From cbe340a31ca81c40be088520ebbaaaca644977ff Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 8 Sep 2026 11:38:18 -0700 Subject: [PATCH] feat(guardrails): deliver tool-call rewrites into buffered streams A post_call pipeline guardrail that rewrites a streamed tool call (its arguments or its name) now has that rewrite written back across the buffered chunks on chat, Responses, and Messages streams, so the client receives the rewritten tool call instead of the original. The chat handler rewrites the first fragment of each tool-call index and blanks the rest, the Responses handler syncs the function_call output items and their argument events, and the Messages handler rewrites the tool_use content_block_start and input_json_delta events in both dict and SSE-bytes chunks. The delivers_ended_stream_text_rewrites flag becomes delivers_ended_stream_rewrites, since the write-back now covers both text and tool calls, and the executor only discards a tool-call rewrite on translations without write-back or on a shape the translation refuses. --- .../chat/guardrail_translation/handler.py | 176 +++++++++++++++--- .../guardrail_translation/base_translation.py | 17 +- .../chat/guardrail_translation/handler.py | 96 +++++++++- .../guardrail_translation/handler.py | 111 ++++++++++- .../proxy/policy_engine/pipeline_executor.py | 22 +-- litellm/proxy/utils.py | 10 +- .../test_anthropic_guardrail_handler.py | 66 +++++++ .../test_openai_guardrail_handler.py | 73 ++++++++ ...test_openai_responses_guardrail_handler.py | 88 +++++++++ .../policy_engine/test_pipeline_executor.py | 22 ++- .../proxy_logging/test_guardrail_pipeline.py | 9 +- 11 files changed, 618 insertions(+), 72 deletions(-) diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py index e486be12fe2..2049866b444 100644 --- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py +++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py @@ -13,10 +13,11 @@ Pattern Overview: """ import json -from collections.abc import Iterator, Mapping, Sequence +from collections.abc import Callable, Mapping, Sequence from copy import deepcopy from dataclasses import dataclass from itertools import chain, repeat +from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Protocol, cast, overload, runtime_checkable from typing_extensions import ReadOnly, TypedDict, assert_never @@ -41,6 +42,7 @@ from litellm.llms.base_llm.guardrail_translation.utils import ( merge_guardrailed_scoped_messages, merge_returned_tools_into_request_tools, scoped_structured_message_indices, + stream_item_field, stream_item_fingerprint, ) from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import ( @@ -153,6 +155,28 @@ class ExtractedInput: EMPTY_EXTRACTED_INPUT: Final = ExtractedInput(scanned=(), images=()) +@dataclass(frozen=True, slots=True) +class _ToolCallShape: + name: str | None + arguments: str + + +_SSEEventRewriter = Callable[[Mapping[str, object]], Mapping[str, object] | None] + + +def _tool_call_shapes(tool_calls: Sequence[object]) -> tuple[_ToolCallShape, ...]: + """The guardrail-visible shape of each tool call, whether the guardrail handed + back the ``ChatCompletionMessageToolCall`` objects it was given or plain dicts.""" + functions: Final = tuple(stream_item_field(tool_call, "function") for tool_call in tool_calls) + return tuple( + _ToolCallShape( + name=name if isinstance(name := stream_item_field(function, "name"), str) else None, + arguments=arguments if isinstance(arguments := stream_item_field(function, "arguments"), str) else "", + ) + for function in functions + ) + + class _AnthropicSSEDelta(TypedDict, total=False): type: ReadOnly[str] text: ReadOnly[str] @@ -170,7 +194,7 @@ class AnthropicMessagesHandler(BaseTranslation): them through guardrail rewrites; downstream provider handling is out of scope. """ - delivers_ended_stream_text_rewrites = True + delivers_ended_stream_rewrites = True def __init__(self): super().__init__() @@ -1050,6 +1074,7 @@ class AnthropicMessagesHandler(BaseTranslation): first_choice.message.tool_calls, ) string_so_far = first_choice.message.content + pre_guardrail_tool_calls: Final = _tool_call_shapes(tool_calls_list or ()) guardrail_inputs: Final = GenericGuardrailAPIInputs() if string_so_far: guardrail_inputs["texts"] = [string_so_far] @@ -1084,6 +1109,19 @@ class AnthropicMessagesHandler(BaseTranslation): and guardrailed_texts[0] != string_so_far ): self._write_ended_stream_text_rewrite(responses_so_far, guardrailed_texts[0]) + if deliver_ended_stream_rewrites: + returned_tool_calls: Final = _guardrailed_inputs.get("tool_calls") + self._write_ended_stream_tool_call_rewrites( + responses_so_far, + pre_guardrail_tool_calls=pre_guardrail_tool_calls, + post_guardrail_tool_calls=_tool_call_shapes( + returned_tool_calls + if isinstance(returned_tool_calls, list) + and len(returned_tool_calls) == len(pre_guardrail_tool_calls) + else tool_calls_list or () + ), + guardrail_name=guardrail_to_apply.guardrail_name or "unknown", + ) else: verbose_proxy_logger.debug("Skipping output guardrail - model response has no choices") return responses_so_far @@ -1212,38 +1250,120 @@ class AnthropicMessagesHandler(BaseTranslation): """Deliver an ended-stream guardrail text rewrite by rewriting the buffered chunks in place: the first ``text_delta`` carries the full rewritten text and every later one is blanked, leaving the surrounding - message and content-block framing untouched. Handles both chunk formats - this stream carries (parsed event dicts and raw SSE bytes).""" + message and content-block framing untouched.""" replacements: Final = chain((rewritten_text,), repeat("")) - for idx, item in