Add read-only production runtime audit

apps/ai-service/tests/test_api.py

@@ -5,7 +5,7 @@ from fastapi.testclient import TestClient
 from adapters.prometheus import PrometheusMetrics
 from adapters.postgres import FeedbackTraceNotFound, RetrievalTrace
 from config import Settings
-from main import create_app
+from main import _route_label, create_app
 from rag.agent import AgentReply
 from rag.answer import DISCLAIMER, Citation, GroundedAnswerService
 from rag.metrics import TRACE_WRITE_FAILED, InMemoryMetrics
@@ -150,6 +150,18 @@ def test_history_for_unknown_conversation_is_empty_not_an_error():
     assert response.json() == {"items": []}
 
 
+def test_history_rejects_an_oversized_conversation_id_before_querying_storage():
+    traces = FakeHistoryTraceWriter({})
+    app = create_app(settings=Settings(), trace_writer=traces)
+
+    response = TestClient(app).get(
+        "/v1/rag/history", params={"conversation_id": "x" * 129}
+    )
+
+    assert response.status_code == 422
+    assert traces.calls == []
+
+
 def test_health_and_fail_closed_rag_response_are_traced():
     traces = MemoryTraceWriter()
     app = create_app(
@@ -204,6 +216,19 @@ def _metrics_app(**settings_kwargs):
     )
 
 
+def test_all_public_rag_endpoints_have_bounded_request_metric_labels():
+    paths = (
+        "/v1/rag/query",
+        "/v1/rag/suggest",
+        "/v1/rag/feedback",
+        "/v1/rag/history",
+        "/v1/rag/sections",
+        "/v1/rag/section-text",
+    )
+
+    assert {_route_label(path) for path in paths} == set(paths)
+
+
 def test_metrics_stays_open_when_no_token_is_configured():
     """The default must not break the existing Compose scrape or local runs —
     the endpoint is not internet-reachable in that topology."""
@@ -252,10 +277,15 @@ class FakeAgent:
 
     def __init__(self, reply: AgentReply) -> None:
         self._reply = reply
-        self.calls: list[tuple[str, str | None]] = []
+        self.calls: list[tuple[str, str | None, str]] = []
 
-    def handle(self, turn: str, conversation_id: str | None = None) -> AgentReply:
-        self.calls.append((turn, conversation_id))
+    def handle(
+        self,
+        turn: str,
+        conversation_id: str | None = None,
+        response_mode: str = "ai",
+    ) -> AgentReply:
+        self.calls.append((turn, conversation_id, response_mode))
         return self._reply
 
     def complete(self, prefix: str, k: int = 8) -> list[str]:
@@ -291,7 +321,34 @@ def test_query_routes_through_the_agent_when_one_is_configured():
     assert body["answer"] == "Liều 500 mg [1]."
     assert body["resolved_drug_id"] == "metformin"
     assert len(body["citations"]) == 1
-    assert agent.calls == [("Liều metformin?", "c1")]
+    assert agent.calls == [("Liều metformin?", "c1", "ai")]
+
+
+def test_query_forwards_monograph_response_mode_to_agent():
+    agent = FakeAgent(AgentReply(
+        decision="clarify",
+        reason="select_drug_sections",
+        clarification="Chọn mục cần xem.",
+        drugs=("metformin",),
+        turn_type="drug_overview",
+    ))
+    app = create_app(
+        settings=Settings(),
+        answer_service=GroundedAnswerService(FixedRouting()),
+        conversational=agent,
+        trace_writer=MemoryTraceWriter(),
+    )
+
+    response = TestClient(app).post("/v1/rag/query", json={
+        "query": "Metformin",
+        "subject_scope": "human",
+        "intent": "fact_lookup",
+        "response_mode": "monograph",
+    })
+
+    assert response.status_code == 200
+    assert response.json()["reason"] == "select_drug_sections"
+    assert agent.calls == [("Metformin", None, "monograph")]
 
 
 def test_query_agent_clarification_is_surfaced_as_the_answer():