From f49ebc93eaf108d2312779346485ce6808c26184 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 9 Sep 2026 14:11:43 -0700 Subject: [PATCH] fix(anthropic): request encrypted reasoning only where the Responses provider returns it The bridge now asks for reasoning.encrypted_content whenever the provider's Responses config lists include, independent of the client's thinking block, and leaves it out for providers such as Perplexity that reject the param. Bridge-tagged blocks are stripped on the chat adapter path too, so a mid session model switch to Gemini or Bedrock no longer forwards them as real signatures, a bare prefix counts as bridge-tagged, and non-mapping messages pass through the strip untouched. --- .../prompt_templates/common_utils.py | 18 ++++++--- litellm/llms/anthropic/common_utils.py | 2 + .../adapters/transformation.py | 4 +- .../responses_adapters/handler.py | 17 ++++++++- .../responses_adapters/transformation.py | 9 ++++- ...ore_utils_prompt_templates_common_utils.py | 3 ++ ...al_pass_through_adapters_transformation.py | 38 +++++++++++++++++++ .../test_responses_adapters_handler.py | 23 +++++++++++ .../test_responses_adapters_transformation.py | 12 +++++- .../anthropic/test_anthropic_common_utils.py | 9 +++++ 10 files changed, 126 insertions(+), 9 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index 918cb982773..75a3aefa98b 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -1853,22 +1853,30 @@ def encrypted_reasoning_signature(encrypted_content: str) -> str: return f"{ENCRYPTED_REASONING_SIGNATURE_PREFIX}{encrypted_content}" +def _carries_encrypted_reasoning(signature: object) -> bool: + return isinstance(signature, str) and signature.startswith(ENCRYPTED_REASONING_SIGNATURE_PREFIX) + + def encrypted_content_from_signature(signature: object) -> str | None: - if not isinstance(signature, str) or not signature.startswith(ENCRYPTED_REASONING_SIGNATURE_PREFIX): + if not isinstance(signature, str) or not _carries_encrypted_reasoning(signature): return None return signature.removeprefix(ENCRYPTED_REASONING_SIGNATURE_PREFIX) or None -def _encrypted_content_of_block(block: Mapping[str, object]) -> str | None: +def _encrypted_reasoning_field(block: Mapping[str, object]) -> object: match block.get("type"): case "thinking": - return encrypted_content_from_signature(block.get("signature")) + return block.get("signature") case "redacted_thinking": - return encrypted_content_from_signature(block.get("data")) + return block.get("data") case _: return None +def _encrypted_content_of_block(block: Mapping[str, object]) -> str | None: + return encrypted_content_from_signature(_encrypted_reasoning_field(block)) + + def is_encrypted_reasoning_block(block: object) -> bool: """A thinking or redacted_thinking block carrying Responses API encrypted reasoning. @@ -1878,7 +1886,7 @@ def is_encrypted_reasoning_block(block: object) -> bool: if not isinstance(block, Mapping): return False mapping: Final = cast(Mapping[str, object], block) # cast-ok: narrowed by isinstance - return _encrypted_content_of_block(mapping) is not None + return _carries_encrypted_reasoning(_encrypted_reasoning_field(mapping)) def _reasoning_replay_group_key(indexed_block: tuple[int, Mapping[str, object]]) -> str: diff --git a/litellm/llms/anthropic/common_utils.py b/litellm/llms/anthropic/common_utils.py index cb24f4153f9..ca520116606 100644 --- a/litellm/llms/anthropic/common_utils.py +++ b/litellm/llms/anthropic/common_utils.py @@ -1203,6 +1203,8 @@ def strip_thinking_blocks_from_anthropic_messages(messages: list[Any]) -> list[A def _without_encrypted_reasoning_blocks(message: dict) -> dict | None: # mutable-ok: Anthropic message payload shape + if not isinstance(message, Mapping): + return message content: Final = message.get("content") if