enumerate(responses_so_far): - if isinstance(item, dict): - delta = item.get("delta") - if item.get("type") == "content_block_delta" and isinstance(delta, dict): - if delta.get("type") == "text_delta": - delta["text"] = next(replacements) - elif isinstance(item, (bytes, bytearray)): - responses_so_far[idx] = ( # rebind-ok: delivers the rewrite into the caller's buffer - AnthropicMessagesHandler._rewrite_sse_text_deltas(bytes(item), replacements) - ) + + def rewrite_text_delta(event: Mapping[str, object]) -> Mapping[str, object] | None: + delta: Final = event.get("delta") + if event.get("type") != "content_block_delta" or not isinstance(delta, Mapping): + return None + if delta.get("type") != "text_delta": + return None + return {**event, "delta": {**delta, "text": next(replacements)}} + + AnthropicMessagesHandler._rewrite_ended_stream_events(responses_so_far, rewrite_text_delta) + + @classmethod + def _write_ended_stream_tool_call_rewrites( + cls, + responses_so_far: list[Any], # mutable-ok: rewrites the caller's buffered chunks in place + *, + pre_guardrail_tool_calls: tuple[_ToolCallShape, ...], + post_guardrail_tool_calls: tuple[_ToolCallShape, ...], + guardrail_name: str, + ) -> None: + """Deliver ended-stream guardrail tool-call rewrites by rewriting the + buffered chunks in place: the rebuilt response lists tool calls in the + order of the stream's ``tool_use`` blocks, so the nth rewritten call lands + on the nth block, its first ``input_json_delta`` carrying the full rewritten + arguments, every later one blanked, and ``content_block_start`` carrying the + rewritten name. Blocks that do not line up with the rebuilt tool calls make + the rewrite undeliverable, so the pipeline executor discards it and releases + the original chunks.""" + if post_guardrail_tool_calls == pre_guardrail_tool_calls: + return + block_indices: Final = tuple( + index + for item in responses_so_far + for event in cls._iter_sse_events(item) + if event.get("type") == "content_block_start" + and isinstance(block := event.get("content_block"), Mapping) + and block.get("type") == "tool_use" + and isinstance(index := event.get("index"), int) + ) + if len(block_indices) != len(post_guardrail_tool_calls): + from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite + + raise UndeliverableStreamRewrite(guardrail_name) + rewrites_by_block: Final = MappingProxyType( + { + index: after + for index, before, after in zip(block_indices, pre_guardrail_tool_calls, post_guardrail_tool_calls) + if after != before + } + ) + argument_replacements: Final = MappingProxyType( + {index: chain((rewrite.arguments,), repeat("")) for index, rewrite in rewrites_by_block.items()} + ) + + def rewrite_tool_use(event: Mapping[str, object]) -> Mapping[str, object] | None: + index: Final = event.get("index") + if not isinstance(index, int) or index not in rewrites_by_block: + return None + match event.get("type"): + case "content_block_start": + block: Final = event.get("content_block") + name: Final = rewrites_by_block[index].name + if not isinstance(block, Mapping) or name is None: + return None + return {**event, "content_block": {**block, "name": name}} + case "content_block_delta": + delta: Final = event.get("delta") + if not isinstance(delta, Mapping) or delta.get("type") != "input_json_delta": + return None + return {**event, "delta": {**delta, "partial_json": next(argument_replacements[index])}} + case _: + return None + + cls._rewrite_ended_stream_events(responses_so_far, rewrite_tool_use) @staticmethod - def _rewrite_sse_text_deltas(sse_bytes: bytes, replacements: "Iterator[str]") -> bytes: - """Rewrite every ``text_delta`` data line in one SSE chunk with the next - replacement text, leaving all other events and framing byte-identical.""" + def _rewrite_ended_stream_events( + responses_so_far: list[Any], # mutable-ok: rewrites the caller's buffered chunks in place + rewrite_event: _SSEEventRewriter, + ) -> None: + """Replace every buffered event ``rewrite_event`` returns a rewrite for, in + both chunk formats this stream carries (parsed event dicts and raw SSE + bytes), leaving every other event and the framing untouched.""" + rewritten_items: Final = tuple( + AnthropicMessagesHandler._rewrite_buffered_item(item, rewrite_event) for item in responses_so_far + ) + responses_so_far[:] = rewritten_items # rebind-ok: delivers the rewrites into the caller's buffer + + @staticmethod + def _rewrite_buffered_item(item: object, rewrite_event: _SSEEventRewriter) -> object: + if isinstance(item, dict): + rewritten: Final = rewrite_event(_as_str_mapping(item)) + return item if rewritten is None else dict(rewritten) + if isinstance(item, (bytes, bytearray)): + return AnthropicMessagesHandler._rewrite_sse_events(bytes(item), rewrite_event) + return item + + @staticmethod + def _rewrite_sse_events(sse_bytes: bytes, rewrite_event: _SSEEventRewriter) -> bytes: + """Rewrite the data lines of one SSE chunk that ``rewrite_event`` rewrites, + leaving all other events and framing byte-identical.""" try: decoded: Final = sse_bytes.decode("utf-8") except UnicodeDecodeError: return sse_bytes return "\n\n".join( - AnthropicMessagesHandler._rewrite_sse_block(block, replacements) for block in decoded.split("\n\n") + "\n".join(AnthropicMessagesHandler._rewrite_sse_line(line, rewrite_event) for line in block.split("\n")) + for block in decoded.split("\n\n") ).encode("utf-8") @staticmethod - def _rewrite_sse_block(block: str, replacements: "Iterator[str]") -> str: - return "\n".join(AnthropicMessagesHandler._rewrite_sse_line(line, replacements) for line in block.split("\n")) - - @staticmethod - def _rewrite_sse_line(line: