From 294bc7eba0d62e4249c811c3bb4308b7d3ff00b4 Mon Sep 17 00:00:00 2001
From: Ishaan <ishaangupta0408@gmail.com>
Date: Fri, 14 Aug 2026 09:11:02 +0000
Subject: [PATCH] fix: preserve prompt_tokens_details when raw
 PromptTokensDetails reaches chunk aggregator

## TLDR

Signed-off-by: Ishaan <ishaangupta0408@gmail.com>
---
 .../streaming_chunk_builder_utils.py          |  3 +++
 .../test_streaming_chunk_builder_utils.py     | 21 +++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py
index ab4017b144b..4d11205aa8c 100644
--- a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py
+++ b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py
@@ -27,6 +27,7 @@ from litellm.types.utils import (
     ServerToolUse,
     Usage,
 )
+from openai.types.completion_usage import PromptTokensDetails
 from litellm.utils import print_verbose, token_counter
 
 if TYPE_CHECKING:
@@ -650,6 +651,8 @@ class ChunkProcessor:
                 prompt_tokens_details = PromptTokensDetailsWrapper(**usage_chunk.prompt_tokens_details)
             elif isinstance(usage_chunk.prompt_tokens_details, PromptTokensDetailsWrapper):
                 prompt_tokens_details = usage_chunk.prompt_tokens_details
+            elif isinstance(usage_chunk.prompt_tokens_details, PromptTokensDetails):
+                prompt_tokens_details = PromptTokensDetailsWrapper(**usage_chunk.prompt_tokens_details.model_dump())
 
         return {
             "prompt_tokens": prompt_tokens,
diff --git a/tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py b/tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py
index 10bd22689d0..8d3b728dd11 100644
--- a/tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py
+++ b/tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py
@@ -1076,6 +1076,27 @@ def test_prompt_tokens_details_survive_later_usage_chunk_without_details():
     assert usage.prompt_tokens_details.cache_write_tokens == 10
 
 
+def test_prompt_tokens_details_base_class_preserved_in_usage_chunk():
+    """Regression for #36882: when the raw streaming usage chunk carries a
+    base-class PromptTokensDetails (not PromptTokensDetailsWrapper), the
+    helper must still extract cached_tokens rather than silently dropping the
+    whole prompt_tokens_details field."""
+    from openai.types.completion_usage import PromptTokensDetails as RawPromptTokensDetails
+
+    from litellm.litellm_core_utils.streaming_chunk_builder_utils import ChunkProcessor
+    from litellm.types.utils import Usage
+
+    raw_details = RawPromptTokensDetails(cached_tokens=0)
+    usage_chunk = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150)
+    usage_chunk.prompt_tokens_details = raw_details  # type: ignore[assignment]
+
+    processor = ChunkProcessor.__new__(ChunkProcessor)
+    result = processor._usage_chunk_calculation_helper(usage_chunk)
+
+    assert result["prompt_tokens_details"] is not None
+    assert result["prompt_tokens_details"].cached_tokens == 0
+
+
 def test_get_combined_tool_content_custom_tool_call():
     from litellm.litellm_core_utils.streaming_chunk_builder_utils import ChunkProcessor
     from litellm.types.utils import ChatCompletionMessageCustomToolCall