Expose DuckLM runtime status

tests/smoke/test_api_health.py

@@ -1,6 +1,8 @@
 from fastapi.testclient import TestClient
 
 from duck_core.api import create_app
+from duck_core.memory.vector_memory import VectorMemory
+from duck_core.model_client import ModelClient
 
 
 def test_health_and_status_endpoints(tmp_path, monkeypatch):
@@ -12,6 +14,59 @@ def test_health_and_status_endpoints(tmp_path, monkeypatch):
     status = client.get("/v1/status").json()
     assert status["name"] == "DuckLM"
     assert status["api_host"] == "127.0.0.1"
+    assert status["api"]["host"] == "127.0.0.1"
+    assert status["api"]["port"] == 8000
+    assert status["paths"]["db_path"] == str(tmp_path / "duck.sqlite3")
+    assert status["models"]["default_provider"] == "llama_server"
+    assert status["models"]["roles"]["thinker"]["model"] == "local-main"
+    assert status["services"]["duck_api"]["ok"] is True
+    assert status["services"]["llama"]["probed"] is False
+    assert status["services"]["vector_memory"]["probed"] is False
+
+
+def test_status_endpoint_can_probe_backends(tmp_path, monkeypatch):
+    monkeypatch.setenv("DUCK_DB_PATH", str(tmp_path / "duck.sqlite3"))
+
+    async def fake_ping(self):
+        return {
+            "thinker": {
+                "ok": True,
+                "base_url": "http://127.0.0.1:8081/v1",
+                "model": "local-main",
+                "latency_ms": 1.2,
+            },
+            "critic": {
+                "ok": False,
+                "base_url": "http://127.0.0.1:8081/v1",
+                "model": "local-main",
+                "error": "offline",
+            },
+        }
+
+    async def fake_vector_health(self):
+        return {
+            "configured": True,
+            "ok": True,
+            "qdrant_url": "http://127.0.0.1:6333",
+            "collection": "duck_memory",
+            "embedding_source": "local:./models/all-MiniLM-L6-v2",
+            "latency_ms": 2.3,
+        }
+
+    monkeypatch.setattr(ModelClient, "ping", fake_ping)
+    monkeypatch.setattr(VectorMemory, "health", fake_vector_health)
+
+    app = create_app()
+    client = TestClient(app)
+
+    status = client.get("/v1/status?probe=true").json()
+
+    assert status["services"]["llama"]["probed"] is True
+    assert status["services"]["llama"]["ok"] is False
+    assert status["services"]["llama"]["roles"]["thinker"]["ok"] is True
+    assert status["services"]["llama"]["roles"]["critic"]["ok"] is False
+    assert status["services"]["vector_memory"]["probed"] is True
+    assert status["services"]["vector_memory"]["ok"] is True
 
 
 def test_webchat_index_renders(tmp_path, monkeypatch):