fix(llm): LLM antwortet jetzt in der konfigurierten Systemsprache

Bisher wurde `language` zwar an STT/TTS übergeben, aber nie an den LLM.
Lokales Modell antwortete deshalb immer auf Englisch/Deutsch, egal ob
Französisch, Spanisch o.a. eingestellt war.

Lösung: `language`-Parameter durch die gesamte LLM-Schicht gezogen:
- base.py: `lang_instruction()` helper + Signatur erweitert
- local_openai_compatible.py: Sprachanweisung ("Respond in Français.")
  wird als letzter System-Part in den Message-Stack eingefügt
- openrouter.py: Explizite Sprachanweisung ergänzt den bestehenden
  "Answer in the same language as the user"-Prompt
- fallback.py: FallbackLLMProvider leitet `language` durch
- orchestrator.py: alle 3 LLM-Aufrufstellen übergeben `language`
- Tests: alle Stub-LLMs um `language=None` ergänzt

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/core/orchestrator.py

@@ -99,6 +99,7 @@ class Orchestrator:
             trace.semantic_response = await self.llm.complete(
                 trace.cleaned_transcript or "",
                 history=history,
+                language=language,
             )
         if not trace.semantic_response:
             raise RuntimeError("LLM returned an empty response")
@@ -158,7 +159,7 @@ class Orchestrator:
         parts: list[str] = []
         stream_fn = getattr(self.llm, "stream", None)
         if stream_fn is not None:
-            async for delta in stream_fn(trace.cleaned_transcript or "", history=history):
+            async for delta in stream_fn(trace.cleaned_transcript or "", history=history, language=language):
                 parts.append(delta)
                 if on_token:
                     await on_token(delta)
@@ -167,7 +168,7 @@ class Orchestrator:
                         await _emit_sentence(sentence)
         else:
             # Provider ohne Streaming -> komplette Antwort als ein Token.
-            result = await self.llm.complete(trace.cleaned_transcript or "", history=history)
+            result = await self.llm.complete(trace.cleaned_transcript or "", history=history, language=language)
             parts.append(result)
             if on_token:
                 await on_token(result)

app/providers/fallback.py

@@ -37,11 +37,13 @@ class FallbackSTTProvider(_Chain):
 
 
 class FallbackLLMProvider(_Chain):
-    async def complete(self, text, history=None, session_id=None) -> str:
+    async def complete(self, text, history=None, session_id=None, language=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)
+                result = await provider.complete(
+                    text, history=history, session_id=session_id, language=language
+                )
                 if index > 0:
                     self._on_fallback()
                 return result
@@ -50,12 +52,14 @@ class FallbackLLMProvider(_Chain):
                 self._on_error(name)
         raise last_exc
 
-    async def stream(self, text, history=None, session_id=None) -> AsyncIterator[str]:
+    async def stream(self, text, history=None, session_id=None, language=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):
+                async for delta in provider.stream(
+                    text, history=history, session_id=session_id, language=language
+                ):
                     produced = True
                     yield delta
                 if index > 0:

app/providers/llm/base.py

@@ -2,6 +2,26 @@ import json
 from abc import ABC, abstractmethod
 from collections.abc import AsyncIterator
 
+_LANG_NAMES: dict[str, str] = {
+    "de": "Deutsch",
+    "en": "English",
+    "fr": "Français",
+    "es": "Español",
+    "it": "Italiano",
+    "nl": "Nederlands",
+    "ru": "Русский",
+    "zh": "中文",
+    "cmn": "中文",
+}
+
+
+def lang_instruction(language: str | None) -> str | None:
+    """'Respond in X.' instruction for the given ISO language code, or None."""
+    if not language:
+        return None
+    name = _LANG_NAMES.get(language, language)
+    return f"Respond in {name}."
+
 
 def sse_delta(line: str) -> str | None:
     """Extrahiert das Token-Delta aus einer OpenAI-kompatiblen SSE-Zeile (oder None)."""
@@ -24,6 +44,7 @@ class LLMProvider(ABC):
         text: str,
         history: list[dict] | None = None,
         session_id: str | None = None,
+        language: str | None = None,
     ) -> str: ...
 
     async def stream(
@@ -31,9 +52,10 @@ class LLMProvider(ABC):
         text: str,
         history: list[dict] | None = None,
         session_id: str | None = None,
+        language: str | None = None,
     ) -> AsyncIterator[str]:
         """Token-Stream. Default: kein echtes Streaming -> komplette Antwort als ein Chunk.
 
