fix(azure_ai): strip provider_specific_fields recursively in tool_calls

The Anthropic /v1/messages adapter attaches the thought signature at
tool_calls[].function.provider_specific_fields
(adapters/transformation.py), not just at the top level of the assistant
message. The previous commit popped provider_specific_fields at the top
level only, so the nested copy inside tool_calls survived serialization
and strict upstreams (Fireworks-on-Azure) still rejected it with
Extra inputs are not permitted, field:
'messages[n].tool_calls[m].function.provider_specific_fields'.
Because Claude Code does multi-turn tool use, this is the next failure a
user hits after thinking_blocks is fixed.

Split the stripped fields into two sets: top-level output-only fields
(thinking_blocks, reasoning_content) removed with a plain pop, and fields
that can nest (cache_control, provider_specific_fields) removed
recursively via filter_value_from_dict so nested tool-call signatures are
caught too. reasoning_content is accepted on input by the Fireworks
backend but is dropped here as a defensive default for a generic
multi-backend provider; the direct fireworks_ai provider strips
thinking_blocks and the mistral provider (MistralConfig) strips both, and
this applies the union since Azure AI Foundry fronts multiple backends
with different strictness.

Added a regression test for the nested
tool_calls[].function.provider_specific_fields case; it fails against a
top-level-only pop and passes only with the recursive strip.

litellm/llms/azure_ai/chat/transformation.py

@@ -29,9 +29,28 @@ class AzureFoundryErrorStrings(str, enum.Enum):
     SET_EXTRA_PARAMETERS_TO_PASS_THROUGH = "Set extra-parameters to 'pass-through'"
 
 
-_OUTPUT_ONLY_MESSAGE_FIELDS: Tuple[str, ...] = (
+# Fields that only ever appear at the top level of a chat message and are
+# removed with a plain ``pop``. ``thinking_blocks`` and ``reasoning_content``
+# are output-only reasoning fields; the Fireworks backend accepts
+# ``reasoning_content`` on input, but the direct ``fireworks_ai`` provider
+# strips ``thinking_blocks`` and Azure AI Foundry fronts multiple backends, so
+# drop both as a defensive default for a generic multi-backend provider.
+_TOP_LEVEL_OUTPUT_ONLY_FIELDS: Tuple[str, ...] = (
     "thinking_blocks",
     "reasoning_content",
+)
+
+# Fields that can nest inside a message (e.g. inside ``tool_calls[].function``
+# or content blocks) and must be removed recursively. ``cache_control`` is an
+# Anthropic prompt-caching annotation; ``provider_specific_fields`` carries
+# Anthropic thought signatures and is attached by the Anthropic
+# ``/v1/messages`` adapter at ``tool_calls[].function.provider_specific_fields``
+# (``adapters/transformation.py``), so a top-level ``pop`` misses it and
+# strict upstreams still reject it with
+# ``Extra inputs are not permitted, field:
+# 'messages[n].tool_calls[m].function.provider_specific_fields'``.
+_RECURSIVELY_STRIPPED_FIELDS: Tuple[str, ...] = (
+    "cache_control",
     "provider_specific_fields",
 )
 
@@ -46,14 +65,15 @@ def _strip_unsupported_message_fields(message: AllMessageValues) -> None:
     ``/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.
+    'messages[n].<field>'``. Top-level output-only fields are popped; fields
+    that can nest (``cache_control``, and ``provider_specific_fields`` which
+    the adapter attaches inside ``tool_calls[].function``) are stripped
+    recursively so nested tool-call signatures are removed too.
     """
-    filter_value_from_dict(cast(dict, message), "cache_control")
     m = cast(dict, message)
-    for field in _OUTPUT_ONLY_MESSAGE_FIELDS:
+    for field in _RECURSIVELY_STRIPPED_FIELDS:
+        filter_value_from_dict(m, field)
+    for field in _TOP_LEVEL_OUTPUT_ONLY_FIELDS:
         m.pop(field, None)
 
 

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

@@ -309,6 +309,11 @@ def test_transform_request_drops_thinking_blocks_for_fireworks_model():
     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'``.
+    Also covers the nested ``tool_calls[].function.provider_specific_fields``
+    signature the Anthropic ``/v1/messages`` adapter attaches on tool-use
+    replays, which otherwise 400s with
+    ``Extra inputs are not permitted, field:
+    'messages[n].tool_calls[m].function.provider_specific_fields'``.
     """
     config = AzureAIStudioConfig()
     messages = [
@@ -324,6 +329,17 @@ def test_transform_request_drops_thinking_blocks_for_fireworks_model():
                     "cache_control": {},
                 }
             ],
+            "tool_calls": [
+                {
+                    "id": "toolu_1",
+                    "type": "function",
+                    "function": {
+                        "name": "read_file",
+                        "arguments": '{"path": "foo.txt"}',
+                        "provider_specific_fields": {"thought_signature": "sig"},
+                    },
+                }
+            ],
         },
         {"role": "user", "content": "go ahead"},
     ]
@@ -336,8 +352,56 @@ def test_transform_request_drops_thinking_blocks_for_fireworks_model():
         headers={},
     )
 
+    def _assert_no_unsupported_fields(node):
+        if isinstance(node, dict):
+            assert "thinking_blocks" not in node
+            assert "reasoning_content" not in node
+            assert "provider_specific_fields" not in node
+            assert "cache_control" not in node
+            for value in node.values():
+                _assert_no_unsupported_fields(value)
+        elif isinstance(node, list):
+            for item in node:
+                _assert_no_unsupported_fields(item)
+
     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
+        _assert_no_unsupported_fields(message)
+
+
+def test_transform_messages_strips_nested_tool_call_provider_specific_fields():
+    """Regression: the Anthropic ``/v1/messages`` adapter attaches the thought
+    signature at ``tool_calls[].function.provider_specific_fields``
+    (``adapters/transformation.py``). A top-level ``pop`` of
+    ``provider_specific_fields`` from the message misses the nested copy, so
+    strict upstreams (Fireworks-on-Azure) still 400 with
+    ``Extra inputs are not permitted, field:
+    'messages[1].tool_calls[0].function.provider_specific_fields'``.
+    ``_transform_messages`` must strip it recursively.
+    """
+    config = AzureAIStudioConfig()
+    messages = [
+        {"role": "user", "content": "Run the tool."},
+        {
+            "role": "assistant",
+            "content": "Calling the tool now.",
+            "tool_calls": [
+                {
+                    "id": "toolu_1",
+                    "type": "function",
+                    "function": {
+                        "name": "read_file",
+                        "arguments": '{"path": "foo.txt"}',
+                        "provider_specific_fields": {"thought_signature": "sig"},
+                    },
+                }
+            ],
+        },
+        {"role": "user", "content": "show me the result"},
+    ]
+
+    out = config._transform_messages(messages, model="glm-5.2")
+
+    tool_call = out[1]["tool_calls"][0]
+    assert "provider_specific_fields" not in tool_call
+    assert "provider_specific_fields" not in tool_call["function"]
+    assert tool_call["function"]["name"] == "read_file"