From 311039d2cbc3d472ecd8aa1632e59d388a4021dc Mon Sep 17 00:00:00 2001
From: Charles Cheng <charlescheng@rezona.ai>
Date: Thu, 18 Jun 2026 16:06:38 +0800
Subject: [PATCH] fix(logging): fall back to adapter dict in generate_content
 logging

The google_genai generate_content adapter path (generate_content_provider_config
is None) routes through litellm.completion and hands the success-logging handler
an already transformed generate_content dict, with no raw httpx_response
recorded. The handler assumed a raw httpx.Response was always present and raised
ValueError('Google GenAI Generate Content: httpx_response is None'), crashing the
background logging worker even though the call itself returned HTTP 200.

Fall back to the transformed dict when no httpx_response is available, mirroring
the existing anthropic_messages logging handler. The native path still uses the
raw httpx_response as before.

Adds regression tests covering both the adapter (no httpx_response) and native
(httpx_response present) logging paths.
---
 litellm/litellm_core_utils/litellm_logging.py | 12 ++-
 .../test_generate_content_logging.py          | 84 +++++++++++++++++++
 2 files changed, 94 insertions(+), 2 deletions(-)
 create mode 100644 tests/test_litellm/google_genai/test_generate_content_logging.py

diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index 36d17596873..e09e1bfa275 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -3294,10 +3294,18 @@ class Logging(LiteLLMLoggingBaseClass):
         """
         import httpx
 
+        # The adapter path (generate_content_provider_config is None) routes
+        # through litellm.completion and hands us an already transformed
+        # generate_content dict with no raw httpx_response recorded. Fall back to
+        # that dict instead of crashing the logging worker (#30707); the native
+        # path still uses the raw httpx_response as before.
         httpx_response = self.model_call_details.get("httpx_response", None)
-        if httpx_response is None:
+        if httpx_response is not None and isinstance(httpx_response, httpx.Response):
+            dict_result = httpx_response.json()
+        elif isinstance(result, dict):
+            dict_result = result
+        else:
             raise ValueError("Google GenAI Generate Content: httpx_response is None")
-        dict_result = httpx_response.json()
         result = litellm.VertexGeminiConfig()._transform_google_generate_content_to_openai_model_response(
             completion_response=dict_result,
             model_response=litellm.ModelResponse(),
diff --git a/tests/test_litellm/google_genai/test_generate_content_logging.py b/tests/test_litellm/google_genai/test_generate_content_logging.py
new file mode 100644
index 00000000000..81bb1a32bc3
--- /dev/null
+++ b/tests/test_litellm/google_genai/test_generate_content_logging.py
@@ -0,0 +1,84 @@
+"""Logging-path tests for the google_genai generate_content adapter (#30707).
+
+The adapter path (``generate_content_provider_config is None``) routes through
+``litellm.completion`` and hands the success-logging handler an already
+transformed generate_content dict, with no raw ``httpx.Response`` recorded. The
+handler must fall back to that dict instead of crashing the logging worker.
+"""
+
+import os
+import sys
+import time
+
+import httpx
+import pytest
+
+sys.path.insert(0, os.path.abspath("../../.."))
+
+from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging
+from litellm.types.utils import ModelResponse
+
+
+def _make_logging_obj() -> LitellmLogging:
+    logging_obj = LitellmLogging(
+        model="gemini-3.1-flash-lite",
+        messages=[{"role": "user", "content": "Hello"}],
+        stream=False,
+        call_type="generate_content",
+        start_time=time.time(),
+        litellm_call_id="30707-test",
+        function_id="30707-test",
+    )
+    # set in update_environment_variables() during a real call; the response
+    # transform reads it, so provide it for the isolated handler test.
+    logging_obj.optional_params = {}
+    return logging_obj
+
+
+def _generate_content_dict() -> dict:
+    return {
+        "candidates": [
+            {
+                "content": {"parts": [{"text": "Hello there"}], "role": "model"},
+                "finishReason": "STOP",
+                "index": 0,
+                "safetyRatings": [],
+            }
+        ],
+        "usageMetadata": {
+            "promptTokenCount": 5,
+            "candidatesTokenCount": 2,
+            "totalTokenCount": 7,
+        },
+        # the adapter adds a convenience top-level text field
+        "text": "Hello there",
+    }
+
+
+def test_adapter_path_logs_without_httpx_response():
+    """No httpx_response (adapter path) must not raise; the transformed dict is used."""
+    logging_obj = _make_logging_obj()
+    logging_obj.model_call_details["httpx_response"] = None
+
+    result = logging_obj._handle_non_streaming_google_genai_generate_content_response_logging(
+        result=_generate_content_dict()
+    )
+
+    assert isinstance(result, ModelResponse)
+    assert result.choices[0].message.content == "Hello there"
+    assert result.usage.total_tokens == 7
+
+
+def test_httpx_response_path_still_used_when_present():
+    """When a raw httpx_response exists it is still the source of truth."""
+    logging_obj = _make_logging_obj()
+    logging_obj.model_call_details["httpx_response"] = httpx.Response(
+        status_code=200, json=_generate_content_dict()
+    )
+
+    result = logging_obj._handle_non_streaming_google_genai_generate_content_response_logging(
+        result=None
+    )
+
+    assert isinstance(result, ModelResponse)
+    assert result.choices[0].message.content == "Hello there"