core: Router bleibt bei Embedding-Ausfall lauffähig (Stufe 0) und erholt sich selbst

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

core/src/knaster/api.py

@@ -71,7 +71,8 @@ def create_app(system: System | None = None, start_prober: bool | None = None) -
 
     @app.get("/health")
     def health() -> dict[str, Any]:
-        return {"status": "ok", "skills": len(system.library.all()), "dry_run": system.settings.dry_run}
+        return {"status": "ok", "skills": len(system.library.all()), "dry_run": system.settings.dry_run,
+                "embedding_index": "ok" if system.router.index_error is None else system.router.index_error}
 
     @app.get("/metrics")
     def metrics() -> Response:

core/src/knaster/router.py

@@ -79,11 +79,18 @@ class Router:
         self.config = config or RouterConfig()
         self.rng = rng or random.Random()
         self.index = SkillIndex(embedder)
+        self.index_error: str | None = None
         self.refresh()
 
-    def refresh(self) -> None:
-        """Index neu aufbauen (nach Skill-Änderungen)."""
-        self.index.rebuild({s.name: s.index_texts() for s in self.library.all()})
+    def refresh(self) -> bool:
+        """Index neu aufbauen (nach Skill-Änderungen). Fällt das Embedding aus, bleibt der Service lauffähig."""
+        try:
+            self.index.rebuild({s.name: s.index_texts() for s in self.library.all()})
+        except Exception as exc:  # noqa: BLE001 – Embedding-Endpunkt nicht erreichbar
+            self.index_error = f"{type(exc).__name__}: {exc}"
+            return False
+        self.index_error = None
+        return True
 
     # --- Routing ----------------------------------------------------------------------------
     def route(self, text: str, hint_skill: str | None = None) -> RouteDecision:
@@ -123,7 +130,17 @@ class Router:
             reasons.append("Keine routbaren Skills")
             signals.append(Signal.NO_SKILL)
             return None, None
-        qvec = self.embedder.embed([text])[0]
+        if self.index_error is not None and not self.refresh():
+            reasons.append(f"Embedding nicht verfügbar ({self.index_error}) – Stufe 0")
+            signals.append(Signal.NO_SKILL)
+            return None, None
+        try:
+            qvec = self.embedder.embed([text])[0]
+        except Exception as exc:  # noqa: BLE001
+            self.index_error = f"{type(exc).__name__}: {exc}"
+            reasons.append(f"Embedding nicht verfügbar ({self.index_error}) – Stufe 0")
+            signals.append(Signal.NO_SKILL)
+            return None, None
         hits = self.index.search(qvec, k=self.config.top_k, only=routable)
         if not hits:
             signals.append(Signal.NO_SKILL)

core/tests/test_router.py

@@ -131,3 +131,25 @@ def test_conflict_with_reranker_trusts_reranker(registry, library, embedder):
     assert Signal.SKILL_CONFLICT in d.signals
     assert "Reranking aktiv" in d.reasons
     assert d.skill == "rechnung-extrahieren" and d.stage == 1
+
+
+class FlakyEmbedder:
+    def __init__(self, inner):
+        self.inner = inner
+        self.down = True
+
+    def embed(self, texts):
+        if self.down:
+            raise ConnectionError("embed-default weg")
+        return self.inner.embed(texts)
+
+
+def test_embedding_outage_degrades_to_strong_and_recovers(registry, library, embedder):
+    flaky = FlakyEmbedder(embedder)
+    r = Router(registry, library, flaky)  # Start ohne Embedding darf nicht abstürzen
+    assert r.index_error
+    d = r.route(RECHNUNG)
+    assert d.stage == 0 and any("Embedding nicht verfügbar" in x for x in d.reasons)
+    flaky.down = False
+    d = r.route(RECHNUNG)
+    assert d.stage == 1 and r.index_error is None