diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py index 98322aecf6a..47a3a903d82 100644 --- a/litellm/completion_extras/litellm_responses_transformation/transformation.py +++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py @@ -3,7 +3,6 @@ Handler for transforming /chat/completions api requests to litellm.responses req """ import json -import os from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence from itertools import accumulate, chain from types import MappingProxyType @@ -43,8 +42,8 @@ from litellm.responses.sse_output_recovery import ( ) from litellm.responses.utils import ( ResponsesAPIRequestUtils, - auto_reasoning_summary, normalize_responses_api_stream_options, + opt_in_auto_reasoning_summary, ) from litellm.types.llms.openai import ( ChatCompletionAnnotation, @@ -1745,19 +1744,11 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): **cast(Reasoning, reasoning_effort) # cast-ok: dict is forwarded verbatim to the provider ) - # Check if auto-summary is enabled via flag or environment variable - # Priority: litellm.reasoning_auto_summary flag > LITELLM_REASONING_AUTO_SUMMARY env var - auto_summary_enabled: Final = ( - litellm.reasoning_auto_summary or os.getenv("LITELLM_REASONING_AUTO_SUMMARY", "false").lower() == "true" - ) - + summary: Final = opt_in_auto_reasoning_summary(model, custom_llm_provider) return ( - Reasoning( - effort=reasoning_effort, - summary="detailed" if model is None else auto_reasoning_summary(model, custom_llm_provider), - ) - if auto_summary_enabled - else Reasoning(effort=reasoning_effort) + Reasoning(effort=reasoning_effort) + if summary is None + else Reasoning(effort=reasoning_effort, summary=summary) ) _map_reasoning_effort = map_reasoning_effort diff --git a/litellm/main.py b/litellm/main.py index 75bd4431120..41c4302499e 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -6124,7 +6124,18 @@ def completion( optional_params, rs_val = strip_reasoning_summary_aliases_from_optional_params(optional_params) - if isinstance(resolved_reasoning_effort, dict) and "summary" in resolved_reasoning_effort: + if custom_llm_provider == "openai" and isinstance(resolved_reasoning_effort, dict): + from litellm.responses.utils import openai_bridge_reasoning + + optional_params["reasoning_effort"] = openai_bridge_reasoning( + requested=resolved_reasoning_effort, + effort_kept="reasoning_effort" in optional_params, + summary_alias=rs_val, + model=model, + drop_params=drop_unsupported_reasoning, + additional_drop_params=routing_kwargs.get("additional_drop_params"), + ) + elif isinstance(resolved_reasoning_effort, dict) and "summary" in resolved_reasoning_effort: optional_params["reasoning_effort"] = resolved_reasoning_effort elif rs_val is not None: eff: Final = optional_params.get("reasoning_effort", resolved_reasoning_effort) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 7199a075612..825de5ac3ee 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -35598,6 +35598,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35675,6 +35682,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35851,6 +35865,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35902,6 +35923,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36017,6 +36045,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36066,6 +36101,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36107,6 +36149,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36194,6 +36243,15 @@ "output_cost_per_token_flex": 2.5e-05, "output_cost_per_token_priority": 0.0001, "output_cost_per_token_ultrafast": 0.0003, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36278,6 +36336,15 @@ "output_cost_per_token_above_272k_tokens_batches": 7.5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36370,6 +36437,15 @@ "output_cost_per_token_above_272k_tokens_batches": 3.75e-07, "output_cost_per_token_flex": 2.5e-07, "output_cost_per_token_priority": 1e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36453,6 +36529,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36534,6 +36619,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36617,6 +36711,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-06, "output_cost_per_token_flex": 6e-06, "output_cost_per_token_priority": 2.