Merge branch 'main' into gl/feat/agent-skills

lib/crewai/src/crewai/flow/flow.py

@@ -3086,25 +3086,35 @@ class Flow(Generic[T], metaclass=FlowMeta):
             logger.warning(
                 f"Structured output failed, falling back to simple prompting: {e}"
             )
-            response = llm_instance.call(messages=prompt)
-            response_clean = str(response).strip()
+            try:
+                response = llm_instance.call(
+                    messages=[{"role": "user", "content": prompt}],
+                )
+                response_clean = str(response).strip()
 
-            # Exact match (case-insensitive)
-            for outcome in outcomes:
-                if outcome.lower() == response_clean.lower():
-                    return outcome
+                # Exact match (case-insensitive)
+                for outcome in outcomes:
+                    if outcome.lower() == response_clean.lower():
+                        return outcome
 
-            # Partial match
-            for outcome in outcomes:
-                if outcome.lower() in response_clean.lower():
-                    return outcome
+                # Partial match
+                for outcome in outcomes:
+                    if outcome.lower() in response_clean.lower():
+                        return outcome
 
-            # Fallback to first outcome
-            logger.warning(
-                f"Could not match LLM response '{response_clean}' to outcomes {list(outcomes)}. "
-                f"Falling back to first outcome: {outcomes[0]}"
-            )
-            return outcomes[0]
+                # Fallback to first outcome
+                logger.warning(
+                    f"Could not match LLM response '{response_clean}' to outcomes {list(outcomes)}. "
+                    f"Falling back to first outcome: {outcomes[0]}"
+                )
+                return outcomes[0]
+
+            except Exception as fallback_err:
+                logger.warning(
+                    f"Simple prompting also failed: {fallback_err}. "
+                    f"Falling back to first outcome: {outcomes[0]}"
+                )
+                return outcomes[0]
 
     def _log_flow_event(
         self,

lib/crewai/src/crewai/flow/human_feedback.py

@@ -76,6 +76,24 @@ if TYPE_CHECKING:
 F = TypeVar("F", bound=Callable[..., Any])
 
 
+def _serialize_llm_for_context(llm: Any) -> str | None:
+    """Serialize a BaseLLM object to a model string with provider prefix.
+
+    When persisting the LLM for HITL resume, we need to store enough info
+    to reconstruct a working LLM on the resume worker. Just storing the bare
+    model name (e.g. "gemini-3-flash-preview") causes provider inference to
+    fail — it defaults to OpenAI. Including the provider prefix (e.g.
+    "gemini/gemini-3-flash-preview") allows LLM() to correctly route.
+    """
+    model = getattr(llm, "model", None)
+    if not model:
+        return None
+    provider = getattr(llm, "provider", None)
+    if provider and "/" not in model:
+        return f"{provider}/{model}"
+    return model
+
+
 @dataclass
 class HumanFeedbackResult:
     """Result from a @human_feedback decorated method.
@@ -412,7 +430,7 @@ def human_feedback(
                 emit=list(emit) if emit else None,
                 default_outcome=default_outcome,
                 metadata=metadata or {},
-                llm=llm if isinstance(llm, str) else getattr(llm, "model", None),
+                llm=llm if isinstance(llm, str) else _serialize_llm_for_context(llm),
             )
 
             # Determine effective provider:

lib/crewai/src/crewai/llms/constants.py

@@ -240,6 +240,7 @@ ANTHROPIC_MODELS: list[AnthropicModels] = [
 
 GeminiModels: TypeAlias = Literal[
     "gemini-3-pro-preview",
+    "gemini-3-flash-preview",
     "gemini-2.5-pro",
     "gemini-2.5-pro-preview-03-25",
     "gemini-2.5-pro-preview-05-06",
@@ -294,6 +295,7 @@ GeminiModels: TypeAlias = Literal[
 ]
 GEMINI_MODELS: list[GeminiModels] = [
     "gemini-3-pro-preview",
+    "gemini-3-flash-preview",
     "gemini-2.5-pro",
     "gemini-2.5-pro-preview-03-25",
     "gemini-2.5-pro-preview-05-06",

lib/crewai/tests/test_async_human_feedback.py

@@ -989,8 +989,10 @@ class TestLLMObjectPreservedInContext:
             persistence = SQLiteFlowPersistence(db_path)
 
