fix: anthropic tool sanitation (#9310)

letta/llm_api/openai.py

@@ -524,7 +524,17 @@ def openai_chat_completions_request(
     log_event(name="llm_request_sent", attributes=data)
     chat_completion = client.chat.completions.create(**data)
     log_event(name="llm_response_received", attributes=chat_completion.model_dump())
-    return ChatCompletionResponse(**chat_completion.model_dump())
+    response = ChatCompletionResponse(**chat_completion.model_dump())
+
+    # Override tool_call IDs to ensure cross-provider compatibility (matches streaming path behavior)
+    # Some models (e.g. Kimi via OpenRouter) generate IDs like 'Read:93' which violate Anthropic's pattern
+    for choice in response.choices:
+        if choice.message.tool_calls:
+            for tool_call in choice.message.tool_calls:
+                if tool_call.id is not None:
+                    tool_call.id = get_tool_call_id()
+
+    return response
 
 
 def prepare_openai_payload(chat_completion_request: ChatCompletionRequest):

letta/schemas/message.py

@@ -64,7 +64,7 @@ from letta.schemas.letta_message_content import (
     get_letta_message_content_union_str_json_schema,
 )
 from letta.system import unpack_message
-from letta.utils import parse_json, validate_function_response
+from letta.utils import parse_json, sanitize_tool_call_id, validate_function_response
 
 
 def truncate_tool_return(content: Optional[str], limit: Optional[int]) -> Optional[str]:
@@ -1973,7 +1973,7 @@ class Message(BaseMessage):
                     content.append(
                         {
                             "type": "tool_use",
-                            "id": tool_call.id,
+                            "id": sanitize_tool_call_id(tool_call.id),
                             "name": tool_call.function.name,
                             "input": tool_call_input,
                         }
@@ -2017,7 +2017,7 @@ class Message(BaseMessage):
                     content.append(
                         {
                             "type": "tool_result",
-                            "tool_use_id": resolved_tool_call_id,
+                            "tool_use_id": sanitize_tool_call_id(resolved_tool_call_id),
                             "content": tool_result_content,
                         }
                     )

letta/utils.py

@@ -491,6 +491,25 @@ def get_tool_call_id() -> str:
     return str(uuid.uuid4())[:TOOL_CALL_ID_MAX_LEN]
 
 
+# Pattern for valid tool_call_id (required by Anthropic: ^[a-zA-Z0-9_-]+$)
+TOOL_CALL_ID_PATTERN = re.compile(r"^[a-zA-Z0-9_-]+$")
+
+
+def sanitize_tool_call_id(tool_id: str) -> str:
+    """Ensure tool_call_id matches cross-provider requirements:
+    - Anthropic: pattern ^[a-zA-Z0-9_-]+$
+    - OpenAI: max length 29 characters
+
+    Some models (e.g. Kimi via OpenRouter) generate IDs like 'Read:93' which
+    contain invalid characters. This sanitizes them for cross-provider compatibility.
+    """
+    # Replace invalid characters with underscores
+    if not TOOL_CALL_ID_PATTERN.match(tool_id):
+        tool_id = re.sub(r"[^a-zA-Z0-9_-]", "_", tool_id)
+    # Truncate to max length
+    return tool_id[:TOOL_CALL_ID_MAX_LEN]
+
+
 def assistant_function_to_tool(assistant_message: dict) -> dict:
     assert "function_call" in assistant_message
     new_msg = copy.deepcopy(assistant_message)