From 307b47190295b4438b57f04d22f5aeb2a7e29f42 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 28 Aug 2026 16:45:45 -0700
Subject: [PATCH] style(guardrails): drop narrative comments from the deferred
 logging path

---
 .../messages/streaming_iterator.py                          | 6 ------
 litellm/proxy/common_request_processing.py                  | 5 -----
 litellm/proxy/utils.py                                      | 6 ------
 3 files changed, 17 deletions(-)

diff --git a/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py b/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py
index e0ddd54c24f..a282d5f4d4f 100644
--- a/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py
+++ b/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py
@@ -367,12 +367,6 @@ class BaseAnthropicMessagesStreamingIterator:
         deferred_dispatch_armed: Final = (
             getattr(self.litellm_logging_obj, "_on_deferred_stream_complete", None) is not None
         )
-        # Post-call guardrails run their end-of-stream scan AFTER this iterator
-        # is exhausted, so enqueueing now would build the spend log before the
-        # scan writes guardrail_information. Park the coroutine instead; the
-        # proxy fires it via _fire_deferred_stream_logging once the guardrail
-        # chain drains. Teardown (client disconnect) keeps enqueueing
-        # immediately: the scan never runs there and billing must not be lost.
         if deferred_dispatch_armed and not stream_teardown:
             self.litellm_logging_obj._deferred_stream_complete_args = (logging_coroutine,)
             return
diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py
index 936ca1607b5..ff6c8d1b1f8 100644
--- a/litellm/proxy/common_request_processing.py
+++ b/litellm/proxy/common_request_processing.py
@@ -2377,11 +2377,6 @@ class ProxyBaseLLMRequestProcessing:
                     and route_type == "anthropic_messages"
                     and self._is_streaming_response(response)
                 ):
-                    # Native /v1/messages SSE streams bypass CSW, so the raw
-                    # iterator parks its logging coroutine at stream end (see
-                    # BaseAnthropicMessagesStreamingIterator._handle_streaming_logging)
-                    # and _fire_deferred_stream_logging hands it here after the
-                    # guardrail end-of-stream scans complete.
                     from litellm.litellm_core_utils.logging_worker import (
                         GLOBAL_LOGGING_WORKER,
                     )
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 516a50610d0..9fbe1b4bd06 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -3227,12 +3227,6 @@ class ProxyLogging:
                     ),
                 )
 
-        # Actually iterate through the chained async generator and yield chunks.
-        # A guardrail block raised after upstream exhaustion (e.g.
-        # unified_guardrail re-raising HTTPException) must still flush any
-        # parked deferred logging, or the blocked stream loses its spend log.
-        # GeneratorExit/CancelledError stay untouched: disconnect cleanup owns
-        # those.
         try:
             async for chunk in current_response:
                 yield chunk