scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-11 22:51:28 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 123

Merge pull request #22904 from Chesars/claude/add-claude-param-default-M4Yic

Browse source 
feat(anthropic): add opt-out flag for default reasoning summary
This commit is contained in:
Cesar Garcia 2026-03-05 11:59:28 -03:00 • committed by GitHub
commit c9e60d9909
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 191 additions and 39 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
litellm/__init__.py
View file

@ -293,6 +293,7 @@ llm_guard_mode: Literal["all", "key-specific", "request-specific"] = "all"
guardrail_name_config_map: Dict[str, GuardrailItem] = {}
include_cost_in_streaming_usage: bool = False
reasoning_auto_summary: bool = False
disable_default_reasoning_summary: bool = False
### PROMPTS ####
from litellm.types.prompts.init_prompts import PromptSpec

18
litellm/llms/anthropic/experimental_pass_through/adapters/handler.py
View file

@ -12,6 +12,9 @@ from typing import (
)
import litellm
from litellm.llms.anthropic.experimental_pass_through.utils import (
is_default_reasoning_summary_disabled,
)
from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
AnthropicAdapter,
)
@ -77,22 +80,27 @@ class LiteLLMMessagesToCompletionTransformationHandler:
# Prefix model with "responses/" to route to OpenAI Responses API
completion_kwargs["model"] = f"responses/{model}"
summary_disabled = is_default_reasoning_summary_disabled()
reasoning_effort = completion_kwargs.get("reasoning_effort")
summary = thinking.get("summary")
if isinstance(reasoning_effort, str) and reasoning_effort:
reasoning_dict: Dict[str, Any] = {"effort": reasoning_effort}
if summary:
reasoning_dict["summary"] = summary
elif not summary_disabled:
reasoning_dict["summary"] = "detailed"
completion_kwargs["reasoning_effort"] = reasoning_dict
elif isinstance(reasoning_effort, dict):
if (
summary
and "summary" not in reasoning_effort
"summary" not in reasoning_effort
and "generate_summary" not in reasoning_effort
):
updated_reasoning_effort = dict(reasoning_effort)
updated_reasoning_effort["summary"] = summary
completion_kwargs["reasoning_effort"] = updated_reasoning_effort
effective_summary = summary if summary else ("detailed" if not summary_disabled else None)
if effective_summary:
updated_reasoning_effort = dict(reasoning_effort)
updated_reasoning_effort["summary"] = effective_summary
completion_kwargs["reasoning_effort"] = updated_reasoning_effort
@staticmethod
def _prepare_completion_kwargs(

8
litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py
View file

@ -8,6 +8,11 @@ path used for OpenAI and Azure models.
import json
from typing import Any, Dict, List, Optional, Union, cast
import litellm
from litellm.llms.anthropic.experimental_pass_through.utils import (
is_default_reasoning_summary_disabled,
)
from litellm.types.llms.anthropic import (
AllAnthropicToolsValues,
AnthopicMessagesAssistantMessageParam,
@ -241,10 +246,13 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
effort = "low"
else:
effort = "minimal"
summary_disabled = is_default_reasoning_summary_disabled()
result: Dict[str, Any] = {"effort": effort}
summary = thinking.get("summary")
if summary:
result["summary"] = summary
elif not summary_disabled:
result["summary"] = "detailed"
return result
def translate_request(

12
litellm/llms/anthropic/experimental_pass_through/utils.py Normal file
View file

@ -0,0 +1,12 @@
import os
import litellm
def is_default_reasoning_summary_disabled() -> bool:
"""Check whether the default 'summary: detailed' injection should be suppressed."""
return (
litellm.disable_default_reasoning_summary
or os.getenv("LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY", "false").lower()
== "true"
)

