Merge pull request #42152 from BerriAI/litellm_claude_code_safeguards_passthrough

fix(anthropic): forward safeguards and anthropic-beta unchanged on native /v1/messages

litellm/llms/anthropic/experimental_pass_through/adapters/handler.py

@@ -35,7 +35,7 @@ if TYPE_CHECKING:
     from litellm.router import Router
 
 # Anthropic-only keys already mapped by the translator; strip on extra_kwargs re-merge.
-ANTHROPIC_ONLY_REQUEST_KEYS: Final[frozenset[str]] = frozenset({"output_config"})
+ANTHROPIC_ONLY_REQUEST_KEYS: Final[frozenset[str]] = frozenset({"output_config", "safeguards"})
 
 _AnthropicMessages: TypeAlias = "list[dict[str, object]]"
 _AnthropicSystem: TypeAlias = "str | list[dict[str, object]] | None"

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

@@ -79,10 +79,14 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             "speed",
             "output_config",
             "reasoning_effort",
+            "safeguards",
             # TODO: Add Anthropic `metadata` support
             # "metadata",
         ]
 
+    def should_filter_anthropic_beta_headers(self) -> bool:
+        return self._resolved_provider != "anthropic"
+
     def _remove_scope_from_cache_control(self, anthropic_messages_request: dict) -> None:
         """
         Remove `scope` field from cache_control blocks.

litellm/types/llms/anthropic.py

@@ -411,6 +411,7 @@ class AnthropicMessagesRequestOptionalParams(TypedDict, total=False):
     output_config: AnthropicOutputConfig | None  # Configuration for Claude's output behavior
     cache_control: dict[str, Any] | None  # Automatic prompt caching
     reasoning_effort: str | None
+    safeguards: ReadOnly[list[dict[str, object]] | None]
 
 
 class AnthropicMessagesRequest(AnthropicMessagesRequestOptionalParams, total=False):
@@ -530,6 +531,7 @@ class AnthropicStopDetails(TypedDict, total=False):
 class MessageDelta(TypedDict, total=False):
     stop_reason: str | None
     stop_details: ReadOnly[AnthropicStopDetails]
+    safeguard_results: ReadOnly[list[dict[str, object]]]
 
 
 class ServerToolUsage(TypedDict, total=False):
@@ -600,6 +602,7 @@ class MessageChunk(TypedDict, total=False):
     stop_reason: str | None
     stop_sequence: str | None
     usage: UsageDelta
+    safeguard_results: ReadOnly[list[dict[str, object]]]
 
 
 class MessageStartBlock(TypedDict):

litellm/types/llms/anthropic_messages/anthropic_response.py

@@ -97,3 +97,4 @@ class AnthropicMessagesResponse(TypedDict, total=False):
     type: Literal["message"] | None
     usage: AnthropicUsage | None
     context_management: NotRequired[ContextManagementResponse]
+    safeguard_results: NotRequired[ReadOnly[list[dict[str, object]]]]

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_output_config_passthrough.py

@@ -110,6 +110,19 @@ class TestOutputConfigStrippedFromCompletionKwargs:
             "reject it with 400 'Extra inputs are not permitted'"
         )
 
+    def test_safeguards_is_stripped_for_non_anthropic_target(self):
+        extra_kwargs = {
+            "custom_llm_provider": "azure",
+            "safeguards": [{"type": "dangerous_tool_use", "classifier_context": {"v": 1}}],
+        }
+
+        result = _call_prepare(extra_kwargs=extra_kwargs)
+
+        completion_kwargs = result[0] if isinstance(result, tuple) else result
+        assert "safeguards" not in completion_kwargs, (
+            "safeguards is an Anthropic-only field; OpenAI-format backends reject it with 400"
+        )
+
     def test_output_config_format_translated_to_response_format(self):
         """When ``output_config`` carries structured-output ``format``, the
         translator now maps it to OpenAI's ``response_format`` so non-Anthropic

tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py

@@ -1438,3 +1438,109 @@ async def test_anthropic_messages_leaves_non_provider_failures_unmapped():
         )
 
