[Bugfix] accept redacted thinking blocks in Anthropic messages (#36992)

Signed-off-by: Benjamin Bartels <benjaminba@tiglab-ubuntu.ilab.local>
Signed-off-by: bbartels <benjamin@bartels.dev>
Co-authored-by: Benjamin Bartels <benjaminba@tiglab-ubuntu.ilab.local>

vllm/entrypoints/anthropic/protocol.py

@@ -34,7 +34,14 @@ class AnthropicUsage(BaseModel):
 class AnthropicContentBlock(BaseModel):
     """Content block in message"""
 
-    type: Literal["text", "image", "tool_use", "tool_result", "thinking"]
+    type: Literal[
+        "text",
+        "image",
+        "tool_use",
+        "tool_result",
+        "thinking",
+        "redacted_thinking",
+    ]
     text: str | None = None
     # For image content
     source: dict[str, Any] | None = None
@@ -48,6 +55,8 @@ class AnthropicContentBlock(BaseModel):
     # For thinking content
     thinking: str | None = None
     signature: str | None = None
+    # For redacted thinking content (safety-filtered by the API)
+    data: str | None = None
 
 
 class AnthropicMessage(BaseModel):

vllm/entrypoints/anthropic/serving.py

@@ -224,6 +224,12 @@ class AnthropicServingMessages(OpenAIServingChat):
             content_parts.append({"type": "image_url", "image_url": {"url": image_url}})
         elif block.type == "thinking" and block.thinking is not None:
             reasoning_parts.append(block.thinking)
+        elif block.type == "redacted_thinking":
+            # Redacted thinking blocks contain safety-filtered reasoning.
+            # We skip them as the content is opaque (base64 'data' field),
+            # but accepting the block prevents a validation error when the
+            # client echoes back the full assistant message.
+            pass
         elif block.type == "tool_use":
             cls._convert_tool_use_block(block, tool_calls)
         elif block.type == "tool_result":