Skip to content

πŸ™Œ OpenHands: AI-Driven Development

Fix issue #9655: [Bug]: CodeActAgent is incompatible with xAI Grok-4 due to hardcoded stop parameter (#9666)

OpenHands8937b3fbfc83
diff --git a/openhands/llm/llm.py b/openhands/llm/llm.py
index 4463241dcc..13786739d8 100644
--- a/openhands/llm/llm.py
+++ b/openhands/llm/llm.py
@@ -104,6 +104,7 @@ MODELS_WITHOUT_STOP_WORDS = [
'o1-preview',
'o1',
'o1-2024-12-17',
+ 'xai/grok-4-0709',
]
diff --git a/tests/unit/test_llm.py b/tests/unit/test_llm.py
index 8b5e106ce3..c5a1c7f48a 100644
--- a/tests/unit/test_llm.py
+++ b/tests/unit/test_llm.py
@@ -264,6 +264,45 @@ def test_llm_init_with_openrouter_model(mock_get_model_info, default_config):
mock_get_model_info.assert_called_once_with('openrouter:gpt-4o-mini')
+@patch('openhands.llm.llm.litellm_completion')
+def test_stop_parameter_handling(mock_litellm_completion, default_config):
+ """Test that stop parameter is only added for supported models."""
+ from litellm.types.utils import ModelResponse
+
+ mock_response = ModelResponse(
+ id='test-id',
+ choices=[{'message': {'content': 'Test response'}}],
+ model='test-model',
+ )
+ mock_litellm_completion.return_value = mock_response
+
+ # Test with a model that supports stop parameter
+ default_config.model = (
+ 'custom-model' # Use a model not in FUNCTION_CALLING_SUPPORTED_MODELS
+ )
+ llm = LLM(default_config)
+ llm.completion(
+ messages=[{'role': 'user', 'content': 'Hello!'}],
+ tools=[
+ {'type': 'function', 'function': {'name': 'test', 'description': 'test'}}
+ ],
+ )
+ # Verify stop parameter was included
+ assert 'stop' in mock_litellm_completion.call_args[1]
+
+ # Test with Grok-4 model that doesn't support stop parameter
+ default_config.model = 'xai/grok-4-0709'
+ llm = LLM(default_config)
+ llm.completion(
+ messages=[{'role': 'user', 'content': 'Hello!'}],
+ tools=[
+ {'type': 'function', 'function': {'name': 'test', 'description': 'test'}}
+ ],
+ )
+ # Verify stop parameter was not included
+ assert 'stop' not in mock_litellm_completion.call_args[1]
+
+
# Tests involving completion and retries