fix(guardrails): keep every response step listed on generateContent

The generateContent adapter logs its nested completion, spend row included, before the response hooks run, so
dropping never-ran steps afterwards left the header and the trace naming fewer steps than the spend row. The route
now keeps main's list. The integration fixtures give each video its own id and accept any non-hit cache_hit value.

litellm/proxy/utils.py

@@ -744,11 +744,12 @@ def _steps_that_ran(result: PipelineExecutionResult) -> frozenset[str]:
 def _whole_response_log_waits_for_its_pipelines(data: Mapping[str, object], call_type: str) -> bool:
     """
     The success log is parked until the response hooks finish, and the headers are built
-    after them, except on the LLM pass-through route.
+    after them. The LLM pass-through route builds its headers first, and the generateContent
+    adapter logs its nested completion before the response hooks run.
     """
     logging_obj: Final = data.get("litellm_logging_obj")
     return (
-        call_type != CallTypes.allm_passthrough_route.value
+        call_type not in (CallTypes.allm_passthrough_route.value, CallTypes.agenerate_content.value)
         and isinstance(logging_obj, Logging)
         and logging_obj.enqueue_deferred_logging is not None
     )

tests/integration/observability/test_guardrail_pipeline_applied_header.py

@@ -90,7 +90,6 @@ _FINE_TUNING_JOB: Final[dict[str, JsonValue]] = {
     "validation_file": None,
 }
 _VIDEO: Final = {
-    "id": "video_pipeline",
     "object": "video",
     "model": "sora-2",
     "status": "queued",
@@ -240,7 +239,7 @@ def _completion(request: Request) -> Reply:
     if request.target in ("/v1/images/generations", "/v1/images/edits"):
         return Reply(body=json.dumps({"created": 1, "data": [{"b64_json": "aW1n", "revised_prompt": None}]}).encode())
     if request.target == "/v1/videos":
-        return Reply(body=json.dumps(_VIDEO).encode())
+        return Reply(body=json.dumps({**_VIDEO, "id": f"video_{uuid.uuid4().hex}"}).encode())
     if request.target == "/v1/fine_tuning/jobs":
         return Reply(body=json.dumps(_FINE_TUNING_JOB).encode())
     if request.target.split("?", 1)[0].endswith(":generateContent"):
@@ -823,7 +822,7 @@ def test_a_cache_hit_names_every_response_step_that_ran_in_the_spend_log_and_the
     expected: Final = [_RESPONSE_FIRST, _RESPONSE_SECOND]
     assert [first.status_code, cached.status_code] == [200, 200], (first.text, cached.text)
     assert _steps_that_ran(rig) == (_RESPONSE_FIRST, _RESPONSE_SECOND) * 2
-    assert [row["cache_hit"] for row in rows] == ["False", "True"], rows
+    assert [row["cache_hit"] == "True" for row in rows] == [False, True], rows
     assert [sorted(response.headers.get(_HEADER, "").split(",")) for response in (first, cached)] == [expected] * 2
     assert [sorted(object_value(row["metadata"]).get("applied_guardrails") or []) for row in rows] == [expected] * 2
     assert [sorted(ast.literal_eval(str(span["attributes"]["metadata.applied_guardrails"]))) for span in spans] == [

tests/unit/proxy/utils/proxy_logging/test_guardrail_pipeline.py

@@ -1366,6 +1366,9 @@ async def test_pre_call_hook_names_a_step_that_ran_in_the_litellm_metadata_a_lat
         pytest.param(
             True, "allm_passthrough_route", "metadata", ["first", "second"], id="pass-through-headers-already-out"
         ),
+        pytest.param(
+            True, "agenerate_content", "metadata", ["first", "second"], id="generate-content-adapter-logged-its-call"
+        ),
     ],
 )
 async def test_post_call_pipeline_drops_the_steps_that_did_not_run_only_while_nothing_has_read_them(