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36699,6 +36802,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-07, "output_cost_per_token_flex": 6e-07, "output_cost_per_token_priority": 2.4e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36936,6 +37048,14 @@ "reasoning_effort_levels": [ "medium" ], + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -36983,6 +37103,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37045,6 +37173,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37099,6 +37235,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37150,6 +37294,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37209,6 +37361,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37266,6 +37426,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37315,6 +37483,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37367,6 +37543,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37421,6 +37605,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37475,6 +37667,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37526,6 +37726,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37578,6 +37786,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37624,6 +37840,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37666,6 +37889,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37716,6 +37946,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37826,6 +38063,13 @@ "mode": "responses", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37908,6 +38152,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37959,6 +38210,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38007,6 +38265,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38055,6 +38320,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43821,6 +44093,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43929,6 +44208,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44008,6 +44294,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44045,6 +44338,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -73288,6 +73588,15 @@ "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "output_cost_per_token_ultrafast": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -76787,6 +77096,15 @@ "output_cost_per_token_batches": 2.5e-06, "output_cost_per_token_flex": 2.5e-06, "output_cost_per_token_priority": 1e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": {"search_context_size_high":0.01,"search_context_size_low":0.01,"search_context_size_medium":0.01}, diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py index 02d07932533..cf993c69814 100644 --- a/litellm/responses/utils.py +++ b/litellm/responses/utils.py @@ -1,8 +1,10 @@ import base64 +import os import re from collections.abc import Callable, Collection, Iterable, Mapping, Sequence from functools import reduce -from typing import Any, Final, Optional, TypeVar, Union, cast, get_type_hints, overload +from types import MappingProxyType +from typing import Any, Final, Optional, TypeVar, Union, cast, get_args, get_type_hints, overload from pydantic import BaseModel, TypeAdapter from typing_extensions import TypedDict as ExtraItemsTypedDict @@ -14,6 +16,7 @@ from litellm.litellm_core_utils.core_helpers import as_str_mapping, attribute_or from litellm.litellm_core_utils.dot_notation_indexing import delete_nested_value, is_nested_path from litellm.litellm_core_utils.get_provider_specific_headers import ProviderSpecificHeaderUtils from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig, ReasoningSummaryLevel +from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig from litellm.types.llms.openai import ( AllMessageValues, InputTokensDetails, @@ -29,6 +32,7 @@ from litellm.types.responses.main import DecodedResponseId from litellm.types.utils import ( CacheCreationTokenDetails, CompletionTokensDetailsWrapper, + ModelInfo, PromptTokensDetailsWrapper, ProviderSpecificHeader, SpecialEnums, @@ -1382,3 +1386,98 @@ def auto_reasoning_summary(model: str, custom_llm_provider: object) -> Reasoning provider, model.removeprefix(f"{provider}/") ) return "detailed" if responses_config is None else responses_config.auto_reasoning_summary() + + +def opt_in_auto_reasoning_summary(model: str | None, custom_llm_provider: object) -> ReasoningSummaryLevel | None: + if not (litellm.reasoning_auto_summary or os.getenv("LITELLM_REASONING_AUTO_SUMMARY", "false").lower() == "true"): + return None + return "detailed" if model is None else auto_reasoning_summary(model, custom_llm_provider) + + +def reasoning_with_summary_default( + reasoning: Mapping[str, object], summary_alias: object, model: str, custom_llm_provider: object +) -> Mapping[str, object]: + if "summary" in reasoning: + return {**reasoning} + summary: Final = ( + summary_alias if summary_alias is not None else opt_in_auto_reasoning_summary(model, custom_llm_provider) + ) + return {**reasoning} if summary is None else {**reasoning, "summary": summary} + + +_SUMMARY_VALUES: Final = frozenset((*get_args(ReasoningSummaryLevel), None)) + + +def _openai_model_info(model: str) -> ModelInfo | None: + from litellm.utils import ModelInfoLookupMiss, lookup_model_info_for_registration + + model_info: Final = lookup_model_info_for_registration(model.removeprefix("responses/"), "openai") + return None if isinstance(model_info, ModelInfoLookupMiss) else model_info + + +def _declared_values(values: Sequence[str] | None) -> frozenset[str | None] | None: + return None if values is None else frozenset(values) + + +def _openai_reasoning_field_values(model: str) -> Mapping[str, frozenset[str | None] | None]: + model_info: Final = _openai_model_info(model) + if model_info is None: + return MappingProxyType({"summary": _SUMMARY_VALUES, "mode": None, "context": None}) + if not model_info.get("supports_reasoning"): + return MappingProxyType({}) + return MappingProxyType( + { + "summary": _SUMMARY_VALUES, + "mode": _declared_values(model_info.get("reasoning_modes")), + "context": _declared_values(model_info.get("reasoning_contexts")), + } + ) + + +def _openai_accepts_reasoning_field( + key: str, value: object, values_by_field: Mapping[str, frozenset[str | None] | None] +) -> bool: + if key == "effort": + return True + if key not in values_by_field: + return False + allowed: Final = values_by_field[key] + return allowed is None or (isinstance(value, str | None) and value in allowed) + + +def _openai_accepted_reasoning(reasoning: Mapping[str, object], model: str) -> Mapping[str, object]: + values_by_field: Final = _openai_reasoning_field_values(model) + return { + key: value for key, value in reasoning.items() if _openai_accepts_reasoning_field(key, value, values_by_field) + } + + +def _additional_drop_paths(additional_drop_params: object) -> tuple[str, ...]: + if additional_drop_params.__class__ is not list: + return () + return tuple(path for path in iter_items(additional_drop_params) if isinstance(path, str)) + + +def openai_bridge_reasoning( + *, + requested: Mapping[str, object], + effort_kept: bool, + summary_alias: object, + model: str, + drop_params: bool, + additional_drop_params: object, +) -> Mapping[str, object] | None: + drop_paths: Final = _additional_drop_paths(additional_drop_params) + if "reasoning_effort" in drop_paths: + return None + kept: Final = requested if effort_kept else {key: value for key, value in requested.items() if key != "effort"} + with_summary: Final = reasoning_with_summary_default(kept, summary_alias, model, "openai") + accepted: Final = ( + _openai_accepted_reasoning(with_summary, model) + if drop_params and OpenAIGPTConfig.is_openai_catalog_model(model) + else {**with_summary} + ) + remaining: Final = as_str_mapping( + _apply_nested_drop_params({"reasoning_effort": accepted}, list(drop_paths)).get("reasoning_effort") + ) + return remaining or None diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 055c18f2e36..4ce35bf86c1 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -225,6 +225,8 @@ class ProviderSpecificModelInfo(TypedDict, total=False): supports_xhigh_reasoning_effort: bool | None supports_max_reasoning_effort: bool | None reasoning_effort_levels: ReadOnly[Sequence[str] | None] + reasoning_modes: ReadOnly[Sequence[str] | None] + reasoning_contexts: ReadOnly[Sequence[str] | None] default_reasoning_effort: ReadOnly[Literal["none", "minimal", "low", "medium", "high", "xhigh"] | None] supports_output_config: bool | None supports_image_size: bool | None diff --git a/litellm/utils.py b/litellm/utils.py index 61b0e17e197..a231644addf 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -7000,6 +7000,8 @@ def get_model_info_helper( supports_xhigh_reasoning_effort=_model_info.get("supports_xhigh_reasoning_effort", None), supports_max_reasoning_effort=_model_info.get("supports_max_reasoning_effort", None), reasoning_effort_levels=_model_info.get("reasoning_effort_levels", None), + reasoning_modes=_model_info.get("reasoning_modes", None), + reasoning_contexts=_model_info.get("reasoning_contexts", None), default_reasoning_effort=_model_info.get("default_reasoning_effort", None), bedrock_output_config_effort_ceiling=_model_info.get("bedrock_output_config_effort_ceiling", None), bedrock_converse_supports_strict_tools=_model_info.get("bedrock_converse_supports_strict_tools", None), diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 