159
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py
View file

@ -180,7 +180,8 @@ def test_openai_model_with_thinking_converts_to_reasoning():
assert "reasoning" in call_kwargs, "reasoning should be passed to litellm.responses"
# budget_tokens=1024 -> effort="minimal" (< 2000 threshold)
expected_reasoning = {"effort": "minimal"}
# summary="detailed" added by default unless disable_default_reasoning_summary is set
expected_reasoning = {"effort": "minimal", "summary": "detailed"}
assert call_kwargs["reasoning"] == expected_reasoning, (
f"reasoning should be {expected_reasoning} for budget_tokens=1024, "
f"got {call_kwargs.get('reasoning')}"
@ -225,7 +226,7 @@ class TestThinkingParameterTransformation:
class TestThinkingSummaryPreservation:
"""Tests for issue #20998: thinking.summary must be preserved when routing to OpenAI Responses API."""
"""Tests for thinking.summary preservation and disable_default_reasoning_summary flag."""
def test_thinking_summary_concise_preserved_for_openai(self):
"""User-provided summary='concise' should not be replaced with 'detailed'."""
@ -253,26 +254,123 @@ class TestThinkingSummaryPreservation:
)
assert completion_kwargs["reasoning_effort"] == {"effort": "high", "summary": "auto"}
def test_thinking_without_summary_does_not_inject_summary(self):
"""When no summary is provided, no summary should be injected (opt-in per OpenAI spec)."""
def test_summary_added_by_default_when_no_user_summary(self):
"""When no user summary and flag is off, summary='detailed' is added by default."""
import litellm
from litellm.llms.anthropic.experimental_pass_through.adapters.handler import (
LiteLLMMessagesToCompletionTransformationHandler,
)
thinking = {"type": "enabled", "budget_tokens": 5000}
completion_kwargs = {"model": "openai/gpt-5.1", "reasoning_effort": "medium"}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking=thinking
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = False
completion_kwargs = {
"model": "responses/gpt-5.2",
"custom_llm_provider": "openai",
"reasoning_effort": "medium",
}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking={"type": "enabled", "budget_tokens": 5000}
)
assert completion_kwargs["reasoning_effort"] == {"effort": "medium", "summary": "detailed"}
finally:
litellm.disable_default_reasoning_summary = original
def test_summary_excluded_when_disable_flag_set_string_reasoning(self):
"""When disable_default_reasoning_summary is True, summary is not added for string reasoning_effort."""
import litellm
from litellm.llms.anthropic.experimental_pass_through.adapters.handler import (
LiteLLMMessagesToCompletionTransformationHandler,
)
assert completion_kwargs["reasoning_effort"] == {"effort": "medium"}
assert "summary" not in completion_kwargs["reasoning_effort"]
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = True
completion_kwargs = {
"model": "responses/gpt-5.2",
"custom_llm_provider": "openai",
"reasoning_effort": "high",
}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking={"type": "enabled", "budget_tokens": 10000}
)
assert completion_kwargs["reasoning_effort"] == {"effort": "high"}
assert "summary" not in completion_kwargs["reasoning_effort"]
finally:
litellm.disable_default_reasoning_summary = original
def test_summary_excluded_when_disable_flag_set_dict_reasoning(self):
"""When disable_default_reasoning_summary is True, summary is not injected into dict reasoning_effort."""
import litellm
from litellm.llms.anthropic.experimental_pass_through.adapters.handler import (
LiteLLMMessagesToCompletionTransformationHandler,
)
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = True
completion_kwargs = {
"model": "responses/gpt-5.2",
"custom_llm_provider": "openai",
"reasoning_effort": {"effort": "medium"},
}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking={"type": "enabled", "budget_tokens": 5000}
)
assert completion_kwargs["reasoning_effort"] == {"effort": "medium"}
assert "summary" not in completion_kwargs["reasoning_effort"]
finally:
litellm.disable_default_reasoning_summary = original
def test_summary_excluded_when_env_var_set(self):
"""When LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY env var is true, summary is not added."""
import litellm
from litellm.llms.anthropic.experimental_pass_through.adapters.handler import (
LiteLLMMessagesToCompletionTransformationHandler,
)
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = False
os.environ["LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY"] = "true"
completion_kwargs = {
"model": "responses/gpt-5.2",
"custom_llm_provider": "openai",
"reasoning_effort": "high",
}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking={"type": "enabled", "budget_tokens": 10000}
)
assert completion_kwargs["reasoning_effort"] == {"effort": "high"}
assert "summary" not in completion_kwargs["reasoning_effort"]
finally:
