fix(spend_logs): store request prompt in LiteLLM_SpendLogs.messages for all call types

Resolves #34747

litellm/litellm_core_utils/redact_messages.py

@@ -10,7 +10,7 @@
 import asyncio
 import copy
 import inspect
-from typing import TYPE_CHECKING, Any, Optional
+from typing import TYPE_CHECKING, Any, Final, Optional
 
 import litellm
 from litellm.integrations.custom_logger import CustomLogger
@@ -34,6 +34,9 @@ else:
     LiteLLMLoggingObject = Any
 
 
+REDACTED_MESSAGES_PLACEHOLDER: Final[tuple[dict[str, str], ...]] = ({"role": "user", "content": "redacted-by-litellm"},)
+
+
 def redact_message_input_output_from_custom_logger(
     litellm_logging_obj: LiteLLMLoggingObject, result, custom_logger: CustomLogger
 ):
@@ -235,7 +238,7 @@ def perform_redaction(model_call_details: dict, result, redact_streaming_respons
     copy via redact_streaming_responses_for_custom_logger instead.
     """
     # Redact model_call_details
-    model_call_details["messages"] = [{"role": "user", "content": "redacted-by-litellm"}]
+    model_call_details["messages"] = list(REDACTED_MESSAGES_PLACEHOLDER)
     model_call_details["prompt"] = ""
     model_call_details["input"] = ""
     _redact_standard_logging_object(model_call_details)

litellm/proxy/spend_tracking/spend_tracking_utils.py

@@ -441,7 +441,7 @@ def get_logging_payload(kwargs, response_obj, start_time, end_time) -> SpendLogs
             requester_ip_address=clean_metadata.get("requester_ip_address", None),
             custom_llm_provider=kwargs.get("custom_llm_provider", ""),
             messages=_get_messages_for_spend_logs_payload(
-                standard_logging_payload=standard_logging_payload, metadata=metadata
+                standard_logging_payload=standard_logging_payload, kwargs=kwargs
             ),
             response=_get_response_for_spend_logs_payload(payload=standard_logging_payload, kwargs=kwargs),
             proxy_server_request=_get_proxy_server_request_for_spend_logs_payload(
@@ -655,19 +655,45 @@ async def get_spend_by_team_and_customer(
 
 def _get_messages_for_spend_logs_payload(
     standard_logging_payload: Optional[StandardLoggingPayload],
-    metadata: Optional[dict] = None,
+    kwargs: Optional[dict] = None,
 ) -> str:
-    if _should_store_prompts_and_responses_in_spend_logs():
-        if standard_logging_payload is not None:
-            call_type = standard_logging_payload.get("call_type", "")
-            if call_type == "_arealtime":
-                messages = standard_logging_payload.get("messages")
-                if messages is not None:
-                    try:
-                        return safe_dumps(messages)
-                    except Exception:
-                        return "{}"
-    return "{}"
+    """
+    Serialize the request prompt into ``LiteLLM_SpendLogs.messages``.
+
+    Only stored when ``store_prompts_in_spend_logs`` is enabled, and replaced
+    with a placeholder when ``turn_off_message_logging`` redaction applies.
+    """
+    if standard_logging_payload is None or not _should_store_prompts_and_responses_in_spend_logs():
+        return "{}"
+
+    messages: Any = standard_logging_payload.get("messages")
+    if messages is None:
+        return "{}"
+
+    if kwargs is not None:
+        from litellm.litellm_core_utils.redact_messages import (
+            REDACTED_MESSAGES_PLACEHOLDER,
+            should_redact_message_logging,
+        )
+
+        model_call_details = {
+            "litellm_params": kwargs.get("litellm_params", {}),
+            "standard_callback_dynamic_params": kwargs.get("standard_callback_dynamic_params"),
+        }
+        if should_redact_message_logging(model_call_details=model_call_details):
+            return safe_dumps(list(REDACTED_MESSAGES_PLACEHOLDER))
+
+    try:
+        sanitized = _sanitize_request_body_for_spend_logs_payload({"messages": messages}).get("messages")
+        messages_json_str = safe_dumps(sanitized)
+    except Exception:
+        return "{}"
+    if LITELLM_TRUNCATED_PAYLOAD_FIELD in messages_json_str:
+        verbose_proxy_logger.info(
+            "Spend Log: messages were truncated before storing in DB. %s",
+            LITELLM_TRUNCATION_DB_SAFEGUARD_NOTE,
+        )
+    return messages_json_str
 
 
 _SENSITIVE_REQUEST_BODY_KEYS = frozenset({"secret_fields"})

tests/logging_callback_tests/test_spend_logs.py

@@ -388,6 +388,9 @@ def test_spend_logs_payload_with_prompts_enabled(monkeypatch):
     assert payload["response"] == json.dumps(
         {"role": "assistant", "content": "Hi there!"}
     )
+    assert json.loads(payload["messages"] or "{}") == [
+        {"role": "user", "content": "Hello!"}
+    ]
     proxy_server_request = json.loads(payload["proxy_server_request"] or "{}")
     assert proxy_server_request["model"] == "gpt-5.5"
     assert proxy_server_request["messages"] == [{"role": "user", "content": "Hello!"}]

tests/test_litellm/proxy/spend_tracking/test_spend_tracking_utils.py

@@ -544,21 +544,85 @@ def test_get_messages_for_spend_logs_realtime_empty_when_disabled(mock_should_st
 @patch(
     "litellm.proxy.spend_tracking.spend_tracking_utils._should_store_prompts_and_responses_in_spend_logs"
 )
-def test_get_messages_for_spend_logs_non_realtime_returns_empty(mock_should_store):
+@pytest.mark.parametrize("call_type", ["acompletion", "aresponses"])
+def test_get_messages_for_spend_logs_stores_prompt_for_all_call_types(
+    mock_should_store, call_type
+):
     """
-    Test that _get_messages_for_spend_logs_payload returns '{}' for non-realtime
-    calls even when store_prompts_in_spend_logs is True.
+    Regression for #34747: with store_prompts_in_spend_logs enabled,
+    LiteLLM_SpendLogs.messages must hold the request prompt for regular
+    /chat/completions and /responses traffic, not just realtime calls.
+    """
+    mock_should_store.return_value = True
+    payload = cast(
+        StandardLoggingPayload,
+        {
+            "call_type": call_type,
+            "messages": [{"role": "user", "content": "Hello"}],
+        },
+    )
+    result = _get_messages_for_spend_logs_payload(payload)
+    assert json.loads(result) == [{"role": "user", "content": "Hello"}]
+
+
+@patch(
+    "litellm.proxy.spend_tracking.spend_tracking_utils._should_store_prompts_and_responses_in_spend_logs"
+)
+def test_get_messages_for_spend_logs_truncates_large_prompt(mock_should_store):
+    """
+    Oversized prompts must be capped by the same DB-storage safeguard used for
+    the request body and response so a single row can't blow up.
+    """
+    from litellm.constants import (
+        LITELLM_TRUNCATED_PAYLOAD_FIELD,
+        MAX_STRING_LENGTH_PROMPT_IN_DB,
+    )
+
+    mock_should_store.return_value = True
+    large_content = "A" * (MAX_STRING_LENGTH_PROMPT_IN_DB + 500)
+    payload = cast(
+        StandardLoggingPayload,
+        {
+            "call_type": "acompletion",
+            "messages": [{"role": "user", "content": large_content}],
+        },
+    )
+
+    result = _get_messages_for_spend_logs_payload(payload)
+
+    assert LITELLM_TRUNCATED_PAYLOAD_FIELD in result
+    assert len(json.loads(result)[0]["content"]) < len(large_content)
+
+
+@patch(
+    "litellm.proxy.spend_tracking.spend_tracking_utils._should_store_prompts_and_responses_in_spend_logs"
+)
+def test_get_messages_for_spend_logs_redacts_when_message_logging_off(
+    mock_should_store,
+):
+    """
+    turn_off_message_logging must win over store_prompts_in_spend_logs, matching
+    how the response and proxy_server_request columns behave.
     """
     mock_should_store.return_value = True
     payload = cast(
         StandardLoggingPayload,
         {
             "call_type": "acompletion",
-            "messages": [{"role": "user", "content": "Hello"}],
+            "messages": [{"role": "user", "content": "secret prompt"}],
         },
     )
-    result = _get_messages_for_spend_logs_payload(payload)
-    assert result == "{}"
+
+    result = _get_messages_for_spend_logs_payload(
+        payload,
+        kwargs={
+            "litellm_params": {},
+            "standard_callback_dynamic_params": {"turn_off_message_logging": True},
+        },
+    )
+
+    assert "secret prompt" not in result
+    assert json.loads(result) == [{"role": "user", "content": "redacted-by-litellm"}]
 
 
 @patch(