diff --git a/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py b/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py index d66631e138c..76544e904d2 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py +++ b/litellm/llms/anthropic/messages/responses_adapters/streaming_iterator.py @@ -3,7 +3,7 @@ import asyncio from collections import deque from collections.abc import AsyncGenerator, Iterator, Mapping -from typing import TYPE_CHECKING, Any, Final, cast +from typing import TYPE_CHECKING, Any, Final from pydantic import ConfigDict, field_validator @@ -139,6 +139,7 @@ class AnthropicResponsesStreamWrapper: self._current_block_index: int = -1 # Map item_id -> content_block_index so we can stop the right block later self._item_id_to_block_index: dict[str, int] = {} + self._reasoning_text_block_indexes: frozenset[int] = frozenset() # Track open function_call items by item_id so we can emit tool_use start self._pending_tool_ids: dict[str, str] = {} # item_id -> call_id / name accumulator self._sent_message_start = False @@ -189,23 +190,13 @@ class AnthropicResponsesStreamWrapper: def _field(source: object, name: str) -> object: return source.get(name) if isinstance(source, dict) else getattr(source, name, None) - @classmethod - def _has_reasoning_text(cls, item: object) -> bool: - content: Final = cls._field(item, "content") - if not isinstance(content, list): - return False - return any( - cls._field(part, "type") == "reasoning_text" and cls._field(part, "text") - for part in cast(list[object], content) # cast-ok: narrowed by isinstance - ) - def _close_reasoning_item(self, item: object, item_id: str | None) -> None: block_idx: Final = self._item_id_to_block_index.get(item_id, -1) if item_id else self._current_block_index encrypted_content: Final = self._field(item, "encrypted_content") encrypted: Final = encrypted_content if isinstance(encrypted_content, str) and encrypted_content else None signature: Final = ( reasoning_text_signature(encrypted) - if block_idx >= 0 and self._has_reasoning_text(item) + if block_idx in self._reasoning_text_block_indexes else encrypted_reasoning_signature(encrypted) if encrypted is not None else None @@ -343,6 +334,8 @@ class AnthropicResponsesStreamWrapper: item_id, {"type": "thinking", "thinking": "", "signature": ""}, ) + if event_type == "response.reasoning_text.delta": + self._reasoning_text_block_indexes = self._reasoning_text_block_indexes | {block_idx} self._chunk_queue.append( { "type": "content_block_delta", diff --git a/litellm/llms/anthropic/messages/responses_adapters/transformation.py b/litellm/llms/anthropic/messages/responses_adapters/transformation.py index 4e028479aa2..4e63729d294 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/messages/responses_adapters/transformation.py @@ -14,8 +14,8 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( TOOL_RESULT_IMAGE_BOUNDARY, TOOL_RESULT_IMAGE_PLACEHOLDER, encrypted_reasoning_signature, - responses_reasoning_items_from_thinking_blocks, reasoning_text_signature, + responses_reasoning_items_from_thinking_blocks, with_prompt_cache_breakpoint, ) from litellm.litellm_core_utils.reasoning_effort_utils import ( @@ -247,14 +247,12 @@ class LiteLLMAnthropicToResponsesAPIAdapter: return str(mapping.get("text") or "") return str(getattr(part, "text", None) or "") - @classmethod - def _reasoning_text_part_text(cls, part: object) -> str: - part_type: Final = ( - cast(Mapping[str, object], part).get("type") # cast-ok: content parts are untyped provider json - if isinstance(part, Mapping) - else getattr(part, "type", None) - ) - return cls._summary_part_text(part) if part_type == "reasoning_text" else "" + @staticmethod + def _reasoning_text_part_text(part: object) -> str: + if isinstance(part, Mapping): + mapping: Final = cast(Mapping[str, object], part) # cast-ok: content parts are untyped provider json + return str(mapping.get("text") or "") if mapping.get("type") == "reasoning_text" else "" + return str(getattr(part, "text", None) or "") if getattr(part, "type", None) == "reasoning_text" else "" @classmethod def _thinking_block_from_reasoning_item( @@ -271,7 +269,7 @@ class LiteLLMAnthropicToResponsesAPIAdapter: since only Anthropic can sign a thinking block. """ reasoning_text: Final = REASONING_SUMMARY_PART_SEPARATOR.join( - part_text for part in content if (part_text := cls._reasoning_text_part_text(part)) + reasoning_part_text for part in content if (reasoning_part_text := cls._reasoning_text_part_text(part)) ) if reasoning_text: return AnthropicResponseContentBlockThinking( diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 166072e9613..57e213a834e 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -48035,6 +48035,13 @@ "ChatCompletionReasoningItem": { "description": "Represents an OpenAI Responses API reasoning item for round-tripping in conversation