scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-10 03:28:53 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 85

fix(cache): align Anthropic reasoning_content with live response on replay

Browse source
This commit is contained in:
nehaaprasaad 2026-05-07 07:56:42 +05:30
parent bd1ea0252a
commit 5313a9c67f
2 changed files with 47 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
View file

@ -595,10 +595,9 @@ def convert_to_model_response_object( # noqa: PLR0915
thinking_blocks = choice["message"]["thinking_blocks"]
provider_specific_fields["thinking_blocks"] = thinking_blocks
# Align with live provider responses: reasoning_content is message-level only.
if reasoning_content:
provider_specific_fields["reasoning_content"] = (
reasoning_content
)
provider_specific_fields.pop("reasoning_content", None)
message = Message(
content=content,

46
tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
View file

@ -902,7 +902,51 @@ def test_convert_to_model_response_object_with_thinking_content():
resp: ModelResponse = convert_to_model_response_object(**args)
assert resp is not None
assert resp.choices[0].message.reasoning_content is not None
msg = resp.choices[0].message
assert msg.reasoning_content is not None
assert "reasoning_content" not in (msg.provider_specific_fields or {})
def test_convert_to_model_response_object_strips_reasoning_content_from_provider_specific_fields():
"""
Stale or merged provider_specific_fields must not duplicate top-level reasoning_content.
Duplication breaks disk cache keys on later turns (see Anthropic extended thinking).
"""
duplicate = "same reasoning"
args = {
"response_object": {
"id": "chatcmpl-dup-reasoning",
"created": 1741057687,
"model": "claude-4-sonnet-20250514",
"object": "chat.completion",
"choices": [
{
"finish_reason": "stop",
"index": 0,
"message": {
"content": "Answer.",
"role": "assistant",
"reasoning_content": duplicate,
"provider_specific_fields": {
"citations": None,
"reasoning_content": duplicate,
"thinking_blocks": [],
},
},
}
],
"usage": {
"completion_tokens": 1,
"prompt_tokens": 1,
"total_tokens": 2,
},
},
"model_response_object": ModelResponse(),
}
resp = convert_to_model_response_object(**args)
msg = resp.choices[0].message
assert msg.reasoning_content == duplicate
assert "reasoning_content" not in (msg.provider_specific_fields or {})
def test_convert_to_model_response_object_with_empty_error_object():