7199a075612..825de5ac3ee 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -35598,6 +35598,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35675,6 +35682,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35851,6 +35865,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35902,6 +35923,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36017,6 +36045,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36066,6 +36101,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36107,6 +36149,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36194,6 +36243,15 @@ "output_cost_per_token_flex": 2.5e-05, "output_cost_per_token_priority": 0.0001, "output_cost_per_token_ultrafast": 0.0003, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36278,6 +36336,15 @@ "output_cost_per_token_above_272k_tokens_batches": 7.5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36370,6 +36437,15 @@ "output_cost_per_token_above_272k_tokens_batches": 3.75e-07, "output_cost_per_token_flex": 2.5e-07, "output_cost_per_token_priority": 1e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36453,6 +36529,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36534,6 +36619,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36617,6 +36711,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-06, "output_cost_per_token_flex": 6e-06, "output_cost_per_token_priority": 2.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36699,6 +36802,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-07, "output_cost_per_token_flex": 6e-07, "output_cost_per_token_priority": 2.4e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36936,6 +37048,14 @@ "reasoning_effort_levels": [ "medium" ], + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -36983,6 +37103,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37045,6 +37173,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37099,6 +37235,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37150,6 +37294,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37209,6 +37361,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37266,6 +37426,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37315,6 +37483,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37367,6 +37543,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37421,6 +37605,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37475,6 +37667,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37526,6 +37726,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37578,6 +37786,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37624,6 +37840,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37666,6 +37889,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37716,6 +37946,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37826,6 +38063,13 @@ "mode": "responses", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37908,6 +38152,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37959,6 +38210,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38007,6 +38265,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38055,6 +38320,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43821,6 +44093,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43929,6 +44208,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44008,6 +44294,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44045,6 +44338,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -73288,6 +73588,15 @@ "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "output_cost_per_token_ultrafast": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -76787,6 +77096,15 @@ "output_cost_per_token_batches": 2.5e-06, "output_cost_per_token_flex": 2.5e-06, "output_cost_per_token_priority": 1e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": {"search_context_size_high":0.01,"search_context_size_low":0.01,"search_context_size_medium":0.01}, diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py index ba39fc3f5ac..f38bbcf4404 100644 --- a/tests/unit/test_main.py +++ b/tests/unit/test_main.py @@ -2127,6 +2127,330 @@ def test_responses_bridge_preserves_reasoning_effort_with_drop_params( assert request_body["reasoning"] == {"effort": "high"} +@pytest.fixture +def