history.", "properties": { + "content": { + "items": { + "$ref": "#/components/schemas/ChatCompletionReasoningTextBlock" + }, + "title": "Content", + "type": "array" + }, "encrypted_content": { "anyOf": [ { @@ -48087,6 +48094,25 @@ "title": "ChatCompletionReasoningSummaryTextBlock", "type": "object" }, + "ChatCompletionReasoningTextBlock": { + "properties": { + "text": { + "title": "Text", + "type": "string" + }, + "type": { + "const": "reasoning_text", + "title": "Type", + "type": "string" + } + }, + "required": [ + "type", + "text" + ], + "title": "ChatCompletionReasoningTextBlock", + "type": "object" + }, "ChatCompletionRedactedThinkingBlock": { "properties": { "cache_control": { diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index 42e0ca586c2..da224842539 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -1,5 +1,5 @@ import builtins -from collections.abc import Iterable, Mapping +from collections.abc import Iterable, Mapping, Sequence from enum import Enum from os import PathLike from typing import IO, TYPE_CHECKING, Any, Final, Generic, Literal, Optional, TypeAlias, Union @@ -723,7 +723,7 @@ class ChatCompletionReasoningItem(TypedDict, total=False): id: str encrypted_content: str | None summary: ReadOnly[list[ChatCompletionReasoningSummaryTextBlock]] - content: ReadOnly[list[ChatCompletionReasoningTextBlock]] + content: ReadOnly[Sequence[ChatCompletionReasoningTextBlock]] class WebSearchOptionsUserLocationApproximate(TypedDict, total=False): diff --git a/tests/integration/messages_endpoint/responses_bridge/test_anthropic_messages_responses_reasoning_replay_wire.py b/tests/integration/messages_endpoint/responses_bridge/test_anthropic_messages_responses_reasoning_replay_wire.py index 68ac00de14a..975850929f4 100644 --- a/tests/integration/messages_endpoint/responses_bridge/test_anthropic_messages_responses_reasoning_replay_wire.py +++ b/tests/integration/messages_endpoint/responses_bridge/test_anthropic_messages_responses_reasoning_replay_wire.py @@ -174,7 +174,11 @@ def test_messages_replays_reasoning_text_with_its_encrypted_content(gateway: Gat "/v1/messages", _messages_request( model, - [user_turn, {"role": "assistant", "content": assistant_content}, _tool_results(identity, {"a": "alpha"})], + [ + user_turn, + {"role": "assistant", "content": assistant_content}, + _tool_results(identity, {"a": "alpha"}), + ], ), ) assert second.status_code == 200, second.text @@ -196,7 +200,7 @@ def test_messages_stream_emits_reasoning_text_deltas_as_thinking(gateway: Gatewa pieces: Final = ("Let me ", "think ", identity) reasoning_item: Final = _reasoning(identity, "".join(pieces)) completed: Final = _response(identity, (reasoning_item, _message(identity, "done"))) - events: Final = ( + events: Final[tuple[Mapping[str, object], ...]] = ( {"type": "response.created", "response": {**completed, "status": "in_progress", "output": []}}, {"type": "response.output_item.added", "output_index": 0, "item": {**reasoning_item, "content": []}}, *( @@ -323,11 +327,26 @@ def test_messages_replays_assistant_text_in_its_original_place_among_tool_calls( "role": "assistant", "content": [ {"type": "text", "text": "First I will look."}, - {"type": "tool_use", "id": f"call_{identity}_a", "name": "Bash", "input": {"command": "ls"}}, + { + "type": "tool_use", + "id": f"call_{identity}_a", + "name": "Bash", + "input": {"command": "ls"}, + }, {"type": "text", "text": "Then two more."}, {"type": "text", "text": "In parallel."}, - {"type": "tool_use", "id": f"call_{identity}_b", "name": "Bash", "input": {"command": "pwd"}}, - {"type": "tool_use", "id": f"call_{identity}_c", "name": "Bash", "input": {"command": "id"}}, + { + "type": "tool_use", + "id": f"call_{identity}_b", + "name": "Bash", + "input": {"command": "pwd"}, + }, + { + "type": "tool_use", + "id": f"call_{identity}_c", + "name": "Bash", + "input": {"command": "id"}, + }, ], }, _tool_results(identity, {"a": "a.txt", "b": "/tmp", "c": "uid=0"}), diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index d03f4b4cffd..c55c46e1dd1 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -28897,6 +28897,8 @@ export interface components { * @description Represents an OpenAI Responses API reasoning item for round-tripping in conversation history. */ ChatCompletionReasoningItem: { + /** Content */ + content?: components["schemas"]["ChatCompletionReasoningTextBlock"][]; /** Encrypted Content */ encrypted_content?: string | null; /** Id */ @@ -28923,6 +28925,18 @@ export interface components { } & { [key: string]: unknown; }; + /** ChatCompletionReasoningTextBlock */ + ChatCompletionReasoningTextBlock: { + /** Text */ + text: string; + /** + * Type + * @constant + */ + type: "reasoning_text"; + } & { + [key: string]: unknown; + }; /** ChatCompletionRedactedThinkingBlock */ ChatCompletionRedactedThinkingBlock: { /** Cache Control */