str, replacements: "Iterator[str]") -> str: + def _rewrite_sse_line(line: str, rewrite_event: _SSEEventRewriter) -> str: if not line.startswith("data:"): return line try: @@ -1252,14 +1372,10 @@ class AnthropicMessagesHandler(BaseTranslation): ) except json.JSONDecodeError: return line - if not isinstance(data, dict) or data.get("type") != "content_block_delta": + if not isinstance(data, dict): return line - delta: Final = data.get("delta") - if not isinstance(delta, dict) or delta.get("type") != "text_delta": - return line - return "data: " + json.dumps( - {**data, "delta": {**delta, "text": next(replacements)}} # mutable-ok: json.dumps needs plain dicts - ) + rewritten: Final = rewrite_event(_as_str_mapping(data)) + return line if rewritten is None else "data: " + json.dumps(rewritten) def get_streaming_scan_key(self, responses_so_far: Sequence[object]) -> StreamingScanKey | None: stream_ended: Final = self._check_streaming_has_ended(responses_so_far) diff --git a/litellm/llms/base_llm/guardrail_translation/base_translation.py b/litellm/llms/base_llm/guardrail_translation/base_translation.py index afd8e0f67f7..6d1a9ab1c3e 100644 --- a/litellm/llms/base_llm/guardrail_translation/base_translation.py +++ b/litellm/llms/base_llm/guardrail_translation/base_translation.py @@ -52,13 +52,14 @@ class StreamingScanKey: class BaseTranslation(ABC): - delivers_ended_stream_text_rewrites: ClassVar[bool] = False + delivers_ended_stream_rewrites: ClassVar[bool] = False """Whether ``process_output_streaming_response`` accepts ``deliver_ended_stream_rewrites=True`` and, on an ended (fully buffered) - stream, writes guardrail text rewrites back across ``responses_so_far`` so - a buffered pipeline can release rewritten chunks. Tool-call rewrites, and - text rewrites on every other translation, are undeliverable: the pipeline - executor discards them and releases the original chunks.""" + stream, writes guardrail text and tool-call rewrites back across + ``responses_so_far`` so a buffered pipeline can release rewritten chunks, + raising ``UndeliverableStreamRewrite`` for a shape it cannot place. Rewrites + on every other translation are undeliverable: the pipeline executor + discards them and releases the original chunks.""" @staticmethod def transform_user_api_key_dict_to_metadata( @@ -175,9 +176,9 @@ class BaseTranslation(ABC): transformations (see ``StreamTransformSink``); base handlers ignore it. ``deliver_ended_stream_rewrites`` is passed True only when the caller holds the whole buffered stream and the subclass declares - ``delivers_ended_stream_text_rewrites``: the handler then writes - guardrail text rewrites back across ``responses_so_far`` instead of - discarding them. + ``delivers_ended_stream_rewrites``: the handler then writes + guardrail text and tool-call rewrites back across ``responses_so_far`` + instead of discarding them. """ return responses_so_far diff --git a/litellm/llms/openai/chat/guardrail_translation/handler.py b/litellm/llms/openai/chat/guardrail_translation/handler.py index 80292aef2cf..42c95ac3316 100644 --- a/litellm/llms/openai/chat/guardrail_translation/handler.py +++ b/litellm/llms/openai/chat/guardrail_translation/handler.py @@ -49,6 +49,8 @@ from litellm.types.proxy.guardrails.guardrail_hooks.generic_guardrail_api import coerce_stream_holdback_value, ) from litellm.types.utils import ( + ChatCompletionDeltaToolCall, + ChatCompletionMessageToolCall, Choices, GenericGuardrailAPIInputs, ModelResponse, @@ -78,7 +80,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation): Methods can be overridden to customize behavior for different message formats. """ - delivers_ended_stream_text_rewrites = True + delivers_ended_stream_rewrites = True def get_structured_messages(self, data: dict) -> list[AllMessageValues] | None: """ @@ -610,13 +612,14 @@ class OpenAIChatCompletionsHandler(BaseTranslation): deliver_ended_stream_rewrites: bool, ) -> None: """Ended-stream path: rebuild the full response, run the non-streaming - output guardrail against it, and (when opted in) write any text rewrite - back across the buffered chunks.""" + output guardrail against it, and (when opted in) write any text or + tool-call rewrite back across the buffered chunks.""" model_response: Final = cast( ModelResponse, stream_chunk_builder(chunks=responses_so_far, logging_obj=litellm_logging_obj), ) pre_guardrail_texts: Final = self._string_choice_contents(model_response) + pre_guardrail_tool_calls: Final = self._function_tool_call_shapes(model_response) await self.process_output_response( response=model_response, guardrail_to_apply=guardrail_to_apply, @@ -624,13 +627,21 @@ class OpenAIChatCompletionsHandler(BaseTranslation): user_api_key_dict=user_api_key_dict, request_data=request_data, ) - if deliver_ended_stream_rewrites: - await self._write_ended_stream_text_rewrites( - responses_so_far=responses_so_far, - guardrailed_response=model_response, - pre_guardrail_texts=pre_guardrail_texts, - guardrail_name=guardrail_to_apply.guardrail_name or "unknown", - ) + if not deliver_ended_stream_rewrites: + return + guardrail_name: Final = guardrail_to_apply.guardrail_name or "unknown" + await self._write_ended_stream_text_rewrites( + responses_so_far=responses_so_far, + guardrailed_response=model_response, + pre_guardrail_texts=pre_guardrail_texts, + guardrail_name=guardrail_name, + ) + self._write_ended_stream_tool_call_rewrites( + responses_so_far=responses_so_far, + guardrailed_response=model_response, + pre_guardrail_tool_calls=pre_guardrail_tool_calls, + guardrail_name=guardrail_name, + ) def build_stream_error_items( self, @@ -1043,6 +1054,71 @@ class