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: append trailing user turn in native Gemini provider (#6973)

Havel Cyrus92eb5f91830c
diff --git a/lib/crewai/src/crewai/llm.py b/lib/crewai/src/crewai/llm.py
index c64445518..c2924e542 100644
--- a/lib/crewai/src/crewai/llm.py
+++ b/lib/crewai/src/crewai/llm.py
@@ -2391,6 +2391,14 @@ class LLM(BaseLLM):
):
return [*messages, {"role": "user", "content": ""}] # type: ignore[list-item]
+ # Handle Gemini models - they require the last message to not be 'assistant'
+ if (
+ ("gemini" in self.model.lower() or "google" in self.model.lower())
+ and messages
+ and messages[-1]["role"] == "assistant"
+ ):
+ return [*messages, {"role": "user", "content": "Please continue."}] # type: ignore[list-item]
+
if not self.is_anthropic:
return messages # type: ignore[return-value]
diff --git a/lib/crewai/src/crewai/llms/providers/gemini/completion.py b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
index 0bdf8ec0e..98edc1238 100644
--- a/lib/crewai/src/crewai/llms/providers/gemini/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
@@ -562,12 +562,19 @@ class GeminiCompletion(BaseLLM):
- System messages are separate system_instruction
- Content is organized as Content objects with Parts
- Roles are 'user' and 'model' (not 'assistant')
+ - History may not end on a model turn; a "Please continue." user turn
+ is appended when it does
Args:
messages: Input messages
Returns:
Tuple of (formatted_contents, system_instruction)
+
+ Raises:
+ ValueError: If the history ends on a model turn with an unresolved
+ function call, which requires a function response rather than a
+ continuation prompt.
"""
base_formatted = super()._format_messages(messages)
@@ -680,6 +687,23 @@ class GeminiCompletion(BaseLLM):
gemini_content = types.Content(role=gemini_role, parts=parts)
contents.append(gemini_content)
+ if contents and contents[-1].role == "model":
+ # Gemini's generateContent API rejects a request whose history ends
+ # on a model turn (agent loops can produce this, e.g. after
+ # max-iteration handling or a guardrail retry).
+ last_parts = contents[-1].parts or []
+ if any(part.function_call for part in last_parts):
+ raise ValueError(
+ "Gemini message history ends on an unresolved function call "
+ "-- a function response must be provided before calling the "
+ "model again."
+ )
+ contents.append(
+ types.Content(
+ role="user", parts=[types.Part.from_text(text="Please continue.")]
+ )
+ )
+
return contents, system_instruction
def _validate_and_emit_structured_output(
diff --git a/lib/crewai/tests/llms/google/test_google.py b/lib/crewai/tests/llms/google/test_google.py
index 2c03dcd23..48c9f3224 100644
--- a/lib/crewai/tests/llms/google/test_google.py
+++ b/lib/crewai/tests/llms/google/test_google.py
@@ -500,6 +500,66 @@ def test_gemini_message_formatting():
assert formatted_contents[1].role == "model"
+def test_gemini_message_formatting_appends_user_turn_after_trailing_model_turn():
+ """
+ Gemini's generateContent API rejects a request whose history ends on a
+ model turn ("Requests ending with a model turn are not supported"). Agent
+ loops can produce this (e.g. after max-iteration handling or a guardrail
+ retry), so formatting must append a synthetic user turn to recover.
+ """
+ llm = LLM(model="google/gemini-2.0-flash-001")
+
+ test_messages = [
+ {"role": "user", "content": "Hello"},
+ {"role": "assistant", "content": "Hi there!"},
+ ]
+
+ formatted_contents, _ = llm._format_messages_for_gemini(test_messages)
+
+ assert [content.role for content in formatted_contents] == [
+ "user",
+ "model",
+ "user",
+ ]
+ assert formatted_contents[0].parts[0].text == "Hello"
+ assert formatted_contents[1].parts[0].text == "Hi there!"
+ assert formatted_contents[2].parts[0].text == "Please continue."
+ assert test_messages == [
+ {"role": "user", "content": "Hello"},
+ {"role": "assistant", "content": "Hi there!"},
+ ]
+
+
+def test_gemini_message_formatting_leaves_user_terminated_history_unchanged():
+ llm = LLM(model="google/gemini-2.0-flash-001")
+
+ test_messages = [{"role": "user", "content": "Hello"}]
+
+ formatted_contents, _ = llm._format_messages_for_gemini(test_messages)
+
+ assert len(formatted_contents) == 1
+ assert formatted_contents[0].role == "user"
+ assert formatted_contents[0].parts[0].text == "Hello"
+
+
+def test_gemini_message_formatting_raises_on_unresolved_function_call():
+ llm = LLM(model="google/gemini-2.0-flash-001")
+
+ test_messages = [
+ {"role": "user", "content": "What's the weather?"},
+ {
+ "role": "assistant",
+ "content": "",
+ "tool_calls": [
+ {"function": {"name": "get_weather", "arguments": "{}"}}
+ ],
+ },
+ ]
+
+ with pytest.raises(ValueError, match="unresolved function call"):
+ llm._format_messages_for_gemini(test_messages)
+
+
def test_gemini_streaming_parameter():
"""
Test that streaming parameter is properly handled
diff --git a/lib/crewai/tests/test_llm.py b/lib/crewai/tests/test_llm.py
index 7f16cd933..ec15e8a3c 100644
--- a/lib/crewai/tests/test_llm.py
+++ b/lib/crewai/tests/test_llm.py
@@ -849,6 +849,33 @@ def test_ollama_does_not_modify_when_last_is_user(ollama_llm):
assert formatted == original_messages
+def gemini_litellm_llm():
+ return LLM(model="gemini/gemini-flash-latest", is_litellm=True)
+
+
+def test_gemini_litellm_appends_user_message_when_last_is_assistant(gemini_litellm_llm):
+ original_messages = [
+ {"role": "user", "content": "Hi there"},
+ {"role": "assistant", "content": "Hello!"},
+ ]
+
+ formatted = gemini_litellm_llm._format_messages_for_provider(original_messages)
+
+ assert len(formatted) == len(original_messages) + 1
+ assert formatted[-1] == {"role": "user", "content": "Please continue."}
+
+
+def test_gemini_litellm_does_not_modify_when_last_is_user(gemini_litellm_llm):
+ original_messages = [
+ {"role": "user", "content": "Tell me a joke."},
+ ]
+
+ formatted = gemini_litellm_llm._format_messages_for_provider(original_messages)
+
+ assert formatted == original_messages
+
+
def test_native_provider_raises_error_when_supported_but_fails():
"""Test that when a native provider is in SUPPORTED_NATIVE_PROVIDERS but fails to instantiate, we raise the error."""
with patch("crewai.llm.SUPPORTED_NATIVE_PROVIDERS", ["openai"]):