not isinstance(content, list): return message diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index db890662132..aaffc84b872 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -113,6 +113,7 @@ from litellm.litellm_core_utils.reasoning_effort_utils import ( from litellm.llms.anthropic.common_utils import ( is_empty_unsigned_thinking_block, normalize_anthropic_tool_use_id, + strip_encrypted_reasoning_blocks_from_anthropic_messages, ) from litellm.llms.anthropic.experimental_pass_through.context_management import ( PolyfillResult, @@ -417,7 +418,8 @@ class LiteLLMAnthropicMessagesAdapter: model: str | None = None, ) -> list: new_messages: Final[list[AllMessageValues]] = [] - for m in messages: + replayable_messages: Final = strip_encrypted_reasoning_blocks_from_anthropic_messages(messages) + for m in replayable_messages: user_message: ChatCompletionUserMessage | None = None tool_message_list: list[ChatCompletionToolMessage] = [] new_user_content_list: list[ChatCompletionTextObject | ChatCompletionImageObject] = [] diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py index 0445c23ed8c..24825cc87e0 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py @@ -19,6 +19,7 @@ from litellm.types.llms.anthropic_messages.anthropic_response import ( AnthropicMessagesResponse, ) from litellm.types.llms.openai import ResponsesAPIResponse +from litellm.utils import ProviderConfigManager from ..utils import litellm_logging_obj_from_kwargs, local_model_name from .streaming_iterator import AnthropicResponsesStreamWrapper @@ -34,6 +35,15 @@ def _forwarded_kwargs(extra_kwargs: Mapping[str, object] | None) -> Mapping[str, return extra_kwargs or {} +def _provider_returns_encrypted_reasoning(model: str, custom_llm_provider: object) -> bool: + provider: Final = ( + custom_llm_provider if isinstance(custom_llm_provider, str) else litellm.get_llm_provider(model=model)[1] + ) + provider_model: Final = local_model_name(model, provider) + responses_config: Final = ProviderConfigManager.get_provider_responses_api_config(provider, provider_model) + return responses_config is not None and "include" in responses_config.get_supported_openai_params(provider_model) + + def _build_responses_kwargs( *, max_tokens: int, @@ -85,8 +95,13 @@ def _build_responses_kwargs( request_data["output_format"] = output_format anthropic_request: Final = AnthropicMessagesRequest(**request_data) - responses_kwargs: Final = _ADAPTER.translate_request(anthropic_request) forwarded_kwargs: Final = _forwarded_kwargs(extra_kwargs) + responses_kwargs: Final = _ADAPTER.translate_request( + anthropic_request, + include_encrypted_reasoning=_provider_returns_encrypted_reasoning( + model, forwarded_kwargs.get("custom_llm_provider") + ), + ) # Normalize reasoning effort based on model capabilities # (e.g. "max" → "xhigh"/"high", "minimal" → "low" if unsupported) diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py index 07419db443c..1fdb0318bab 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py @@ -505,10 +505,16 @@ class LiteLLMAnthropicToResponsesAPIAdapter: def translate_request( self, anthropic_request: AnthropicMessagesRequest, + include_encrypted_reasoning: bool = True, ) -> dict[str, Any]: """ Translate a full Anthropic /v1/messages request dict to litellm.responses() / litellm.aresponses() kwargs. + + ``include_encrypted_reasoning`` asks the provider for ``reasoning.encrypted_content`` + on every call, so a reasoning model's items can be replayed intact next turn even + when the client sent no ``thinking`` block; pass False for a provider whose + Responses API rejects ``include``. """ model: Final[str] = anthropic_request["model"] messages_list: Final = cast( @@ -538,6 +544,8 @@ class LiteLLMAnthropicToResponsesAPIAdapter: "model": model, "input": input_items, } + if include_encrypted_reasoning: + responses_kwargs["include"] = [RESPONSES_INCLUDE_ENCRYPTED_REASONING] # mutable-ok: API request payload if system and not developer_parts: if isinstance(system, str): @@ -582,7 +590,6 @@ class LiteLLMAnthropicToResponsesAPIAdapter: ) if reasoning: responses_kwargs["reasoning"] = reasoning - responses_kwargs["include"] = [RESPONSES_INCLUDE_ENCRYPTED_REASONING] # mutable-ok: json list # output_format / output_config.format -> text format # output_format: {"type": "json_schema", "schema": {...