OpenAIChatCompletionsHandler(BaseTranslation): task_mappings=[(target_choice_index, None) for _ in changed], # mutable-ok: callee takes lists ) + @staticmethod + def _function_tool_call_shapes(response: "ModelResponse") -> tuple[tuple[str | None, str], ...]: + return tuple( + (tool_call.function.name, tool_call.function.arguments) + for choice in response.choices + for tool_call in choice.message.tool_calls or () + if isinstance(tool_call, ChatCompletionMessageToolCall) + ) + + @staticmethod + def _function_tool_call_fragments( + responses_so_far: Sequence["ModelResponseStream"], + ) -> tuple[tuple[ChatCompletionDeltaToolCall, ...], ...]: + """Group the stream's function tool-call fragments by their tool-call index, in + the index order ``stream_chunk_builder`` lists the rebuilt tool calls, keeping + only the indices the builder keeps (an id and a name somewhere in the stream).""" + fragments: Final = tuple( + tool_call + for response in responses_so_far + for choice in response.choices + for tool_call in choice.delta.tool_calls or () + if isinstance(tool_call, ChatCompletionDeltaToolCall) + ) + identified: Final = frozenset(fragment.index for fragment in fragments if fragment.id) + named: Final = frozenset(fragment.index for fragment in fragments if fragment.function.name) + return tuple( + tuple(fragment for fragment in fragments if fragment.index == index) for index in sorted(identified & named) + ) + + def _write_ended_stream_tool_call_rewrites( + self, + responses_so_far: list["ModelResponseStream"], # mutable-ok: rewrites the caller's buffered chunks in place + guardrailed_response: "ModelResponse", + pre_guardrail_tool_calls: tuple[tuple[str | None, str], ...], + guardrail_name: str, + ) -> None: + """Write ended-stream guardrail tool-call rewrites back across the buffered + chunks: the rewritten name and full arguments land in the tool call's first + fragment and the arguments of its later fragments are blanked, mirroring the + text write-back. A rewrite on a stream carrying more than one distinct choice + index, or whose fragments do not line up with the rebuilt tool calls, is + reported as undeliverable, so the pipeline executor discards it and releases + the original chunks.""" + post_guardrail_tool_calls: Final = self._function_tool_call_shapes(guardrailed_response) + if post_guardrail_tool_calls == pre_guardrail_tool_calls: + return + stream_choice_indices: Final = frozenset( + choice.index for response in responses_so_far for choice in response.choices + ) + fragments_by_tool_call: Final = self._function_tool_call_fragments(responses_so_far) + if len(stream_choice_indices) != 1 or len(fragments_by_tool_call) != len(post_guardrail_tool_calls): + from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite + + raise UndeliverableStreamRewrite(guardrail_name) + for before, (name, arguments), fragments in zip( + pre_guardrail_tool_calls, post_guardrail_tool_calls, fragments_by_tool_call + ): + if (name, arguments) == before: + continue + head, *tail = fragments + head.function.name = name + head.function.arguments = arguments + for fragment in tail: + fragment.function.arguments = "" + async def _apply_guardrail_responses_to_output_streaming( self, responses: list["ModelResponseStream"], diff --git a/litellm/llms/openai/responses/guardrail_translation/handler.py b/litellm/llms/openai/responses/guardrail_translation/handler.py index b0f79552bc5..447870fc0be 100644 --- a/litellm/llms/openai/responses/guardrail_translation/handler.py +++ b/litellm/llms/openai/responses/guardrail_translation/handler.py @@ -101,6 +101,18 @@ if TYPE_CHECKING: from litellm.types.llms.openai import ResponseInputParam +class _ToolCallShape(NamedTuple): + name: str | None + arguments: str + + +def _tool_call_shapes(tool_calls: Sequence[ChatCompletionToolCallChunk]) -> tuple[_ToolCallShape, ...]: + return tuple( + _ToolCallShape(name=tool_call["function"].get("name"), arguments=tool_call["function"].get("arguments", "")) + for tool_call in tool_calls + ) + + class ResponseOutputEnvelope(TypedDict, total=False): """Dict form of a Responses API response, as far as guardrail write-back reads it.""" @@ -340,7 +352,7 @@ class OpenAIResponsesHandler(BaseTranslation): Methods can be overridden to customize behavior for different message formats. """ - delivers_ended_stream_text_rewrites = True + delivers_ended_stream_rewrites = True def get_structured_messages(self, data: dict) -> list[AllMessageValues] | None: """ @@ -754,6 +766,7 @@ class OpenAIResponsesHandler(BaseTranslation): if response_model: inputs["model"] = response_model + pre_guardrail_tool_calls: Final = _tool_call_shapes(tool_calls_to_check) guardrailed_inputs: Final = await guardrail_to_apply.apply_guardrail( inputs=inputs, request_data=request_data, @@ -762,6 +775,12 @@ class OpenAIResponsesHandler(BaseTranslation): ) guardrailed_texts: Final = guardrailed_inputs.get("texts", []) + returned_tool_calls: Final = guardrailed_inputs.get("tool_calls") + post_guardrail_tool_calls: Final = _tool_call_shapes( + returned_tool_calls + if isinstance(returned_tool_calls, list) and len(returned_tool_calls) == len(tool_calls_to_check) + else tool_calls_to_check + ) # Write guardrailed texts back into the output items in-place. # final_chunk is a reference into responses_so_far so this @@ -784,6 +803,13 @@ class OpenAIResponsesHandler(BaseTranslation): stream_events=responses_so_far[:-1], rewrites_by_position=rewrites_by_position, ) + self._deliver_ended_stream_tool_call_rewrites( + responses_so_far=responses_so_far, + outputs=outputs, + pre_guardrail_tool_calls=pre_guardrail_tool_calls, + post_guardrail_tool_calls=post_guardrail_tool_calls, + guardrail_name=guardrail_to_apply.guardrail_name