         Provider mit SSE-Unterstuetzung ueberschreiben diese Methode.
         """
-        yield await self.complete(text, history=history, session_id=session_id)
+        yield await self.complete(text, history=history, session_id=session_id, language=language)

app/providers/llm/local_openai_compatible.py

@@ -2,7 +2,7 @@ from collections.abc import AsyncIterator
 
 import httpx
 
-from app.providers.llm.base import LLMProvider, sse_delta
+from app.providers.llm.base import LLMProvider, lang_instruction, sse_delta
 
 
 class LocalOpenAICompatibleLLM(LLMProvider):
@@ -24,7 +24,9 @@ class LocalOpenAICompatibleLLM(LLMProvider):
         self.max_tokens = max_tokens
         self.temperature = temperature
 
-    def _build_messages(self, text: str, history: list[dict] | None) -> list[dict]:
+    def _build_messages(
+        self, text: str, history: list[dict] | None, language: str | None = None
+    ) -> list[dict]:
         # Manche Chat-Templates (z. B. Qwen3) erlauben nur EINE System-Nachricht,
         # ganz am Anfang. Daher den Sprach-System-Prompt und etwaige System-
         # Nachrichten aus der History (z. B. Nutzer-Erinnerungen) zu einer einzigen
@@ -40,6 +42,9 @@ class LocalOpenAICompatibleLLM(LLMProvider):
                     system_parts.append(content)
             else:
                 rest.append(msg)
+        instr = lang_instruction(language)
+        if instr:
+            system_parts.append(instr)
 
         messages: list[dict] = []
         if system_parts:
@@ -48,10 +53,12 @@ class LocalOpenAICompatibleLLM(LLMProvider):
         messages.append({"role": "user", "content": text})
         return messages
 
-    def _payload(self, text: str, history: list[dict] | None, stream: bool) -> dict:
+    def _payload(
+        self, text: str, history: list[dict] | None, stream: bool, language: str | None = None
+    ) -> dict:
         payload: dict = {
             "model": self.model,
-            "messages": self._build_messages(text, history),
+            "messages": self._build_messages(text, history, language=language),
             "temperature": self.temperature,
         }
         if stream:
@@ -68,12 +75,13 @@ class LocalOpenAICompatibleLLM(LLMProvider):
         text: str,
         history: list[dict] | None = None,
         session_id: str | None = None,
+        language: str | None = None,
     ) -> str:
         async with httpx.AsyncClient(timeout=120) as client:
             response = await client.post(
                 f"{self.base_url}/chat/completions",
                 headers={"Authorization": f"Bearer {self.api_key}"},
-                json=self._payload(text, history, stream=False),
+                json=self._payload(text, history, stream=False, language=language),
             )
             response.raise_for_status()
             data = response.json()
@@ -84,13 +92,14 @@ class LocalOpenAICompatibleLLM(LLMProvider):
         text: str,
         history: list[dict] | None = None,
         session_id: str | None = None,
+        language: str | None = None,
     ) -> AsyncIterator[str]:
         async with httpx.AsyncClient(timeout=120) as client:
             async with client.stream(
                 "POST",
                 f"{self.base_url}/chat/completions",
                 headers={"Authorization": f"Bearer {self.api_key}"},
-                json=self._payload(text, history, stream=True),
+                json=self._payload(text, history, stream=True, language=language),
             ) as response:
                 if response.status_code >= 400:
                     body = await response.aread()

app/providers/llm/openrouter.py

@@ -2,7 +2,7 @@ from collections.abc import AsyncIterator
 
 import httpx
 
-from app.providers.llm.base import LLMProvider, sse_delta
+from app.providers.llm.base import LLMProvider, lang_instruction, sse_delta
 
