fix: add TOOL_CHOICE to ModelOptions like TEMPERATURE not a sentinel

The pass-thru behavior was not clear enough, so adding it to ModelOptions
where important options are known.  Most of these are sentinels which are
removed (because @@@) but this will be like TEMPERATURE which is passed
through to the backends.

No behavior change, but give a handly constant and a place to look for these.
This does not address all the other possible pass through args.

Signed-off-by: Mark Sturdevant <mark.sturdevant@ibm.com>

Author: markstur

cli/serve/app.py

@@ -115,14 +115,14 @@ def _build_model_options(request: ChatCompletionRequest) -> dict:
         "response_format",  # Response format (json_object) - not yet implemented
         "functions",  # Legacy function calling - not yet implemented
         "function_call",  # Legacy function calling - not yet implemented
-        # Tool choice is passed through as-is (not a ModelOption sentinel)
     }
     openai_to_model_option = {
         "temperature": ModelOption.TEMPERATURE,
         "max_tokens": ModelOption.MAX_NEW_TOKENS,
         "seed": ModelOption.SEED,
         "stream": ModelOption.STREAM,
         "tools": ModelOption.TOOLS,
+        "tool_choice": ModelOption.TOOL_CHOICE,
     }
 
     # Get all non-None fields

docs/examples/m_serve/m_serve_example_tool_calling.py

@@ -126,12 +126,13 @@ def serve(
     """Serve function that handles tool calling.
 
     This function demonstrates how to use tools with m serve. The tools
-    are passed via model_options and the model can request to call them.
+    are passed via model_options using ModelOption.TOOLS, and tool_choice
+    can be specified using ModelOption.TOOL_CHOICE.
 
     Args:
         input: List of chat messages
         requirements: Optional list of requirement strings
-        model_options: Model options including tools and tool_choice
+        model_options: Model options including ModelOption.TOOLS and ModelOption.TOOL_CHOICE
 
     Returns:
         ModelOutputThunk with potential tool calls
@@ -169,9 +170,10 @@ def serve(
     # Example usage (for testing purposes)
     test_messages = [ChatMessage(role="user", content="What's the weather in Paris?")]
 
-    # Simulate tool definitions being passed
+    # Simulate tool definitions being passed with tool_choice
     test_model_options = {
-        ModelOption.TOOLS: [weather_tool.as_json_tool, stock_price_tool.as_json_tool]
+        ModelOption.TOOLS: [weather_tool.as_json_tool, stock_price_tool.as_json_tool],
+        ModelOption.TOOL_CHOICE: "auto",  # Can be "none", "auto", or specific tool
     }
 
     response = serve(input=test_messages, model_options=test_model_options)

mellea/backends/model_options.py

@@ -22,6 +22,7 @@ class ModelOption:
 
     Attributes:
         TOOLS (str): Sentinel key for a list or dict of tools to expose for tool calling.
+        TOOL_CHOICE (str): Key for tool choice strategy (passed through to the backend).
         MAX_NEW_TOKENS (str): Sentinel key for the maximum number of new tokens to generate.
         SYSTEM_PROMPT (str): Sentinel key for the system prompt string.
         TEMPERATURE (str): Key for the sampling temperature (passed through to the backend).
@@ -34,6 +35,9 @@ class ModelOption:
     TOOLS = "@@@tools@@@"
     """Must be a list[Callable] or a dict[str, Callable] where str is the name of the function."""
 
+    TOOL_CHOICE = "tool_choice"
+    """Controls which tool the model should use. Can be "none", "auto", or a specific tool name."""
+
     MAX_NEW_TOKENS = "@@@max_new_tokens@@@"
     SYSTEM_PROMPT = "@@@system_prompt@@@"
     TEMPERATURE = "temperature"

test/cli/test_serve.py

@@ -505,9 +505,9 @@ async def test_tool_params_passed_to_model_options(self, mock_module):
 
         # Tools should be passed with ModelOption.TOOLS key
         assert ModelOption.TOOLS in model_options
-        # tool_choice should be passed through as-is
-        assert "tool_choice" in model_options
-        assert model_options["tool_choice"] == "auto"
+        # tool_choice should be passed through using ModelOption.TOOL_CHOICE
+        assert ModelOption.TOOL_CHOICE in model_options
+        assert model_options[ModelOption.TOOL_CHOICE] == "auto"
         # Legacy function calling parameters should still be excluded
         assert "functions" not in model_options
         assert "function_call" not in model_options

test/cli/test_serve_tool_calling.py

@@ -15,6 +15,7 @@
     FunctionParameters,
     ToolFunction,
 )
+from mellea.backends import ModelOption
 from mellea.core.base import AbstractMelleaTool, ModelOutputThunk, ModelToolCall
 
 
@@ -223,9 +224,9 @@ async def test_tool_choice_passed_to_model_options(
         assert call_args is not None
         model_options = call_args.kwargs["model_options"]
 
-        # tool_choice should be passed through as-is
-        assert "tool_choice" in model_options
-        assert model_options["tool_choice"] == "auto"
+        # tool_choice should be passed through using ModelOption.TOOL_CHOICE
+        assert ModelOption.TOOL_CHOICE in model_options
+        assert model_options[ModelOption.TOOL_CHOICE] == "auto"
 
     @pytest.mark.asyncio
     async def test_tool_calls_with_complex_arguments(