}} diff --git a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py index 9df124576a0..ecb07d8c937 100644 --- a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py +++ b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py @@ -7,6 +7,7 @@ import pytest from litellm.litellm_core_utils.prompt_templates.common_utils import ( + ENCRYPTED_REASONING_SIGNATURE_PREFIX, TOOL_RESULT_IMAGE_BOUNDARY, TOOL_RESULT_IMAGE_PLACEHOLDER, add_system_prompt_to_messages, @@ -1610,6 +1611,8 @@ class TestEncryptedReasoningReplay: [ ({"type": "thinking", "thinking": "x", "signature": encrypted_reasoning_signature("g")}, True), ({"type": "redacted_thinking", "data": encrypted_reasoning_signature("g")}, True), + ({"type": "thinking", "thinking": "x", "signature": ENCRYPTED_REASONING_SIGNATURE_PREFIX}, True), + ({"type": "redacted_thinking", "data": ENCRYPTED_REASONING_SIGNATURE_PREFIX}, True), ({"type": "thinking", "thinking": "x", "signature": "ErcBCkgIValid"}, False), ({"type": "redacted_thinking", "data": "EmwKAhgBEgy"}, False), ({"type": "text", "text": encrypted_reasoning_signature("g")}, False), diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py index 30465ca25ba..279ed7a8b33 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py @@ -9,6 +9,7 @@ import litellm from litellm.litellm_core_utils.prompt_templates.common_utils import ( TOOL_RESULT_IMAGE_PLACEHOLDER, + encrypted_reasoning_signature, ) from litellm.litellm_core_utils.prompt_templates.factory import ( THOUGHT_SIGNATURE_SEPARATOR, @@ -407,6 +408,43 @@ def test_translate_anthropic_messages_to_openai_thinking_blocks(): assert result[1]["tool_calls"][0]["id"] == "toolu_01234" +def test_translate_anthropic_messages_to_openai_drops_bridge_encrypted_reasoning_blocks(): + """A session that moves from an OpenAI reasoning model to a chat provider replays reasoning only OpenAI can read. + + Gemini rejects the whole request when such a block reaches it as a thought_signature, so the + adapter drops those blocks and keeps the provider-signed ones. + """ + + anthropic_messages = [ + AnthropicMessagesUserMessageParam( + role="user", + content=[{"type": "text", "text": "Who drinks water?"}], + ), + AnthopicMessagesAssistantMessageParam( + role="assistant", + content=[ + {"type": "thinking", "thinking": "plan", "signature": encrypted_reasoning_signature("gAAAA_1")}, + {"type": "redacted_thinking", "data": encrypted_reasoning_signature("gAAAA_2")}, + {"type": "text", "text": "The Norwegian."}, + ], + ), + AnthopicMessagesAssistantMessageParam( + role="assistant", + content=[ + {"type": "thinking", "thinking": "native", "signature": "EqQBCkYIAxgCIkA_signed"}, + {"type": "text", "text": "Still the Norwegian."