 
 SYSTEM_PROMPT = """
@@ -51,7 +51,9 @@ class OpenRouterLLMProvider(LLMProvider):
         self.api_key = (api_key or "").strip()
         self.model = (model or "").strip()
 
-    def _build_messages(self, text: str, history: list[dict] | None) -> list[dict]:
+    def _build_messages(
+        self, text: str, history: list[dict] | None, language: str | None = None
+    ) -> list[dict]:
         if not self.api_key:
             raise ValueError("OPENROUTER_API_KEY is empty")
         if not self.model:
@@ -59,7 +61,12 @@ class OpenRouterLLMProvider(LLMProvider):
         if not text or not text.strip():
             raise ValueError("LLM input text is empty")
 
-        messages = [{"role": "system", "content": SYSTEM_PROMPT}]
+        system_content = SYSTEM_PROMPT
+        instr = lang_instruction(language)
+        if instr:
+            system_content = f"{SYSTEM_PROMPT}\n\n{instr}"
+
+        messages = [{"role": "system", "content": system_content}]
         if history:
             messages.extend(history)
         messages.append({"role": "user", "content": text.strip()})
@@ -70,10 +77,11 @@ class OpenRouterLLMProvider(LLMProvider):
         text: str,
         history: list[dict] | None = None,
         session_id: str | None = None,
+        language: str | None = None,
     ) -> str:
         payload = {
             "model": self.model,
-            "messages": self._build_messages(text, history),
+            "messages": self._build_messages(text, history, language=language),
         }
 
         timeout = httpx.Timeout(connect=10.0, read=120.0, write=30.0, pool=10.0)
@@ -115,10 +123,11 @@ class OpenRouterLLMProvider(LLMProvider):
         text: str,
         history: list[dict] | None = None,
         session_id: str | None = None,
+        language: str | None = None,
     ) -> AsyncIterator[str]:
         payload = {
             "model": self.model,
-            "messages": self._build_messages(text, history),
+            "messages": self._build_messages(text, history, language=language),
             "stream": True,
         }
         timeout = httpx.Timeout(connect=10.0, read=120.0, write=30.0, pool=10.0)

tests/test_emergency_llm.py

@@ -12,7 +12,7 @@ class StubLLM:
         self.raw = raw
         self.calls = 0
 
-    async def complete(self, text, history=None, session_id=None):
+    async def complete(self, text, history=None, session_id=None, language=None):
         self.calls += 1
         return self.raw
 

tests/test_endpoints_e2e.py

@@ -59,7 +59,7 @@ def test_session_route_applies(monkeypatch):
 
 def test_chat_per_request_override_and_loopback(monkeypatch):
     class StubLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "Mir geht es gut, danke."
 
     class StubTTS:

tests/test_memory.py

@@ -8,7 +8,7 @@ client = TestClient(app)
 
 def _install_stubs(monkeypatch, captured):
     class MemLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             captured["history"] = list(history or [])
             return f"Antwort auf: {text}"
 

tests/test_memory_extraction.py

@@ -12,7 +12,7 @@ class StubLLM:
         self.raw = raw
         self.calls = 0
 
-    async def complete(self, text, history=None, session_id=None):
+    async def complete(self, text, history=None, session_id=None, language=None):
         self.calls += 1
         return self.raw
 

tests/test_quota_safety.py

@@ -10,7 +10,7 @@ client = TestClient(app)
 
 def _stub_chat(monkeypatch, answer="ok"):
     class StubLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return answer
 
     class StubTTS:

tests/test_realtime.py

@@ -42,10 +42,10 @@ def test_energy_vad_ignores_silence_without_speech():
 
 def _install_slow_stream(monkeypatch):
     class SlowLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "fertig"
 
