fallback for json.loads tool function args

some openai-compatible models (such as openrouter's anthropic models)
do not default to an empty json object when there are no arguments,
leading to a parsing error.

`function.arguments == '{}'`:

```bash
% llm --tool llm_version "What is the version?" --td -m openrouter/openai/gpt-4o

Tool call: llm_version({})
  0.26

The version of the LLM is 0.26.
```

`function.arguments == ''`:

```bash
% llm --tool llm_version "What is the version?" --td -m openrouter/anthropic/claude-sonnet-4
I'll check the version of llm for you.Error: Expecting value: line 1 column 1 (char 0)
```

Author: thermokarst

llm/default_plugins/openai_models.py

@@ -690,6 +690,8 @@ def execute(self, prompt, stream, response, conversation=None, key=None):
                     usage = chunk.usage.model_dump()
                 if chunk.choices and chunk.choices[0].delta:
                     for tool_call in chunk.choices[0].delta.tool_calls or []:
+                        if tool_call.function.arguments is None:
+                            tool_call.function.arguments = ""
                         index = tool_call.index
                         if index not in tool_calls:
                             tool_calls[index] = tool_call
@@ -711,7 +713,7 @@ def execute(self, prompt, stream, response, conversation=None, key=None):
                         llm.ToolCall(
                             tool_call_id=value.id,
                             name=value.function.name,
-                            arguments=json.loads(value.function.arguments),
+                            arguments=json.loads(value.function.arguments or '{}'),
                         )
                     )
         else:
@@ -728,7 +730,7 @@ def execute(self, prompt, stream, response, conversation=None, key=None):
                     llm.ToolCall(
                         tool_call_id=tool_call.id,
                         name=tool_call.function.name,
-                        arguments=json.loads(tool_call.function.arguments),
+                        arguments=json.loads(tool_call.function.arguments or '{}'),
                     )
                 )
             if completion.choices[0].message.content is not None:
@@ -774,6 +776,8 @@ async def execute(
                     usage = chunk.usage.model_dump()
                 if chunk.choices and chunk.choices[0].delta:
                     for tool_call in chunk.choices[0].delta.tool_calls or []:
+                        if tool_call.function.arguments is None:
+                            tool_call.function.arguments = ""
                         index = tool_call.index
                         if index not in tool_calls:
                             tool_calls[index] = tool_call
@@ -794,7 +798,7 @@ async def execute(
                         llm.ToolCall(
                             tool_call_id=value.id,
                             name=value.function.name,
-                            arguments=json.loads(value.function.arguments),
+                            arguments=json.loads(value.function.arguments or '{}'),
                         )
                     )
             response.response_json = remove_dict_none_values(combine_chunks(chunks))
@@ -812,7 +816,7 @@ async def execute(
                     llm.ToolCall(
                         tool_call_id=tool_call.id,
                         name=tool_call.function.name,
-                        arguments=json.loads(tool_call.function.arguments),
+                        arguments=json.loads(tool_call.function.arguments or '{}'),
                     )
                 )
             if completion.choices[0].message.content is not None: