fix(azure_ai): strip thinking_blocks and other output-only fields from chat messages

AzureAIStudioConfig._transform_messages only converted content lists to
strings and passed image/audio blocks through. It did not drop the
output-only / Anthropic-only fields that LiteLLM's Anthropic /v1/messages
adapter attaches to the OpenAI-format assistant messages it builds when
replaying multi-turn history from an Anthropic-format client such as
Claude Code. Those fields (thinking_blocks, reasoning_content,
provider_specific_fields, and the Anthropic cache_control annotation) are
not part of the OpenAI chat-completions message schema, so they reached
Azure AI Foundry-hosted backends whose chat-message schema sets
additionalProperties: false (for example Fireworks GLM) and triggered
400 Extra inputs are not permitted, field: 'messages[n].thinking_blocks'.

Strip those fields from each message before the request is serialized,
mirroring what the direct fireworks_ai provider
(FireworksAIConfig._transform_messages_helper) and the mistral provider
(MistralChatConfig._strip_output_only_fields) already do. The strip
reuses the existing filter_value_from_dict helper for cache_control and
pops the flat output-only fields directly. Regression tests cover both
_transform_messages and the full transform_request path that the reported
end-to-end scenario hits.

litellm/llms/azure_ai/chat/transformation.py

@@ -11,6 +11,7 @@ from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     _audio_or_image_in_message_content,
     convert_content_list_to_str,
+    filter_value_from_dict,
 )
 from litellm.llms.azure.common_utils import BaseAzureLLM
 from litellm.llms.base_llm.chat.transformation import LiteLLMLoggingObj
@@ -28,6 +29,34 @@ class AzureFoundryErrorStrings(str, enum.Enum):
     SET_EXTRA_PARAMETERS_TO_PASS_THROUGH = "Set extra-parameters to 'pass-through'"
 
 
+_OUTPUT_ONLY_MESSAGE_FIELDS: Tuple[str, ...] = (
+    "thinking_blocks",
+    "reasoning_content",
+    "provider_specific_fields",
+)
+
+
+def _strip_unsupported_message_fields(message: AllMessageValues) -> None:
+    """Remove input-forbidden fields from a chat message before it is sent to
+    an Azure AI Foundry hosted model.
+
+    Azure AI Foundry fronts many OpenAI-compatible backends (Fireworks,
+    Mistral, xAI, ...) whose chat-message schemas set ``additionalProperties:
+    false``. LiteLLM-internal / output-only fields that the Anthropic
+    ``/v1/messages`` adapter attaches when replaying assistant turns from an
+    Anthropic-format client (e.g. Claude Code) are not part of that schema and
+    trigger ``400 Extra inputs are not permitted, field:
+    'messages[n].<field>'``. ``cache_control`` is an Anthropic prompt-caching
+    annotation with the same problem. ``thinking_blocks`` and
+    ``reasoning_content`` are output-only; ``provider_specific_fields`` carries
+    Anthropic thought signatures. None are valid OpenAI input, so drop them.
+    """
+    filter_value_from_dict(cast(dict, message), "cache_control")
+    m = cast(dict, message)
+    for field in _OUTPUT_ONLY_MESSAGE_FIELDS:
+        m.pop(field, None)
+
+
 class AzureAIStudioConfig(OpenAIConfig):
     def get_supported_openai_params(self, model: str) -> List:
         model_supports_tool_choice = True  # azure ai supports this by default
@@ -171,6 +200,7 @@ class AzureAIStudioConfig(OpenAIConfig):
             2. If message contains an image or audio, send as is (user-intended)
         """
         for message in messages:
+            _strip_unsupported_message_fields(message)
             # Do nothing if the message contains an image or audio
             if _audio_or_image_in_message_content(message):
                 continue

tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py

@@ -262,3 +262,82 @@ def test_drop_tool_level_extra_fields_strips_copilot_mcp_server_name():
         assert "copilot_mcp_server_name" not in tool
     assert result["tools"][0]["type"] == "function"
     assert result["tools"][1]["function"]["name"] == "read_file"
+
+
+def test_transform_messages_strips_anthropic_replay_fields():
+    """Regression: Anthropic /v1/messages adapter attaches ``thinking_blocks``
+    (and ``provider_specific_fields`` / ``cache_control``) to the OpenAI-format
+    assistant messages it builds when replaying Claude Code turns. Azure AI
+    Foundry hosted model backends with strict chat-message schemas (e.g.
+    Fireworks GLM) reject them with
+    ``Extra inputs are not permitted, field: 'messages[n].thinking_blocks'``.
+    ``_transform_messages`` must drop them before the request is sent.
+    """
+    config = AzureAIStudioConfig()
+    messages = [
+        {"role": "user", "content": "Read a file."},
+        {
+            "role": "assistant",
+            "content": "I can help.",
+            "thinking_blocks": [
+                {
+                    "type": "thinking",
+                    "thinking": "The user wants me to read a file.",
+                    "signature": "",
+                    "cache_control": {},
+                }
+            ],
+            "reasoning_content": "internal reasoning",
+            "provider_specific_fields": {"thought_signature": "sig"},
+            "cache_control": {"type": "ephemeral"},
+        },
+        {"role": "user", "content": "go ahead"},
+    ]
+
+    out = config._transform_messages(messages, model="glm-5.2")
+
+    assert "thinking_blocks" not in out[1]
+    assert "reasoning_content" not in out[1]
+    assert "provider_specific_fields" not in out[1]
+    assert "cache_control" not in out[1]
+    assert out[1]["content"] == "I can help."
+
+
+def test_transform_request_drops_thinking_blocks_for_fireworks_model():
+    """End-to-end of the reported scenario: a Fireworks model deployed on Azure
+    AI Foundry receives a replayed assistant turn carrying ``thinking_blocks``.
+    The outgoing request payload built by ``transform_request`` must not carry
+    it, or the Fireworks backend 400s with
+    ``Extra inputs are not permitted, field: 'messages[2].thinking_blocks'``.
+    """
+    config = AzureAIStudioConfig()
+    messages = [
+        {"role": "user", "content": "Read a file."},
+        {
+            "role": "assistant",
+            "content": "I can help.",
+            "thinking_blocks": [
+                {
+                    "type": "thinking",
+                    "thinking": "The user wants me to read a file.",
+                    "signature": "",
+                    "cache_control": {},
+                }
+            ],
+        },
+        {"role": "user", "content": "go ahead"},
+    ]
+
+    payload = config.transform_request(
+        model="glm-5.2",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+
+    for message in payload["messages"]:
+        assert "thinking_blocks" not in message
+        assert "reasoning_content" not in message
+        assert "provider_specific_fields" not in message
+        assert "cache_control" not in message