Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

fix(llm): preserve tool_calls when response also contains text

Greyson LaLonde9f132350379b
diff --git a/lib/crewai/src/crewai/llm.py b/lib/crewai/src/crewai/llm.py
index 57b00b408..1c03eddc9 100644
--- a/lib/crewai/src/crewai/llm.py
+++ b/lib/crewai/src/crewai/llm.py
@@ -1235,8 +1235,12 @@ class LLM(BaseLLM):
# --- 4) Check for tool calls
tool_calls = response_message.tool_calls or []
- # --- 5) If no tool calls or no available functions, return the text response directly as long as there is a text response
- if (not tool_calls or not available_functions) and text_response:
+ # --- 5) If there are tool calls but no available functions, return the tool calls
+ if tool_calls and not available_functions:
+ return tool_calls
+
+ # --- 6) If there are no tool calls to execute, return the text response directly
+ if not tool_calls and text_response:
self._handle_emit_call_events(
response=text_response,
call_type=LLMCallType.LLM_CALL,
@@ -1247,11 +1251,6 @@ class LLM(BaseLLM):
)
return text_response
- # --- 6) If there are tool calls but no available functions, return the tool calls
- # This allows the caller (e.g., executor) to handle tool execution
- if tool_calls and not available_functions:
- return tool_calls
-
# --- 7) Handle tool calls if present (execute when available_functions provided)
if tool_calls and available_functions:
tool_result = self._handle_tool_call(
@@ -1384,7 +1383,10 @@ class LLM(BaseLLM):
tool_calls = response_message.tool_calls or []
- if (not tool_calls or not available_functions) and text_response:
+ if tool_calls and not available_functions:
+ return tool_calls
+
+ if not tool_calls and text_response:
self._handle_emit_call_events(
response=text_response,
call_type=LLMCallType.LLM_CALL,
@@ -1395,11 +1397,6 @@ class LLM(BaseLLM):
)
return text_response
- # If there are tool calls but no available functions, return the tool calls
- # This allows the caller (e.g., executor) to handle tool execution
- if tool_calls and not available_functions:
- return tool_calls
-
# Handle tool calls if present (execute when available_functions provided)
if tool_calls and available_functions:
tool_result = self._handle_tool_call(
diff --git a/lib/crewai/tests/test_llm.py b/lib/crewai/tests/test_llm.py
index dff40bdc6..b4002b8e9 100644
--- a/lib/crewai/tests/test_llm.py
+++ b/lib/crewai/tests/test_llm.py
@@ -177,6 +177,7 @@ def test_llm_passes_additional_params():
# Create mocks for response structure
mock_message = MagicMock()
mock_message.content = "Test response"
+ mock_message.tool_calls = None
mock_choice = MagicMock()
mock_choice.message = mock_message
mock_response = MagicMock()
@@ -1146,3 +1147,52 @@ async def test_usage_info_streaming_with_acall():
assert llm._token_usage["total_tokens"] > 0
assert len(result) > 0
+
+
+def _build_response_with_text_and_tool_calls():
+ """Mimic a litellm ModelResponse that contains both content and tool_calls."""
+ from litellm.types.utils import ChatCompletionMessageToolCall, Function
+
+ response_message = MagicMock()
+ response_message.content = "I will search for the given query."
+ response_message.tool_calls = [
+ ChatCompletionMessageToolCall(
+ id="call_123",
+ type="function",
+ function=Function(name="search", arguments='{"q": "x"}'),
+ )
+ ]
+ choice = MagicMock(message=response_message)
+ response = MagicMock(choices=[choice], model_extra=None)
+ return response
+
+
+def test_non_streaming_returns_tool_calls_when_text_also_present():
+ """A response with both text and tool_calls must not drop the tool_calls
+ when available_functions is None (executor-managed tool execution path).
+ """
+ llm = LLM(model="gpt-4o-mini", is_litellm=True)
+ response = _build_response_with_text_and_tool_calls()
+
+ with patch("crewai.llm.litellm.completion", return_value=response):
+ result = llm.call("anything", available_functions=None)
+
+ assert isinstance(result, list)
+ assert len(result) == 1
+ assert result[0].function.name == "search"
+
+
+async def test_non_streaming_async_returns_tool_calls_when_text_also_present():
+ llm = LLM(model="openai/gpt-4o-mini", is_litellm=True, stream=False)
+ response = _build_response_with_text_and_tool_calls()
+
+ async def _ret(*args, **kwargs):
+ return response
+
+ with patch("crewai.llm.litellm.acompletion", side_effect=_ret):
+ result = await llm.acall("anything", available_functions=None)
+
+ assert isinstance(result, list)
+ assert len(result) == 1
+ assert result[0].function.name == "search"