fix(anthropic): unwrap double-encoded tool_call arguments for Anthropic providers

When agent frameworks store/load tool call history they sometimes apply
json.dumps twice to function.arguments, producing a JSON string whose first
decode still returns a str. Anthropic tool_use.input must be an object;
the extra string layer causes Azure AI Anthropic to return:
messages.x.content.y.tool_use.input: Input should be an object

parse_tool_call_arguments now detects this case: when json.loads() returns a
str it attempts one additional decode and logs a warning naming the tool so
callers can trace the source of the double-encoding.

Fixes #42739

Signed-off-by: Aftabbs <aftabbs.wwe@gmail.com>

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -1,4 +1,4 @@
-"""
+"""
 Common utility functions used for translating messages across providers
 """
 
@@ -2382,7 +2382,24 @@ def parse_tool_call_arguments(
         return {}
 
     try:
-        return json.loads(arguments)
+        parsed = json.loads(arguments)
+        if isinstance(parsed, str):
+            # Double-encoded JSON: json.dumps applied twice by the caller (common
+            # in agent frameworks that serialise tool arguments through a store).
+            # Anthropic's tool_use.input must be an object, not a string, so
+            # unwrap one extra layer and warn so callers can fix the root cause.
+            verbose_logger.warning(
+                "Tool call arguments for tool '%s' (%s) decoded to a string "
+                "instead of an object — double-encoded JSON detected. "
+                "Attempting second decode.",
+                tool_name or "<unknown>",
+                context or "unknown context",
+            )
+            try:
+                parsed = json.loads(parsed)
+            except json.JSONDecodeError:
+                pass
+        return parsed
     except json.JSONDecodeError as original_error:
         repaired: Final = _attempt_json_repair(arguments)
         if repaired is not None: