fix(guardrails): keep the privaite restore map out of background request state

Background Responses results are fetched by a later poll that the post-call
success and streaming hooks never see, so the fake-to-original map stashed in
request metadata would persist in stored request state with no consumer left
to pop it. The pre-call hook now skips the stash for background requests;
those responses keep their placeholders, which is the fail-closed behavior.
Regression test included.

litellm/proxy/guardrails/guardrail_hooks/privaite/privaite.py

@@ -282,10 +282,18 @@ class PrivaiteGuardrail(CustomGuardrail):
 
             raise HTTPException(status_code=400, detail={"error": str(exc)}) from exc
 
-        if mapping is not None and self.deanonymize and not mapping.is_empty:
+        if (
+            mapping is not None
+            and self.deanonymize
+            and not mapping.is_empty
+            and not data.get("background")
+        ):
             # Carry a plain fake->original dict to the post-call hook on the same
             # request, the same channel LiteLLM's Presidio guardrail uses. The
             # post-call hook pops it again so it does not linger in metadata.
+            # Background requests are excluded: their result is fetched by a later
+            # poll the post-call hooks never see, so the map would persist in
+            # request state with no consumer; those responses keep placeholders.
             data.setdefault("metadata", {})["privaite_map"] = dict(mapping.get_all_fakes())
         return data
 

tests/unit/proxy/guardrails/guardrail_hooks/test_privaite.py

@@ -347,6 +347,21 @@ async def test_pre_call_without_pii_does_not_stash():
     assert "privaite_map" not in out.get("metadata", {})
 
 
+@pytest.mark.asyncio
+async def test_pre_call_background_request_anonymizes_but_does_not_stash():
+    # Background results are fetched by a later poll the post-call hooks never
+    # see, so the fake->original map must not persist in request state.
+    gr = _make_guardrail()
+    data = {
+        "input": "Email Marie Dupont at marie@acme.com",
+        "background": True,
+    }
+    out = await gr.async_pre_call_hook(None, None, data, "aresponses")
+    assert "Marie Dupont" not in out["input"]
+    assert "<PERSON_1>" in out["input"]
+    assert "privaite_map" not in out.get("metadata", {})
+
+
 @pytest.mark.asyncio
 async def test_pre_call_clears_client_supplied_map():
     # metadata is caller-controlled: a client-supplied privaite_map must be