or "unknown", + ) return responses_so_far # ------------------------------------------------------------------ # @@ -894,6 +920,89 @@ class OpenAIResponsesHandler(BaseTranslation): continue OpenAIResponsesHandler._write_event_field(content[content_idx], "text", rewritten) + def _deliver_ended_stream_tool_call_rewrites( + self, + responses_so_far: Sequence[object], + outputs: Sequence[object], + pre_guardrail_tool_calls: tuple[_ToolCallShape, ...], + post_guardrail_tool_calls: tuple[_ToolCallShape, ...], + guardrail_name: str, + ) -> None: + """Write ended-stream guardrail tool-call rewrites into the completed + envelope's ``function_call`` items and sync the earlier stream events, + keyed by ``output_index``. The guardrail sees the envelope's function + calls in output order, which is how a rewritten call finds its item; a + rewrite whose calls do not line up with the envelope is reported as + undeliverable, so the pipeline executor discards it and releases the + original events.""" + if post_guardrail_tool_calls == pre_guardrail_tool_calls: + return + function_call_indices: Final = tuple( + output_idx + for output_idx, output_item in enumerate(outputs) + if stream_item_field(output_item, "type") == "function_call" + ) + if len(function_call_indices) != len(post_guardrail_tool_calls): + from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite + + raise UndeliverableStreamRewrite(guardrail_name) + rewrites_by_output_index: Final = MappingProxyType( + { + output_idx: after + for output_idx, before, after in zip( + function_call_indices, pre_guardrail_tool_calls, post_guardrail_tool_calls + ) + if after != before + } + ) + for output_idx, rewrite in rewrites_by_output_index.items(): + self._write_function_call_item(outputs[output_idx], rewrite.name, rewrite.arguments) + self._sync_stream_events_with_tool_call_rewrites( + stream_events=responses_so_far[:-1], + rewrites_by_output_index=rewrites_by_output_index, + ) + + def _sync_stream_events_with_tool_call_rewrites( + self, + stream_events: Sequence[object], + rewrites_by_output_index: Mapping[int, _ToolCallShape], + ) -> None: + """Sync pre-completion function-call events with the rewritten completed + response: the first ``function_call_arguments.delta`` for a rewritten call + carries the full rewritten arguments and the rest are blanked, while + ``function_call_arguments.done`` and ``output_item.done`` carry the full + rewritten arguments and ``output_item.added`` / ``output_item.done`` the + rewritten name, so every event a client may read agrees with the + rewritten ``response.completed`` payload.""" + delta_replacements: Final = MappingProxyType( + {index: chain((rewrite.arguments,), repeat("")) for index, rewrite in rewrites_by_output_index.items()} + ) + for event in stream_events: + output_index = stream_item_field(event, "output_index") + if not isinstance(output_index, int) or output_index not in rewrites_by_output_index: + continue + rewrite = rewrites_by_output_index[output_index] + match stream_item_field(event, "type"): + case "response.function_call_arguments.delta": + self._write_event_field(event, "delta", next(delta_replacements[output_index])) + case "response.function_call_arguments.done": + self._write_event_field(event, "arguments", rewrite.arguments) + case "response.output_item.added": + self._write_function_call_item(stream_item_field(event, "item"), rewrite.name, None) + case "response.output_item.done": + self._write_function_call_item(stream_item_field(event, "item"), rewrite.name, rewrite.arguments) + case _: + pass + + @staticmethod + def _write_function_call_item(item: object, name: str | None, arguments: str | None) -> None: + if not (isinstance(item, dict) or hasattr(item, "get")): + return + if name is not None: + OpenAIResponsesHandler._write_event_field(item, "name", name) + if arguments is not None: + OpenAIResponsesHandler._write_event_field(item, "arguments", arguments) + def _check_streaming_has_ended(self, responses_so_far: Sequence[object]) -> bool: """ Check if the streaming has ended. diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py index 9bc10949e9f..3e112bf8e67 100644 --- a/litellm/proxy/policy_engine/pipeline_executor.py +++ b/litellm/proxy/policy_engine/pipeline_executor.py @@ -85,10 +85,10 @@ def _logged_by_inner_guardrail(method: _GuardrailMethodT) -> _GuardrailMethodT: class _StreamRewriteObserver(CustomGuardrail): """Stand-in handed to the endpoint translation in place of a streaming pipeline step's guardrail. It records whether the guardrail returned different output than it was given, - which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text - rewrites are deliverable on translations that write them back across the buffered chunks - (``delivers_ended_stream_text_rewrites``); tool-call rewrites and text rewrites on any - other translation are discarded by the executor, which releases the original chunks. + which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text and + tool-call rewrites are deliverable on translations that write them back across the + buffered chunks (``delivers_ended_stream_rewrites``); rewrites on any other translation + are discarded by the executor, which releases the original chunks. The inner guardrail's ``apply_guardrail`` already records the guardrail information and span, so the observer's stays out of ``log_guardrail_information``.""" @@ -290,13 +290,13 @@ class PipelineExecutor: litellm_logging_obj: "LiteLLMLoggingObj | None", ) -> None: """Run one streaming post_call step through the endpoint translation, delivering - text rewrites