openai_responses_route( + restore_model_registry, respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> respx.Route: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.delenv("LITELLM_REASONING_AUTO_SUMMARY", raising=False) + litellm.register_model( + { + "gpt-6.1-terra": { + "litellm_provider": "openai", + "mode": "chat", + "supports_reasoning": True, + "supports_minimal_reasoning_effort": False, + "reasoning_modes": ["standard", "pro"], + "reasoning_contexts": ["auto", "current_turn", "all_turns"], + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + }, + persist_across_reloads=False, + ) + return respx_mock.post("https://api.openai.com/v1/responses").respond( + json={ + "id": "resp_test", + "object": "response", + "created_at": 1734366691, + "status": "completed", + "model": "gpt-6.1-terra", + "output": [ + { + "type": "message", + "id": "msg_1", + "status": "completed", + "role": "assistant", + "content": [{"type": "output_text", "text": "Done.", "annotations": []}], + } + ], + "usage": {"input_tokens": 1, "output_tokens": 1, "total_tokens": 2}, + } + ) + + +@pytest.mark.parametrize( + ("reasoning_effort", "summary_alias", "expected_reasoning"), + [ + ( + {"effort": "low", "mode": "pro", "context": "current_turn"}, + {}, + {"effort": "low", "mode": "pro", "context": "current_turn"}, + ), + ( + {"effort": "low", "mode": "pro", "context": "current_turn"}, + {"reasoning_summary": "auto"}, + {"effort": "low", "mode": "pro", "context": "current_turn", "summary": "auto"}, + ), + ( + {"effort": "low", "mode": "pro", "summary": "concise"}, + {"reasoningSummary": "auto"}, + {"effort": "low", "mode": "pro", "summary": "concise"}, + ), + ( + {"effort": "low", "context": "current_turn", "summary": None}, + {"reasoning_summary": "auto"}, + {"effort": "low", "context": "current_turn", "summary": None}, + ), + ({"mode": "pro", "context": "all_turns"}, {}, {"mode": "pro", "context": "all_turns"}), + ("low", {}, {"effort": "low"}), + ], +) +def test_openai_responses_bridge_sends_the_whole_reasoning_effort_dict( + reasoning_effort: str | dict[str, str | None], + summary_alias: dict[str, str], + expected_reasoning: dict[str, str | None], + openai_responses_route: respx.Route, +): + caller_reasoning_effort: Final = copy.deepcopy(reasoning_effort) + + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=reasoning_effort, + api_key="fake-key", + **summary_alias, + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + assert reasoning_effort == caller_reasoning_effort + + +@pytest.mark.parametrize( + "reasoning_effort, summary_alias, expected_reasoning", + [ + ( + {"effort": "minimal", "mode": "pro", "context": "current_turn"}, + {}, + {"mode": "pro", "context": "current_turn"}, + ), + ({"effort": "minimal", "mode": "pro", "summary": "auto"}, {}, {"mode": "pro", "summary": "auto"}), + ({"effort": "minimal", "mode": "pro"}, {"reasoning_summary": "auto"}, {"mode": "pro", "summary": "auto"}), + ], +) +def test_openai_responses_bridge_drops_only_a_refused_effort_under_drop_params( + reasoning_effort: dict[str, str], + summary_alias: dict[str, str], + expected_reasoning: dict[str, str], + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=reasoning_effort, + drop_params=True, + api_key="fake-key", + **summary_alias, + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + + +def test_openai_responses_bridge_drops_a_reasoning_effort_dict_for_a_model_without_reasoning( + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-4.1", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "current_turn"}, + drop_params=True, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert "reasoning" not in request_body + + +@pytest.mark.parametrize( + "drop_params, expected_reasoning", + [ + (True, {"effort": "low", "context": "current_turn"}), + ( + False, + { + "effort": "low", + "exclude": False, + "mode": "turbo", + "context": "current_turn", + "generate_summary": "concise", + }, + ), + ], +) +def test_openai_responses_bridge_drops_reasoning_fields_openai_refuses_only_under_drop_params( + drop_params: bool, + expected_reasoning: dict[str, object], + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={ + "effort": "low", + "exclude": False, + "mode": "turbo", + "context": "current_turn", + "generate_summary": "concise", + }, + drop_params=drop_params, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + + +@pytest.mark.parametrize( + "declared, drop_params, expected_reasoning", + [ + ( + {"reasoning_modes": ["standard"], "reasoning_contexts": ["auto", "current_turn"]}, + True, + {"effort": "low"}, + ), + ( + {"reasoning_modes": ["pro"], "reasoning_contexts": ["all_turns"]}, + True, + {"effort": "low", "mode": "pro", "context": "all_turns"}, + ), + ({}, True, {"effort": "low", "mode": "pro", "context": "all_turns"}), + ({"reasoning_modes": ["standard"]}, True, {"effort": "low", "context": "all_turns"}), + ( + {"reasoning_modes": ["standard"], "reasoning_contexts": ["auto", "current_turn"]}, + False, + {"effort": "low", "mode": "pro", "context": "all_turns"}, + ), + ], +) +def test_openai_responses_bridge_drops_a_mode_or_context_the_model_does_not_list_only_under_drop_params( + declared: dict[str, list[str]], + drop_params: bool, + expected_reasoning: dict[str, str], + openai_responses_route: respx.Route, +): + litellm.register_model( + { + "gpt-6.1-reasoning-test": { + "litellm_provider": "openai", + "mode": "chat", + "supports_reasoning": True, + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + **declared, + } + }, + persist_across_reloads=False, + ) + + litellm.completion( + model="openai/responses/gpt-6.1-reasoning-test", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns"}, + drop_params=drop_params, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + + +def test_openai_responses_bridge_forwards_mode_and_context_under_drop_params_for_a_catalog_model_with_no_cost_entry( + openai_responses_route: respx.Route, +): + litellm.open_ai_chat_completion_models.add("unpriced-reasoning-model") + + litellm.completion( + model="openai/responses/unpriced-reasoning-model", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns", "generate_summary": "auto"}, + drop_params=True, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == {"effort": "low", "mode": "pro", "context": "all_turns"} + + +def test_openai_responses_bridge_forwards_a_reasoning_effort_dict_whole_for_a_model_outside_the_openai_catalog( + openai_responses_route: respx.Route, +): + reasoning_effort: Final = {"effort": "high", "exclude": True, "budget_tokens": 2000} + + litellm.completion( + model="openai/responses/my-proxy-alias", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=reasoning_effort, + drop_params=True, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == reasoning_effort + + +@pytest.mark.parametrize( + "additional_drop_params, summary_alias, expected_reasoning", + [ + (["reasoning_effort.mode"], {}, {"effort": "low", "context": "current_turn"}), + ( + ["reasoning_effort.summary"], + {"reasoning_summary": "auto"}, + {"effort": "low", "mode": "pro", "context": "current_turn"}, + ), + (["reasoning_effort"], {"reasoning_summary": "auto"}, None), + ("reasoning_effort", {}, {"effort": "low", "mode": "pro", "context": "current_turn"}), + ], +) +def test_openai_responses_bridge_applies_additional_drop_params_to_a_reasoning_effort_dict( + additional_drop_params: list[str] | str, + summary_alias: dict[str, str], + expected_reasoning: dict[str, str] | None, + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "current_turn"}, + additional_drop_params=additional_drop_params, + api_key="fake-key", + **summary_alias, + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body.get("reasoning") == expected_reasoning + + +def test_openai_responses_bridge_gives_a_reasoning_effort_dict_the_auto_summary_a_string_gets( + openai_responses_route: respx.Route, monkeypatch: pytest.MonkeyPatch +): + monkeypatch.setenv("LITELLM_REASONING_AUTO_SUMMARY", "true") + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort="low", + api_key="fake-key", + ) + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "current_turn"}, + api_key="fake-key", + ) + + string_reasoning, dict_reasoning = ( + json.loads(call.request.content)["reasoning"] for call in openai_responses_route.calls + ) + assert string_reasoning["summary"] is not None + assert dict_reasoning == { + "effort": "low", + "mode": "pro", + "context": "current_turn", + "summary": string_reasoning["summary"], + } + + _BRIDGED_RESPONSE_ANSWER: Final = json.dumps( { "id": "resp_slow", diff --git a/tests/unit/test_utils.py b/tests/unit/test_utils.py index 38287b25169..146e2df6413 100644 --- a/tests/unit/test_utils.py +++ b/tests/unit/test_utils.py @@ -1195,6 +1195,8 @@ _COST_MAP_ROW_SCHEMA: Final = { "type": "array", "items": {"type": "string", "enum": ["none", "minimal", "low", "medium", "high", "xhigh", "max"]}, }, + "reasoning_modes": {"type": "array", "items": {"type": "string"}}, + "reasoning_contexts": {"type": "array", "items": {"type": "string"}}, "default_reasoning_effort": { "type": "string", "enum": ["none", "minimal", "low", "medium", "high", "xhigh"],