}, + ], + ), + ] + + result = LiteLLMAnthropicMessagesAdapter().translate_anthropic_messages_to_openai(messages=anthropic_messages) + + assert [m["role"] for m in result] == ["user", "assistant", "assistant"] + assert not result[1].get("thinking_blocks") + assert result[1]["content"] == "The Norwegian." + assert [b["signature"] for b in result[2]["thinking_blocks"]] == ["EqQBCkYIAxgCIkA_signed"] + + def test_translate_anthropic_messages_to_openai_sets_reasoning_content(): """Reasoning-aware chat providers read reasoning_content, so thinking text must land there. diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py index 3383813245a..2cbfaa17d23 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py @@ -54,6 +54,29 @@ def test_build_responses_kwargs_prefers_explicit_prompt_cache_key_over_derived() assert responses_kwargs["prompt_cache_key"] == "explicit-key" +def test_build_responses_kwargs_asks_openai_for_encrypted_reasoning_without_thinking(): + responses_kwargs = _build_responses_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="openai/gpt-5.6-luna", + extra_kwargs={"custom_llm_provider": "openai"}, + ) + assert responses_kwargs["include"] == ["reasoning.encrypted_content"] + assert "reasoning" not in responses_kwargs + + +def test_build_responses_kwargs_skips_include_for_a_responses_provider_that_rejects_it(): + responses_kwargs = _build_responses_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="perplexity/sonar", + thinking={"type": "enabled", "budget_tokens": 4096}, + extra_kwargs={"custom_llm_provider": "perplexity"}, + ) + assert "include" not in responses_kwargs + assert "reasoning" in responses_kwargs + + def test_build_responses_kwargs_without_metadata_sets_no_prompt_cache_key(): responses_kwargs = _build_responses_kwargs( max_tokens=1024, diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py index 79eaf6b60f2..9b7b15ace61 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py @@ -1162,7 +1162,6 @@ class TestTranslateRequestBroaderCoverage: req = _make_request(thinking={"type": "disabled"}) kwargs = _ADAPTER.translate_request(req) assert "reasoning" not in kwargs - assert "include" not in kwargs def test_thinking_asks_for_the_encrypted_reasoning(self): """The documented way to get reasoning that survives store=false is to ask for it.""" @@ -1170,6 +1169,17 @@ class TestTranslateRequestBroaderCoverage: kwargs = _ADAPTER.translate_request(req) assert kwargs["include"] == ["reasoning.encrypted_content"] + def test_encrypted_reasoning_is_asked_for_without_a_thinking_block(self): + """A reasoning model reasons whether or not the client sent `thinking`, so the replay needs it either way.""" + kwargs = _ADAPTER.translate_request(_make_request()) + assert kwargs["include"] == ["reasoning.encrypted_content"] + + def test_encrypted_reasoning_is_not_asked_for_when_the_provider_rejects_include(self): + req = _make_request(thinking={"type": "enabled", "budget_tokens": 12000}) + kwargs = _ADAPTER.translate_request(req, include_encrypted_reasoning=False) + assert kwargs["reasoning"] == {"effort": "high"} + assert "include" not in kwargs + def test_metadata_user_id_mapped_to_user(self): req = _make_request(metadata={"user_id": "user-42"}) kwargs = _ADAPTER.translate_request(req) diff --git a/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py b/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py index 97d22274b1c..91652c89092 100644 --- a/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py +++ b/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py @@ -1600,6 +1600,15 @@ class TestAnthropicThinkingSignatureSelfHeal: assert len(msgs[1]["content"]) == 2 assert len(msgs[2]["content"]) == 4 + def test_strip_encrypted_reasoning_leaves_malformed_messages_for_the_provider_to_reject(self): + """A bare string in messages must reach Anthropic as a 400, not die in the stripper as a 500.""" + from litellm.llms.anthropic.common_utils import ( + strip_encrypted_reasoning_blocks_from_anthropic_messages, + ) + + msgs = ["hi", {"role": "user", "content": "hello"}] + assert strip_encrypted_reasoning_blocks_from_anthropic_messages(msgs) == msgs + def test_strip_empty_text_blocks_treats_null_text_as_empty(self): from litellm.llms.anthropic.common_utils import ( strip_empty_content_blocks_from_anthropic_messages,