     assert "Traceback" not in str(excinfo.value)
+
+
+@pytest.mark.asyncio
+async def test_anthropic_messages_forwards_safeguards_and_unknown_beta_to_anthropic():
+    """Shapes are what Claude Code 2.1.278 sends and api.anthropic.com returns, captured 2026-09-21."""
+    from litellm.llms.anthropic.experimental_pass_through.messages import handler
+
+    safeguards = [{"type": "dangerous_tool_use", "classifier_context": {"v": 1, "permission_mode": "auto"}}]
+    client_betas = "dangerous-tool-use-2026-09-03,interleaved-thinking-2025-05-14"
+    safeguard_results = [{"type": "dangerous_tool_use", "status": {"type": "available", "tool_uses": {}}}]
+    captured: dict[str, object] = {}
+
+    def upstream_records_the_request(request: httpx.Request) -> httpx.Response:
+        captured["body"] = json.loads(request.content)
+        captured["anthropic-beta"] = request.headers.get("anthropic-beta")
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-haiku-4-5",
+                "content": [{"type": "text", "text": "ok"}],
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+                "safeguard_results": safeguard_results,
+            },
+            request=request,
+        )
+
+    upstream = AsyncHTTPHandler()
+    upstream.client = httpx.AsyncClient(transport=httpx.MockTransport(upstream_records_the_request))
+
+    response = await handler.anthropic_messages(
+        max_tokens=16,
+        messages=[{"role": "user", "content": "hi"}],
+        model="anthropic/claude-haiku-4-5",
+        custom_llm_provider="anthropic",
+        api_key="sk-test",
+        client=upstream,
+        safeguards=safeguards,
+        extra_headers={"anthropic-beta": client_betas},
+    )
+
+    assert captured["body"]["safeguards"] == safeguards
+    assert set(captured["anthropic-beta"].split(",")) == set(client_betas.split(","))
+    assert response["safeguard_results"] == safeguard_results
+
+
+@pytest.mark.asyncio
+async def test_anthropic_messages_streaming_forwards_safeguards_and_keeps_safeguard_results():
+    """Shapes are what Claude Code 2.1.278 sends and api.anthropic.com returns, captured 2026-09-21."""
+    from litellm.llms.anthropic.experimental_pass_through.messages import handler
+
+    safeguards = [{"type": "dangerous_tool_use", "classifier_context": {"v": 1, "permission_mode": "auto"}}]
+    tool_verdicts = {"toolu_01": {"type": "evaluated", "outcome": "not_flagged"}}
+    safeguard_results = [{"type": "dangerous_tool_use", "status": {"type": "available", "tool_uses": tool_verdicts}}]
+    captured: dict[str, object] = {}
+    message_start = {
+        "type": "message_start",
+        "message": {
+            "id": "msg_1",
+            "type": "message",
+            "role": "assistant",
+            "model": "claude-haiku-4-5",
+            "content": [],
+            "stop_reason": None,
+            "stop_sequence": None,
+            "usage": {"input_tokens": 1, "output_tokens": 0},
+            "safeguard_results": safeguard_results,
+        },
+    }
+    message_delta = {
+        "type": "message_delta",
+        "delta": {"stop_reason": "end_turn", "stop_sequence": None, "safeguard_results": safeguard_results},
+        "usage": {"output_tokens": 1},
+    }
+    sse = "".join(
+        f"event: {event['type']}\ndata: {json.dumps(event)}\n\n"
+        for event in (message_start, message_delta, {"type": "message_stop"})
+    )
+
+    def upstream_streams_safeguard_results(request: httpx.Request) -> httpx.Response:
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(200, headers={"content-type": "text/event-stream"}, content=sse.encode(), request=request)
+
+    upstream = AsyncHTTPHandler()
+    upstream.client = httpx.AsyncClient(transport=httpx.MockTransport(upstream_streams_safeguard_results))
+
+    stream = await handler.anthropic_messages(
+        max_tokens=16,
+        messages=[{"role": "user", "content": "hi"}],
+        model="anthropic/claude-haiku-4-5",
+        custom_llm_provider="anthropic",
+        api_key="sk-test",
+        client=upstream,
+        stream=True,
+        safeguards=safeguards,
+    )
+    raw = b"".join([chunk async for chunk in stream]).decode()
+    events = [json.loads(line[len("data: ") :]) for line in raw.splitlines() if line.startswith("data: ")]
+
+    assert captured["body"]["safeguards"] == safeguards
+    assert events[0]["message"]["safeguard_results"] == safeguard_results
+    assert [e for e in events if e["type"] == "message_delta"][0]["delta"]["safeguard_results"] == safeguard_results