on translations that support ended-stream write-back. A rewrite that - cannot reach the client yet (a tool-call rewrite, a text rewrite on a translation - without write-back, or one the translation refused with - ``UndeliverableStreamRewrite``) is discarded: the buffered chunks go back to the - originals and the step passes, so the client gets the stream the merge base sent.""" + text and tool-call rewrites on translations that support ended-stream write-back. A + rewrite that cannot reach the client yet (one on a translation without write-back, or + one the translation refused with ``UndeliverableStreamRewrite``) is discarded: the + buffered chunks go back to the originals and the step passes, so the client gets the + stream the merge base sent.""" observer: Final = _StreamRewriteObserver(callback) - deliver_rewrites: Final = type(endpoint_translation).delivers_ended_stream_text_rewrites + deliver_rewrites: Final = type(endpoint_translation).delivers_ended_stream_rewrites originals: Final = copy.deepcopy(streaming_chunks) try: if deliver_rewrites: @@ -319,7 +319,7 @@ class PipelineExecutor: except UndeliverableStreamRewrite: _release_original_chunks(step.guardrail, streaming_chunks, originals) else: - if observer.rewrote_tool_calls or (observer.rewrote_texts and not deliver_rewrites): + if not deliver_rewrites and (observer.rewrote_texts or observer.rewrote_tool_calls): _release_original_chunks(step.guardrail, streaming_chunks, originals) if not callback.records_own_guardrail_information: add_guardrail_to_applied_guardrails_header(request_data=hook_input, guardrail_name=step.guardrail) diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py index 8b36061c6be..76b9b4b44ed 100644 --- a/litellm/proxy/utils.py +++ b/litellm/proxy/utils.py @@ -3445,11 +3445,11 @@ class ProxyLogging: assembled output through the endpoint guardrail translation, the same machinery flat post_call guardrails use at end of stream. An allow releases the buffered chunks: verbatim when no guardrail rewrote the - output, rewritten in place when one rewrote text and the translation - delivers ended-stream rewrites (later steps then re-scan the rewritten - chunks, so rewrites chain). A rewrite the translation cannot deliver - yet (a tool-call rewrite, or a text rewrite on a route without - write-back) is discarded by the executor and the original chunks are + output, rewritten in place when one rewrote text or a tool call and the + translation delivers ended-stream rewrites (later steps then re-scan the + rewritten chunks, so rewrites chain). A rewrite the translation cannot + deliver yet (one on a route without write-back, or a shape the route + refuses) is discarded by the executor and the original chunks are released, as is a buffered shape no translation resolves; a block or modify_response terminates with the translation's block chunks or the raised error. diff --git a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py index bf40f781fa3..bd2dff9b33c 100644 --- a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py +++ b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py @@ -315,6 +315,72 @@ class TestAnthropicMessagesHandlerStreamingOutputProcessing: assert "event: message_start" in raw and "event: message_stop" in raw assert '"stop_reason": "end_turn"' in raw + @staticmethod + def _ended_tool_use_sse_chunks() -> list: + events = [ + ("message_start", {"type": "message_start", "message": {"id": "msg_1", "type": "message", "role": "assistant", "model": "claude-sonnet-4-5", "content": [], "stop_reason": None, "usage": {"input_tokens": 1, "output_tokens": 0}}}), + ("content_block_start", {"type": "content_block_start", "index": 0, "content_block": {"type": "tool_use", "id": "toolu_1", "name": "lookup_fruit", "input": {}}}), + ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "input_json_delta", "partial_json": ""}}), + ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "input_json_delta", "partial_json": '{"fruit": "persim'}}), + ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "input_json_delta", "partial_json": 'mon"}'}}), + ("content_block_stop", {"type": "content_block_stop", "index": 0}), + ("message_delta", {"type": "message_delta", "delta": {"stop_reason": "tool_use", "stop_sequence": None}, "usage": {"output_tokens": 2}}), + ("message_stop", {"type": "message_stop"}), + ] + return [f"event: {name}\ndata: {json.dumps(payload)}\n\n".encode() for name, payload in events] + + @staticmethod + def _argument_masking_guardrail() -> CustomGuardrail: + class MaskArguments(CustomGuardrail): + async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None): + for tool_call in inputs.get("tool_calls", []): + tool_call.function.arguments = '{"fruit": "[MASKED]"}' + return inputs + + return MaskArguments(guardrail_name="test") + + @staticmethod + def _partial_jsons(chunks: list) -> list: + return [ + json.loads(line[len("data:") :].strip())["delta"]["partial_json"] + for chunk in chunks + for line in chunk.decode().split("\n") + if line.startswith("data:") and json.loads(line[len("data:") :].strip()).get("type") == "content_block_delta" + ] + + @pytest.mark.asyncio + async def test_deliver_ended_stream_rewrites_writes_tool_use_input_back_into_sse_chunks(self): + handler = AnthropicMessagesHandler() + chunks = self._ended_tool_use_sse_chunks() + + result = await handler.process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=self._argument_masking_guardrail(), + litellm_logging_obj=MagicMock(), + deliver_ended_stream_rewrites=True, + ) + + assert result is chunks + assert self._partial_jsons(chunks) == ['{"fruit": "[MASKED]"}', "", ""] + raw = b"".join(chunks).decode() + assert '"name": "lookup_fruit"' in raw and '"id": "toolu_1"' in raw + assert '"stop_reason": "tool_use"' in raw + assert "persim" not in raw + + @pytest.mark.asyncio + async def test_ended_stream_tool_use_rewrite_leaves_chunks_untouched_by_default(self): + handler = AnthropicMessagesHandler() + chunks = self._ended_tool_use_sse_chunks() + original = [bytes(chunk) for chunk in chunks] + + await handler.process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=self._argument_masking_guardrail(), + litellm_logging_obj=MagicMock(), + ) + + assert chunks == original + @pytest.mark.asyncio async def test_ended_stream_rewrite_leaves_chunks_untouched_by_default(self): handler = AnthropicMessagesHandler() diff --git a/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py b/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py index aff0530ee8b..be168ec83e6 100644 --- a/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py +++ b/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py @@ -1113,6 +1113,79 @@ class TestOpenAIChatCompletionsHandlerStreamingOutput: assert chunks[1].choices[0].delta.content in (None, "") assert chunks[1].choices[0].finish_reason == "stop" + @staticmethod + def _ended_tool_call_stream_chunks() -> list: + from litellm.types.utils import ( + ChatCompletionDeltaToolCall, + Delta, + Function, + ModelResponseStream, + StreamingChoices, + ) + + def chunk(tool_call: ChatCompletionDeltaToolCall | None, finish_reason: Optional[str] = None): + return ModelResponseStream( + id="chatcmpl-123", + created=1234567890, + model="gpt-4", + object="chat.completion.chunk", + choices=[ + StreamingChoices( + index=0, + delta=Delta(tool_calls=[tool_call] if tool_call else None), + finish_reason=finish_reason, + ) + ], + ) + + def fragment(arguments: str, name: Optional[str] = None, call_id: Optional[str] = None): + return ChatCompletionDeltaToolCall( + id=call_id, index=0, type="function", function=Function(name=name, arguments=arguments) + ) + + return [ + chunk(fragment("", name="lookup_fruit", call_id="call_1")), + chunk(fragment('{"fruit":')), + chunk(fragment(' "persimmon"}')), + chunk(None, finish_reason="tool_calls"), + ] + + @pytest.mark.asyncio + async def test_deliver_ended_stream_rewrites_writes_tool_call_arguments_back_into_chunks(self): + handler = OpenAIChatCompletionsHandler() + guardrail = MockGuardrail(guardrail_name="test") + chunks = self._ended_tool_call_stream_chunks() + + result = await handler.process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=guardrail, + litellm_logging_obj=None, + deliver_ended_stream_rewrites=True, + ) + + assert result is chunks + fragments = [chunk.choices[0].delta.tool_calls for chunk in chunks[:3]] + assert [fragment[0].function.arguments for fragment in fragments] == ['{"fruit": "PERSIMMON"}', "", ""] + assert fragments[0][0].function.name == "lookup_fruit" + assert fragments[0][0].id == "call_1" + assert chunks[3].choices[0].delta.tool_calls is None + assert chunks[3].choices[0].finish_reason == "tool_calls" + + @pytest.mark.asyncio + async def test_ended_stream_tool_call_rewrite_leaves_chunks_untouched_by_default(self): + handler = OpenAIChatCompletionsHandler() + guardrail = MockGuardrail(guardrail_name="test") + chunks = self._ended_tool_call_stream_chunks() + + await handler.process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=guardrail, + litellm_logging_obj=None, + ) + + fragments = [chunk.choices[0].delta.tool_calls for chunk in chunks[:3]] + assert [fragment[0].function.arguments for fragment in fragments] == ["", '{"fruit":', ' "persimmon"}'] + @pytest.mark.asyncio async def test_ended_stream_rewrite_leaves_chunks_untouched_by_default(self): handler = OpenAIChatCompletionsHandler() diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py index 33c8c97fea7..a024fc7ff81 100644 --- a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py +++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py @@ -1195,6 +1195,94 @@ class TestOpenAIResponsesHandlerStreamingOutputProcessing: assert events[4]["item"]["content"][0]["text"] == "hello [MASKED]" assert events[5]["response"]["output"][0]["content"][0]["text"] == "hello [MASKED]" + @staticmethod + def _ended_function_call_stream_events() -> List[dict]: + def item(arguments: str, status: str) -> dict: + return { + "type": "function_call", + "id": "fc_123", + "call_id": "call_123", + "name": "lookup_fruit", + "arguments": arguments, + "status": status, + } + + return [ + {"type": "response.output_item.added", "output_index": 0, "item": item("", "in_progress")}, + {"type": "response.function_call_arguments.delta", "item_id": "fc_123", "output_index": 0, "delta": '{"fruit":'}, + {"type": "response.function_call_arguments.delta", "item_id": "fc_123", "output_index": 0, "delta": ' "persimmon"}'}, + { + "type": "response.function_call_arguments.done", + "item_id": "fc_123", + "output_index": 0, + "arguments": '{"fruit": "persimmon"}', + }, + {"type": "response.output_item.done", "output_index": 0, "item": item('{"fruit": "persimmon"}', "completed")}, + { + "type": "response.completed", + "response": { + "id": "resp_123", + "model": "gpt-4o", + "output": [item('{"fruit": "persimmon"}', "completed")], + "status": "completed", + }, + }, + ] + + @staticmethod + def _argument_masking_guardrail() -> CustomGuardrail: + class MaskArguments(CustomGuardrail): + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict, + input_type: Literal["request", "response"], + logging_obj: Optional[Any] = None, + ) -> GenericGuardrailAPIInputs: + tool_calls = [ + {**tool_call, "function": {**tool_call["function"], "arguments": '{"fruit": "[MASKED]"}'}} + for tool_call in inputs.get("tool_calls", []) + ] + return {**inputs, "tool_calls": tool_calls} + + return MaskArguments(guardrail_name="test-mask-arguments") + + @pytest.mark.asyncio + async def test_deliver_ended_stream_rewrites_syncs_function_call_events(self): + handler = OpenAIResponsesHandler() + events = self._ended_function_call_stream_events() + + result = await handler.process_output_streaming_response( + responses_so_far=events, + guardrail_to_apply=self._argument_masking_guardrail(), + litellm_logging_obj=None, + deliver_ended_stream_rewrites=True, + ) + + assert result is events + assert events[0]["item"]["arguments"] == "" + assert events[1]["delta"] == '{"fruit": "[MASKED]"}' + assert events[2]["delta"] == "" + assert events[3]["arguments"] == '{"fruit": "[MASKED]"}' + assert events[4]["item"]["arguments"] == '{"fruit": "[MASKED]"}' + assert events[5]["response"]["output"][0]["arguments"] == '{"fruit": "[MASKED]"}' + assert events[5]["response"]["output"][0]["name"] == "lookup_fruit" + + @pytest.mark.asyncio + async def test_ended_stream_function_call_rewrite_leaves_events_untouched_by_default(self): + handler = OpenAIResponsesHandler() + events = self._ended_function_call_stream_events() + + await handler.process_output_streaming_response( + responses_so_far=events, + guardrail_to_apply=self._argument_masking_guardrail(), + litellm_logging_obj=None, + ) + + assert events[1]["delta"] == '{"fruit":' + assert events[3]["arguments"] == '{"fruit": "persimmon"}' + assert events[5]["response"]["output"][0]["arguments"] == '{"fruit": "persimmon"}' + @pytest.mark.asyncio @pytest.mark.parametrize("terminal_type", ["response.incomplete", "response.failed"]) async def test_deliver_ended_stream_rewrites_syncs_non_completed_terminals(self, terminal_type): diff --git a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py index 54ef1f79f4d..ee75003db9f 100644 --- a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py +++ b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py @@ -924,7 +924,7 @@ class _TextReturningGuardrail(CustomGuardrail): class _TextTranslation: - delivers_ended_stream_text_rewrites = False + delivers_ended_stream_rewrites = False def __init__(self): self.seen_guardrail_names = [] @@ -946,7 +946,7 @@ class _WritingTranslation: """Writes the guardrail's text (and tool-call) outputs back into the buffered chunks the way the chat/Responses/Messages handlers do on an ended stream.""" - delivers_ended_stream_text_rewrites = True + delivers_ended_stream_rewrites = True async def process_output_streaming_response( self, @@ -970,7 +970,7 @@ class _WritingTranslation: class _RefusingTranslation: - delivers_ended_stream_text_rewrites = True + delivers_ended_stream_rewrites = True async def process_output_streaming_response( self, @@ -1088,13 +1088,27 @@ async def test_streaming_step_delivers_text_rewrite_through_writing_translation( @pytest.mark.asyncio -async def test_streaming_step_discards_tool_call_rewrite_and_restores_written_text(monkeypatch, caplog): +async def test_streaming_step_delivers_tool_call_rewrite_through_writing_translation(monkeypatch, caplog): monkeypatch.setattr(litellm, "callbacks", [_TextAndToolCallRewritingGuardrail(rewrite_tool_call=True)]) chunks = [_chunk()] with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): result = await _run_streaming_step(_WritingTranslation(), chunks) + assert result.terminal_action == "allow" + assert chunks[0]["text"] == "hello [MASKED]" + assert chunks[0]["tool_call"]["function"]["arguments"] == '{"ssn": "[MASKED]"}' + assert not any("discarded" in record.getMessage() for record in caplog.records) + + +@pytest.mark.asyncio +async def test_streaming_step_discards_tool_call_rewrite_when_translation_lacks_write_back(monkeypatch, caplog): + monkeypatch.setattr(litellm, "callbacks", [_TextAndToolCallRewritingGuardrail(rewrite_tool_call=True)]) + chunks = [_chunk()] + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + result = await _run_streaming_step(_TextTranslation(), chunks) + _assert_passed_with_discard_warning(result, caplog) assert chunks == [_chunk()] diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py index 73dd6746b29..d27f504ba6d 100644 --- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py +++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py @@ -1836,7 +1836,7 @@ def _echoed_tool_call_dicts(arguments: str) -> List[Dict[str, Any]]: @pytest.mark.asyncio @pytest.mark.parametrize("on_fail, on_error", [("block", None), ("next", "next")]) -async def test_streaming_iterator_hook_pipeline_releases_originals_on_runtime_tool_call_rewrite( +async def test_streaming_iterator_hook_pipeline_delivers_runtime_tool_call_rewrite( proxy_logging, make_user_api_key_auth, monkeypatch, on_fail, on_error, caplog ): transform = lambda inputs: {"tool_calls": _echoed_tool_call_dicts('{"ssn": "[MASKED]"}')} # noqa: E731 @@ -1855,9 +1855,12 @@ async def test_streaming_iterator_hook_pipeline_releases_originals_on_runtime_to delivered.append(item) assert len(delivered) == 2 - assert delivered[0].choices[0].delta.tool_calls[0].function.arguments == '{"ssn": "123"}' + delivered_tool_call = delivered[0].choices[0].delta.tool_calls[0] + assert delivered_tool_call.function.arguments == '{"ssn": "[MASKED]"}' + assert delivered_tool_call.function.name == "lookup" + assert delivered_tool_call.id == "call_1" assert delivered[1].choices[0].finish_reason == "tool_calls" - assert any("'gr-post'" in message and "discarded" in message for message in _warnings(caplog)) + assert not any("discarded" in message for message in _warnings(caplog)) @pytest.mark.asyncio