fix(anthropic): strip beta-gated system fields before an older model's system turn becomes a user turn

Models without supports_mid_conversation_system send no mid-conversation beta,
so tool_addition and tool_removal blocks are dropped (and logged) before the
turn is converted instead of riding along as user content. The Invoke
output_config regression test reads the real cost map instead of patching
_bedrock_model_supports

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/llms/anthropic/messages/native/transformation.py

@@ -7,8 +7,10 @@ from pydantic import TypeAdapter
 import litellm
 from litellm._logging import verbose_logger
 from litellm.anthropic_beta_gated_fields import (
+    StrippedMessages,
     log_beta_gated_strips,
     strip_beta_gated_native_request,
+    strip_beta_gated_system_messages,
     unstripped_native_request,
 )
 from litellm.anthropic_beta_headers_manager import get_unsupported_headers, is_beta_header_supported_by_bundled_config
@@ -53,6 +55,8 @@ DEFAULT_ANTHROPIC_API_VERSION: Final = "2023-06-01"
 
 _CALLER_CREDENTIAL_HEADERS: Final = frozenset({"x-api-key", "authorization"})
 
+_MESSAGE_MAPPINGS: Final = TypeAdapter(tuple[Mapping[str, object], ...])
+
 
 def _carries_caller_credential(headers: Mapping[str, str]) -> bool:
     """Whether the caller sent their own Anthropic credential, in which case this passthrough
@@ -265,14 +269,22 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             len(messages),
         )
         hoisted: Final = messages[:leading_count]
+        keeps_system_turns: Final = supports_factory(
+            model=model,
+            custom_llm_provider=self.custom_llm_provider,
+            key="supports_mid_conversation_system",
+        )
+        trailing_messages: Final = _MESSAGE_MAPPINGS.validate_python(messages[leading_count:])  # pyright: ignore[reportUnknownArgumentType]  # messages comes from the bare request dict
+        gated: Final = (
+            StrippedMessages(messages=trailing_messages, stripped=())
+            if keeps_system_turns
+            else strip_beta_gated_system_messages(trailing_messages, frozenset())
+        )
+        log_beta_gated_strips(gated.stripped, model=model, provider=self.custom_llm_provider or "anthropic")
         remaining: Final = (
             messages[leading_count:]
-            if supports_factory(
-                model=model,
-                custom_llm_provider=self.custom_llm_provider,
-                key="supports_mid_conversation_system",
-            )
-            else list(convert_mid_conversation_system_turns(messages[leading_count:]))
+            if keeps_system_turns
+            else list(convert_mid_conversation_system_turns(_MESSAGE_MAPPINGS.validate_python(gated.messages)))
         )
         if hoisted or remaining != messages:
             anthropic_messages_request["messages"] = remaining

tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude_transformation.py

@@ -16,7 +16,7 @@ import httpx
 import pytest
 import requests_mock as requests_mock_lib
 import respx
-from pydantic import TypeAdapter
+from pydantic import BaseModel, ConfigDict, TypeAdapter
 
 import litellm
 from litellm.constants import (
@@ -4200,10 +4200,9 @@ def test_bedrock_messages_without_guardrail_config_sends_no_guardrail_headers()
     assert not any(name.startswith("X-Amzn-Bedrock-Guardrail") for name in headers)
 
 
-def test_bedrock_messages_keeps_mid_conversation_output_config_its_auto_added_beta_gates():
+@pytest.mark.usefixtures("local_model_cost_map", "local_beta_headers_config")
+def test_bedrock_messages_keeps_mid_conversation_output_config_its_auto_added_beta_gates() -> None:
     """Invoke learns its body betas after the shared transform, so the gated field it supports must survive."""
-    from unittest.mock import patch
-
     from litellm.types.router import GenericLiteLLMParams
 
     messages: Final = [
@@ -4212,14 +4211,58 @@ def test_bedrock_messages_keeps_mid_conversation_output_config_its_auto_added_be
         {"role": "system", "content": "Answer briefly", "output_config": {"effort": "high"}},
         {"role": "user", "content": [{"type": "text", "text": "Go on"}]},
     ]
-    with patch("litellm.llms.bedrock.common_utils._bedrock_model_supports", return_value=True):
-        result: Final = AmazonAnthropicClaudeMessagesConfig().transform_anthropic_messages_request(  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]  # the base signature returns a bare dict
-            model="anthropic.claude-opus-4-8-v1:0",
-            messages=messages,
-            anthropic_messages_optional_request_params={"max_tokens": 4096},
-            litellm_params=GenericLiteLLMParams(),
-            headers={},
-        )
+    result: Final = AmazonAnthropicClaudeMessagesConfig().transform_anthropic_messages_request(
+        model="anthropic.claude-opus-4-8-v1:0",
+        messages=messages,
+        anthropic_messages_optional_request_params={"max_tokens": 4096},
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
 
     assert "mid-conversation-output-config-2026-07-01" in result["anthropic_beta"]
     assert {"role": "system", "content": "Answer briefly", "output_config": {"effort": "high"}} in result["messages"]
+
+
+class _SentMessage(BaseModel):
+    model_config = ConfigDict(frozen=True)
+    role: str
+    content: list[dict[str, object]]
+
+
+_SENT_MESSAGES: Final = TypeAdapter(list[_SentMessage])
+
+
+@pytest.mark.usefixtures("local_model_cost_map", "local_beta_headers_config")
+def test_bedrock_messages_drops_tool_changes_from_a_system_turn_an_older_model_gets_as_a_user_turn() -> None:
+    """A model without mid-conversation system support sends no tool-changes beta, so the converted turn
+    must not carry the blocks that beta gates."""
+    from litellm.types.router import GenericLiteLLMParams
+
+    messages: Final = [
+        {"role": "user", "content": [{"type": "text", "text": "Hello"}]},
+        {"role": "assistant", "content": [{"type": "text", "text": "Hi"}]},
+        {
+            "role": "system",
+            "content": [
+                {"type": "text", "text": "A weather tool is now available"},
+                {"type": "tool_addition", "tool": {"type": "tool_reference", "name": "get_weather"}},
+            ],
+        },
+        {"role": "user", "content": [{"type": "text", "text": "Go on"}]},
+    ]
+    result: Final = AmazonAnthropicClaudeMessagesConfig().transform_anthropic_messages_request(
+        model="anthropic.claude-3-7-sonnet-20250219-v1:0",
+        messages=messages,
+        anthropic_messages_optional_request_params={"max_tokens": 4096},
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    sent: Final = _SENT_MESSAGES.validate_python(result["messages"])
+    blocks: Final = [
+        block for message in sent for block in message.content
+    ]  # comprehension-ok: flattening the sent content blocks
+    assert "mid-conversation-tool-changes-2026-07-01" not in result.get("anthropic_beta", [])
+    assert all(block["type"] != "tool_addition" for block in blocks)
+    assert {"type": "text", "text": "A weather tool is now available"} in blocks
+    assert all(message.role != "system" for message in sent)