+        for index, (name, provider) in enumerate(self.entries):
+            try:
+                result = await provider.synthesize(text, voice=voice, audio_format=audio_format)
+                if index > 0:
+                    self._on_fallback()
+                return result
+            except Exception as exc:  # noqa: BLE001
+                last_exc = exc
+                self._on_error(name)
+        raise last_exc

tests/conftest.py

@@ -3,16 +3,18 @@ import pytest
 import app.dependencies as deps
 from app.config import settings
 from app.store import SQLiteStore
+from app.metrics import metrics
 
 
 @pytest.fixture(autouse=True)
 def reset_state(tmp_path, monkeypatch):
-    """Pro Test: frische SQLite-DB, frischer Singleton-Audio-Router, Auth aus.
+    """Pro Test: frische SQLite-DB, frischer Singleton-Audio-Router, Auth aus, Metriken leer.
 
     Auth-Tests schalten `settings.auth_enabled` selbst wieder ein.
     """
     deps._store = SQLiteStore(str(tmp_path / "test.db"))
     deps._audio_router = None
+    metrics.reset()
     monkeypatch.setattr(settings, "auth_enabled", False)
     monkeypatch.setattr(settings, "admin_api_key", "")
     yield

tests/test_resilience.py

@@ -0,0 +1,129 @@
+import asyncio
+
+import pytest
+from fastapi.testclient import TestClient
+
+import app.dependencies as deps
+from app.main import app
+from app.config import settings
+from app.metrics import metrics
+from app.providers.fallback import FallbackLLMProvider
+
+client = TestClient(app)
+
+
+def _run(coro):
+    return asyncio.run(coro)
+
+
+# --- Fallback-Einheiten ----------------------------------------------------
+
+def test_llm_fallback_uses_second_on_error():
+    class BadLLM:
+        async def complete(self, text, history=None, session_id=None):
+            raise RuntimeError("down")
+
+    class GoodLLM:
+        async def complete(self, text, history=None, session_id=None):
+            return "ok"
+
+    chain = FallbackLLMProvider("llm", [("bad", BadLLM()), ("good", GoodLLM())])
+    assert _run(chain.complete("x")) == "ok"
+
+    counters = metrics.snapshot()["counters"]
+    assert any("provider_fallback_total" in key for key in counters)
+    assert any('provider_error_total{module="llm",provider="bad"}' in key for key in counters)
+
+
+def test_llm_fallback_all_fail_raises():
+    class BadLLM:
+        async def complete(self, text, history=None, session_id=None):
+            raise RuntimeError("x")
+
+    chain = FallbackLLMProvider("llm", [("a", BadLLM()), ("b", BadLLM())])
+    with pytest.raises(RuntimeError):
+        _run(chain.complete("x"))
+
+
+def test_llm_stream_fallback_before_first_token():
+    class BadStream:
+        async def complete(self, text, history=None, session_id=None):
+            return "x"
+
+        async def stream(self, text, history=None, session_id=None):
+            raise RuntimeError("boom")
+            yield  # macht die Funktion zum Generator
+
+    class GoodStream:
+        async def complete(self, text, history=None, session_id=None):
+            return "ok"
+
+        async def stream(self, text, history=None, session_id=None):
+            yield "he"
+            yield "llo"
+
+    chain = FallbackLLMProvider("llm", [("bad", BadStream()), ("good", GoodStream())])
+
+    async def collect():
+        return [delta async for delta in chain.stream("x")]
+
+    assert _run(collect()) == ["he", "llo"]
+
+
+# --- Fallback ueber Config + Endpunkt --------------------------------------
+
+def test_config_llm_fallback_applied(monkeypatch):
+    class BadLLM:
+        async def complete(self, text, history=None, session_id=None):
+            raise RuntimeError("primary down")
+
+    class GoodLLM:
+        async def complete(self, text, history=None, session_id=None):
+            return "rescued"
+
+    class StubTTS:
+        async def synthesize(self, text, voice=None, audio_format="pcm"):
+            return b"A"
+
+    monkeypatch.setitem(deps.LLM_REGISTRY, "bad", lambda s: BadLLM())
+    monkeypatch.setitem(deps.LLM_REGISTRY, "good", lambda s: GoodLLM())
+    monkeypatch.setitem(deps.TTS_REGISTRY, "t", lambda s: StubTTS())
+    monkeypatch.setattr(settings, "llm_fallback", "good")
+
+    resp = client.post(
+        "/api/chat?debug=true",
+        json={"text": "x", "llm_provider": "bad", "tts_provider": "t"},
+    )
+    assert resp.status_code == 200
+    assert resp.json()["trace"]["semantic_response"] == "rescued"
+
+
+# --- Metriken --------------------------------------------------------------
+
+def test_metrics_endpoint_records_requests_and_stages(monkeypatch):
+    class StubLLM:
+        async def complete(self, text, history=None, session_id=None):
+            return "hi"
+
+    class StubTTS:
+        async def synthesize(self, text, voice=None, audio_format="pcm"):
+            return b"A"
+
+    monkeypatch.setitem(deps.LLM_REGISTRY, "l", lambda s: StubLLM())
+    monkeypatch.setitem(deps.TTS_REGISTRY, "t", lambda s: StubTTS())
+
+    resp = client.post(
+        "/api/chat?debug=true", json={"text": "x", "llm_provider": "l", "tts_provider": "t"}
+    )
+    assert resp.status_code == 200
+
+    snap = client.get("/api/metrics").json()
+    assert any("http_requests_total" in k and "chat" in k for k in snap["counters"])
+    assert any('stage_duration_seconds{stage="llm"}' in k for k in snap["timers"])
+    assert any('stage_duration_seconds{stage="tts"}' in k for k in snap["timers"])
+
+
+def test_metrics_prometheus_format(monkeypatch):
+    client.get("/health")
+    text = client.get("/api/metrics?format=prometheus").text
+    assert "http_requests_total" in text