fix(guardrails): honor skip_tool_message_in_guardrail on the bedrock during_call path

Anthropic tool_result blocks were always flattened into the ApplyGuardrail
INPUT scan, so skip_tool_message_in_guardrail had no effect on the native
during_call path and masked write-back could shift indices onto the wrong
parts. The leaf walk now takes skip_tool_results, driven by
effective_skip_tool_message_for_guardrail, the masking rewrite skips
tool_result items when the flag is on, and the bedrock initializer passes
the litellm_params flag through. The refusal walk stays unfiltered so
attachments inside a skipped tool_result are still refused fail-closed.

Co-authored-by: feng.tsai <fengtsai@berri.ai>
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
yucheng 2026-09-26 08:52:19 +00:00
parent be1b0d5d83
commit f06bc06a0e
2 changed files with 24 additions and 7 deletions

View file

@ -45,6 +45,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import BedrockImageProc
from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler
from litellm.llms.base_llm.guardrail_translation.utils import (
effective_scan_only_tool_results_for_guardrail,
effective_skip_tool_message_for_guardrail,
)
from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, bedrock_bearer_token, run_aws_signing
from litellm.llms.custom_httpx.http_handler import (
@ -283,13 +284,15 @@ def _function_call_output_parts(data: Mapping[str, object]) -> tuple[Mapping[str
)
def _content_leaf_parts(content: object) -> tuple[object, ...]:
def _content_leaf_parts(content: object, *, skip_tool_results: bool = False) -> tuple[object, ...]:
"""Leaf content parts with Anthropic tool_result blocks flattened.
A tool_result carrying nested blocks hands the model their contents, so a nested
document or image must be counted, scanned, or refused exactly like a top-level
part. Its string content becomes a text leaf. Iterative (no recursion): the
code-quality check bans new recursive functions.
part. Its string content becomes a text leaf. skip_tool_results=True drops the
whole tool_result part; the refusal walk stays unfiltered so attachments inside
a skipped tool_result are still refused fail-closed. Iterative (no recursion):
the code-quality check bans new recursive functions.
"""
if not isinstance(content, list):
return ()
@ -298,6 +301,8 @@ def _content_leaf_parts(content: object) -> tuple[object, ...]:
while stack:
part = stack.pop()
if isinstance(part, dict) and part.get("type") == "tool_result":
if skip_tool_results:
continue
inner = part.get("content")
if isinstance(inner, str):
leaves.append(inner)
@ -359,6 +364,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
self.chunk_budget_chars = chunk_budget_chars
self.contextual_grounding_from_messages = contextual_grounding_from_messages
self.experimental_use_latest_role_message_only = bool(kwargs.get("experimental_use_latest_role_message_only"))
self.skip_tool_message_in_guardrail: bool | None = kwargs.get("skip_tool_message_in_guardrail")
# Resource-less, detect-only InvokeGuardrailChecks mode. Present `checks`
# routes the guardrail to InvokeGuardrailChecks; absent => ApplyGuardrail.
@ -506,7 +512,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
return (BedrockContentItem(text=BedrockTextContent(text=content)),)
if not isinstance(content, list):
return ()
parts: Final = _content_leaf_parts(content)
parts: Final = _content_leaf_parts(content, skip_tool_results=effective_skip_tool_message_for_guardrail(self))
items: Final = await asyncio.gather(*(self._build_input_content_item(item=item) for item in parts))
return tuple(item for item in items if item is not None)
@ -3318,6 +3324,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
"""
updated_messages: Final = []
masking_index = 0
skip_tool_results: Final = effective_skip_tool_message_for_guardrail(self)
for message in messages:
new_message = message.copy()
@ -3339,6 +3346,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
content_list=content,
masked_texts=masked_texts,
masking_index=masking_index,
skip_tool_results=skip_tool_results,
)
updated_messages.append(new_message)
@ -3363,8 +3371,13 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
@staticmethod
def _mask_tool_result(
item: Mapping[str, object], masked_texts: Sequence[str], masking_index: int
item: Mapping[str, object],
masked_texts: Sequence[str],
masking_index: int,
skip_tool_results: bool,
) -> tuple[Mapping[str, object], int]:
if skip_tool_results:
return item, masking_index
inner: Final = item.get("content")
if isinstance(inner, str):
masked, masking_index = BedrockGuardrail._mask_leaf(
@ -3386,7 +3399,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
return item, masking_index
def _mask_content_list(
self, content_list: Sequence[object], masked_texts: list[str], masking_index: int
self, content_list: Sequence[object], masked_texts: list[str], masking_index: int, skip_tool_results: bool
) -> tuple[list[Any], int]:
"""
Apply masking to a list of content items, walking the same leaves
@ -3405,7 +3418,10 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
for item in content_list:
if isinstance(item, dict) and item.get("type") == "tool_result":
new_item, masking_index = self._mask_tool_result(
item=item, masked_texts=masked_texts, masking_index=masking_index
item=item,
masked_texts=masked_texts,
masking_index=masking_index,
skip_tool_results=skip_tool_results,
)
new_content.append(new_item)
else:

View file

@ -41,6 +41,7 @@ def initialize_bedrock(litellm_params: LitellmParams, guardrail: Guardrail):
aws_bedrock_runtime_endpoint=litellm_params.aws_bedrock_runtime_endpoint,
experimental_use_latest_role_message_only=litellm_params.experimental_use_latest_role_message_only,
only_scan_new_messages=litellm_params.only_scan_new_messages or False,
skip_tool_message_in_guardrail=litellm_params.skip_tool_message_in_guardrail,
streaming_buffer_until_moderated=streaming_params.streaming_buffer_until_moderated,
streaming_sampling_rate=streaming_params.streaming_sampling_rate,
streaming_end_of_stream_only=streaming_params.streaming_end_of_stream_only,