-        async def stream(self, text, history=None, session_id=None):
+        async def stream(self, text, history=None, session_id=None, language=None):
             for i in range(200):
                 await asyncio.sleep(0.005)
                 yield f"t{i} "
@@ -89,7 +89,7 @@ def _install_voice_stubs(monkeypatch):
             return "ok"
 
     class LLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "antwort"
 
     class TTS:

tests/test_resilience.py

@@ -20,11 +20,11 @@ def _run(coro):
 
 def test_llm_fallback_uses_second_on_error():
     class BadLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             raise RuntimeError("down")
 
     class GoodLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "ok"
 
     chain = FallbackLLMProvider("llm", [("bad", BadLLM()), ("good", GoodLLM())])
@@ -37,7 +37,7 @@ def test_llm_fallback_uses_second_on_error():
 
 def test_llm_fallback_all_fail_raises():
     class BadLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             raise RuntimeError("x")
 
     chain = FallbackLLMProvider("llm", [("a", BadLLM()), ("b", BadLLM())])
@@ -47,18 +47,18 @@ def test_llm_fallback_all_fail_raises():
 
 def test_llm_stream_fallback_before_first_token():
     class BadStream:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "x"
 
-        async def stream(self, text, history=None, session_id=None):
+        async def stream(self, text, history=None, session_id=None, language=None):
             raise RuntimeError("boom")
             yield  # macht die Funktion zum Generator
 
     class GoodStream:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "ok"
 
-        async def stream(self, text, history=None, session_id=None):
+        async def stream(self, text, history=None, session_id=None, language=None):
             yield "he"
             yield "llo"
 
@@ -74,11 +74,11 @@ def test_llm_stream_fallback_before_first_token():
 
 def test_config_llm_fallback_applied(monkeypatch):
     class BadLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             raise RuntimeError("primary down")
 
     class GoodLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "rescued"
 
     class StubTTS:
@@ -102,7 +102,7 @@ def test_config_llm_fallback_applied(monkeypatch):
 
 def test_metrics_endpoint_records_requests_and_stages(monkeypatch):
     class StubLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "hi"
 
     class StubTTS:

tests/test_streaming.py

@@ -40,7 +40,7 @@ def test_sse_delta_parsing():
 
 def test_base_stream_default_yields_full_completion():
     class P(LLMProvider):
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "ganze Antwort"
 
     async def run():
@@ -51,10 +51,10 @@ def test_base_stream_default_yields_full_completion():
 
 def _install_streaming(monkeypatch, tokens):
     class StreamLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "".join(tokens)
 
-        async def stream(self, text, history=None, session_id=None):
+        async def stream(self, text, history=None, session_id=None, language=None):
             for tok in tokens:
                 yield tok
 
@@ -95,7 +95,7 @@ def test_ws_without_stream_flag_has_no_tokens(monkeypatch):
 
 def test_ws_stream_fallback_for_nonstreaming_llm(monkeypatch):
     class OnlyComplete:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "komplett"
 
     class StubTTS:
@@ -118,10 +118,10 @@ def test_ws_audio_stream_sends_chunks_per_sentence(monkeypatch):
     tts_calls = []
 
     class StreamLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return "Satz eins. Satz zwei."
 
-        async def stream(self, text, history=None, session_id=None):
+        async def stream(self, text, history=None, session_id=None, language=None):
             for tok in ["Satz ", "eins. ", "Satz ", "zwei."]:
                 yield tok
 

tests/test_ws.py

@@ -11,7 +11,7 @@ client = TestClient(app)
 
 def _install_stubs(monkeypatch, captured):
     class MemLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             captured["history"] = list(history or [])
             return f"Echo: {text}"
 
@@ -74,7 +74,7 @@ def _install_voice_stubs(monkeypatch):
             return f"erkannt({len(audio_bytes)})"
 
     class StubLLM:
-        async def complete(self, text, history=None, session_id=None):
+        async def complete(self, text, history=None, session_id=None, language=None):
             return f"Antwort zu {text}"
 
     class StubTTS: