From 2fead4f6aa0064434ac649c0ea8a9f22686e1be3 Mon Sep 17 00:00:00 2001
From: jibanez-staticduo <staticduo@gmail.com>
Date: Thu, 1 Oct 2026 04:39:11 +0200
Subject: [PATCH] fix(live): filter terminal response content in realtime logs

---
 .../litellm_core_utils/realtime_streaming.py  | 24 +++++++++++-
 .../test_realtime_streaming.py                | 38 +++++++++++++++++++
 2 files changed, 61 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/realtime_streaming.py b/litellm/litellm_core_utils/realtime_streaming.py
index 1713bc11102..98dd7abf903 100644
--- a/litellm/litellm_core_utils/realtime_streaming.py
+++ b/litellm/litellm_core_utils/realtime_streaming.py
@@ -283,7 +283,29 @@ class RealTimeStreaming:
         if message_obj.get("type") == "response.event" and isinstance(message_obj.get("event"), dict):
             nested: Final = message_obj["event"]
             if nested.get("type") in ("response.completed", "response.incomplete", "response.failed"):
-                self.messages.append(TypeAdapter(OpenAILiveResponseEvent).validate_python(message_obj))
+                response: Final = nested.get("response")
+                # Retain billing evidence even when response content is excluded from logging.
+                stored: Final = (
+                    message_obj
+                    if self._should_store_message(message_obj)
+                    else {
+                        "type": "response.event",
+                        "event": {
+                            "type": nested["type"],
+                            "response": {
+                                **{
+                                    key: value
+                                    for key, value in response.items()
+                                    if key in ("id", "created_at", "model", "usage", "service_tier")
+                                },
+                                "output": [],
+                            }
+                            if isinstance(response, Mapping)
+                            else None,
+                        },
+                    }
+                )
+                self.messages.append(TypeAdapter(OpenAILiveResponseEvent).validate_python(stored))
             return
         if not self._should_store_message(message_obj):
             return
diff --git a/tests/unit/litellm_core_utils/test_realtime_streaming.py b/tests/unit/litellm_core_utils/test_realtime_streaming.py
index d74f3fcf89b..fbd3639eac0 100644
--- a/tests/unit/litellm_core_utils/test_realtime_streaming.py
+++ b/tests/unit/litellm_core_utils/test_realtime_streaming.py
@@ -3640,6 +3640,7 @@ def test_public_live_accounting_survives_filtered_logging(monkeypatch):
                 "response": {
                     "id": "resp_one",
                     "model": "gpt-backend",
+                    "output": [],
                     "usage": {"total_tokens": 12},
                 },
             },
@@ -3654,6 +3655,43 @@ def test_public_live_accounting_survives_filtered_logging(monkeypatch):
     assert stream.messages == events
 
 
+@pytest.mark.parametrize("terminal", ["response.completed", "response.incomplete", "response.failed"])
+@pytest.mark.parametrize("allowed", [[], ["response.event"], "*"])
+def test_live_terminal_logging_filters_content_and_preserves_accounting(
+    monkeypatch: pytest.MonkeyPatch, terminal: str, allowed: list[str] | str
+) -> None:
+    from litellm.cost_calculator import _live_backend_responses
+
+    monkeypatch.setattr(litellm, "logged_real_time_event_types", allowed)
+    stream = RealTimeStreaming(MagicMock(), MagicMock(), MagicMock())
+    response = {
+        "id": "resp_private",
+        "created_at": 1,
+        "model": "gpt-backend",
+        "output": [{"type": "message", "content": [{"type": "output_text", "text": "private answer"}]}],
+        "instructions": "private instructions",
+        "metadata": {"private": "metadata"},
+        "usage": {"input_tokens": 20, "output_tokens": 10, "total_tokens": 30},
+    }
+    event = {"type": "response.event", "event": {"type": terminal, "response": response}}
+    stream.store_message(event)
+
+    stored = stream.messages[0]["event"]["response"]
+    if allowed:
+        assert stored == response
+    else:
+        assert stored == {
+            key: value for key, value in response.items() if key not in ("output", "instructions", "metadata")
+        } | {"output": []}
+    measured = _live_backend_responses(stream.messages)
+    assert len(measured) == 1
+    assert measured[0].id == "resp_private"
+    assert measured[0].model == "gpt-backend"
+    assert measured[0].usage.total_tokens == 30
+    assert response["instructions"] == "private instructions"
+    assert response["output"][0]["content"][0]["text"] == "private answer"
+
+
 @pytest.mark.parametrize("account_usage,expected", [(True, 1), (False, 0)])
 def test_live_initialization_is_retained_only_by_accounting_owner(account_usage, expected):
     stream = RealTimeStreaming(