litellm.disable_default_reasoning_summary = original
os.environ.pop("LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY", None)
def test_user_provided_summary_preserved_even_when_flag_off(self):
"""When user already set summary in dict reasoning_effort, it's preserved regardless of flag."""
import litellm
from litellm.llms.anthropic.experimental_pass_through.adapters.handler import (
LiteLLMMessagesToCompletionTransformationHandler,
)
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = False
completion_kwargs = {
"model": "responses/gpt-5.2",
"custom_llm_provider": "openai",
"reasoning_effort": {"effort": "high", "summary": "concise"},
}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking={"type": "enabled", "budget_tokens": 10000}
)
assert completion_kwargs["reasoning_effort"]["summary"] == "concise"
finally:
litellm.disable_default_reasoning_summary = original
def test_openai_model_with_thinking_summary_end_to_end(self):
"""End-to-end: anthropic_messages_handler should preserve thinking.summary for OpenAI models.
OpenAI models are routed to litellm.responses(), so we verify the
reasoning dict passed to it contains the user's summary value.
"""
"""End-to-end: anthropic_messages_handler should preserve thinking.summary for OpenAI models."""
from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
anthropic_messages_handler,
)
@ -309,16 +407,22 @@ class TestThinkingSummaryPreservation:
result = LiteLLMAnthropicToResponsesAPIAdapter.translate_thinking_to_reasoning(thinking)
assert result == {"effort": "medium", "summary": "concise"}
def test_responses_adapter_no_summary_when_not_provided(self):
"""translate_thinking_to_reasoning should not include summary when not provided."""
def test_responses_adapter_no_summary_when_disabled(self):
"""translate_thinking_to_reasoning should not include summary when flag is set and no user summary."""
import litellm
from litellm.llms.anthropic.experimental_pass_through.responses_adapters.transformation import (
LiteLLMAnthropicToResponsesAPIAdapter,
)
thinking = {"type": "enabled", "budget_tokens": 5000}
result = LiteLLMAnthropicToResponsesAPIAdapter.translate_thinking_to_reasoning(thinking)
assert result == {"effort": "medium"}
assert "summary" not in result
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = True
thinking = {"type": "enabled", "budget_tokens": 5000}
result = LiteLLMAnthropicToResponsesAPIAdapter.translate_thinking_to_reasoning(thinking)
assert result == {"effort": "medium"}
assert "summary" not in result
finally:
litellm.disable_default_reasoning_summary = original
def test_translate_thinking_for_model_preserves_summary(self):
"""translate_thinking_for_model should include summary in reasoning_effort dict when user provides it."""
@ -332,16 +436,3 @@ class TestThinkingSummaryPreservation:
model="openai/gpt-5.2",
)
assert result == {"reasoning_effort": {"effort": "medium", "summary": "concise"}}
def test_translate_thinking_for_model_no_summary_when_not_provided(self):
"""translate_thinking_for_model should return plain string reasoning_effort when no summary provided."""
from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
LiteLLMAnthropicMessagesAdapter,
)
thinking = {"type": "enabled", "budget_tokens": 5000}
result = LiteLLMAnthropicMessagesAdapter.translate_thinking_for_model(
thinking=thinking,
model="openai/gpt-5.2",
)
assert result == {"reasoning_effort": "medium"}

32
tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py
View file

@ -658,6 +658,38 @@ class TestTranslateThinkingToReasoning:
result = _ADAPTER.translate_thinking_to_reasoning({"type": "enabled"})
assert result == {"effort": "minimal", "summary": "detailed"}
def test_summary_excluded_when_disable_flag_set(self):
"""When disable_default_reasoning_summary is True, summary is not included."""
import litellm
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = True
result = _ADAPTER.translate_thinking_to_reasoning(
{"type": "enabled", "budget_tokens": 10000}
)
assert result == {"effort": "high"}
assert "summary" not in result
finally:
litellm.disable_default_reasoning_summary = original
def test_summary_excluded_when_env_var_set(self):
"""When LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY env var is true, summary is not included."""
import litellm
original = litellm.disable_default_reasoning_summary
try:
litellm.disable_default_reasoning_summary = False
os.environ["LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY"] = "true"
result = _ADAPTER.translate_thinking_to_reasoning(
{"type": "enabled", "budget_tokens": 5000}
)
assert result == {"effort": "medium"}
assert "summary" not in result
finally:
litellm.disable_default_reasoning_summary = original
os.environ.pop("LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY", None)
# ---------------------------------------------------------------------------
# translate_request – broader coverage