             # Create a mock BaseLLM object (not a string)
+            # Simulates LLM(model="gemini-2.0-flash", provider="gemini")
             mock_llm_obj = MagicMock()
-            mock_llm_obj.model = "gemini/gemini-2.0-flash"
+            mock_llm_obj.model = "gemini-2.0-flash"
+            mock_llm_obj.provider = "gemini"
 
             class PausingProvider:
                 def __init__(self, persistence: SQLiteFlowPersistence):
@@ -1086,11 +1088,36 @@ class TestLLMObjectPreservedInContext:
 
     def test_none_llm_when_no_model_attr(self) -> None:
         """Test that llm is None when object has no model attribute."""
-        mock_obj = MagicMock(spec=[])  # No attributes
+        from crewai.flow.human_feedback import _serialize_llm_for_context
 
-        # Simulate what the decorator does
-        llm_value = mock_obj if isinstance(mock_obj, str) else getattr(mock_obj, "model", None)
-        assert llm_value is None
+        mock_obj = MagicMock(spec=[])  # No attributes
+        assert _serialize_llm_for_context(mock_obj) is None
+
+    def test_provider_prefix_added_to_bare_model(self) -> None:
+        """Test that provider prefix is added when model has no slash."""
+        from crewai.flow.human_feedback import _serialize_llm_for_context
+
+        mock_obj = MagicMock()
+        mock_obj.model = "gemini-3-flash-preview"
+        mock_obj.provider = "gemini"
+        assert _serialize_llm_for_context(mock_obj) == "gemini/gemini-3-flash-preview"
+
+    def test_provider_prefix_not_doubled_when_already_present(self) -> None:
+        """Test that provider prefix is not added when model already has a slash."""
+        from crewai.flow.human_feedback import _serialize_llm_for_context
+
+        mock_obj = MagicMock()
+        mock_obj.model = "gemini/gemini-2.0-flash"
+        mock_obj.provider = "gemini"
+        assert _serialize_llm_for_context(mock_obj) == "gemini/gemini-2.0-flash"
+
+    def test_no_provider_attr_falls_back_to_bare_model(self) -> None:
+        """Test that bare model is used when no provider attribute exists."""
+        from crewai.flow.human_feedback import _serialize_llm_for_context
+
+        mock_obj = MagicMock(spec=[])
+        mock_obj.model = "gpt-4o-mini"
+        assert _serialize_llm_for_context(mock_obj) == "gpt-4o-mini"
 
 
 class TestAsyncHumanFeedbackEdgeCases:

lib/crewai/tests/test_human_feedback_decorator.py

@@ -400,6 +400,45 @@ class TestCollapseToOutcome:
 
         assert result == "approved"  # First in list
 
+    def test_both_llm_calls_fail_returns_first_outcome(self):
+        """When both structured and simple prompting fail, return outcomes[0]."""
+        flow = Flow()
+
+        with patch("crewai.llm.LLM") as MockLLM:
+            mock_llm = MagicMock()
+            # Both calls raise — simulates wrong provider / auth failure
+            mock_llm.call.side_effect = RuntimeError("Model not found")
+            MockLLM.return_value = mock_llm
+
+            result = flow._collapse_to_outcome(
+                feedback="looks great, approve it",
+                outcomes=["needs_changes", "approved"],
+                llm="gemini-3-flash-preview",
+            )
+
+        assert result == "needs_changes"  # First in list (safe fallback)
+
+    def test_structured_fails_but_simple_succeeds(self):
+        """When structured output fails but simple prompting works, use that."""
+        flow = Flow()
+
+        with patch("crewai.llm.LLM") as MockLLM:
+            mock_llm = MagicMock()
+            # First call (structured) fails, second call (simple) succeeds
+            mock_llm.call.side_effect = [
+                RuntimeError("Function calling not supported"),
+                "approved",
+            ]
+            MockLLM.return_value = mock_llm
+
+            result = flow._collapse_to_outcome(
+                feedback="looks great",
+                outcomes=["needs_changes", "approved"],
+                llm="gpt-4o-mini",
+            )
+
+        assert result == "approved"
+
 
 # -- HITL Learning tests --