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: update llm parameter handling in human_feedback function (#4801)

João Mouraf070ce8abd6a
diff --git a/lib/crewai/src/crewai/flow/human_feedback.py b/lib/crewai/src/crewai/flow/human_feedback.py
index 4a191da99..096687d7a 100644
--- a/lib/crewai/src/crewai/flow/human_feedback.py
+++ b/lib/crewai/src/crewai/flow/human_feedback.py
@@ -408,7 +408,7 @@ def human_feedback(
emit=list(emit) if emit else None,
default_outcome=default_outcome,
metadata=metadata or {},
- llm=llm if isinstance(llm, str) else None,
+ llm=llm if isinstance(llm, str) else getattr(llm, "model", None),
)
# Determine effective provider:
diff --git a/lib/crewai/tests/test_async_human_feedback.py b/lib/crewai/tests/test_async_human_feedback.py
index 9bb3d0045..035f29dcc 100644
--- a/lib/crewai/tests/test_async_human_feedback.py
+++ b/lib/crewai/tests/test_async_human_feedback.py
@@ -971,6 +971,128 @@ class TestCollapseToOutcomeJsonParsing:
assert mock_llm.call.call_count == 2
+class TestLLMObjectPreservedInContext:
+ """Tests that BaseLLM objects have their model string preserved in PendingFeedbackContext."""
+
+ @patch("crewai.flow.flow.crewai_event_bus.emit")
+ def test_basellm_object_model_string_survives_roundtrip(self, mock_emit: MagicMock) -> None:
+ """Test that when llm is a BaseLLM object, its model string is stored in context
+ so that outcome collapsing works after async pause/resume.
+
+ This is the exact bug: locally the sync path keeps the LLM object in memory,
+ but in production the async path serializes the context and the LLM object was
+ discarded (stored as None), causing resume to skip classification and always
+ fall back to emit[0].
+ """
+ with tempfile.TemporaryDirectory() as tmpdir:
+ db_path = os.path.join(tmpdir, "test_flows.db")
+ persistence = SQLiteFlowPersistence(db_path)
+
+ # Create a mock BaseLLM object (not a string)
+ mock_llm_obj = MagicMock()
+ mock_llm_obj.model = "gemini/gemini-2.0-flash"
+
+ class PausingProvider:
+ def __init__(self, persistence: SQLiteFlowPersistence):
+ self.persistence = persistence
+ self.captured_context: PendingFeedbackContext | None = None
+
+ def request_feedback(
+ self, context: PendingFeedbackContext, flow: Flow
+ ) -> str:
+ self.captured_context = context
+ self.persistence.save_pending_feedback(
+ flow_uuid=context.flow_id,
+ context=context,
+ state_data=flow.state if isinstance(flow.state, dict) else flow.state.model_dump(),
+ )
+ raise HumanFeedbackPending(context=context)
+
+ provider = PausingProvider(persistence)
+
+ class TestFlow(Flow):
+ result_path: str = ""
+
+ @start()
+ @human_feedback(
+ message="Approve?",
+ emit=["needs_changes", "approved"],
+ llm=mock_llm_obj,
+ default_outcome="approved",
+ provider=provider,
+ )
+ def review(self):
+ return "content for review"
+
+ @listen("approved")
+ def handle_approved(self):
+ self.result_path = "approved"
+ return "Approved!"
+
+ @listen("needs_changes")
+ def handle_changes(self):
+ self.result_path = "needs_changes"
+ return "Changes needed"
+
+ # Phase 1: Start flow (should pause)
+ flow1 = TestFlow(persistence=persistence)
+ result = flow1.kickoff()
+ assert isinstance(result, HumanFeedbackPending)
+
+ # Verify the context stored the model STRING, not None
+ assert provider.captured_context is not None
+ assert provider.captured_context.llm == "gemini/gemini-2.0-flash"
+
+ # Verify it survives persistence roundtrip
+ flow_id = result.context.flow_id
+ loaded = persistence.load_pending_feedback(flow_id)
+ assert loaded is not None
+ _, loaded_context = loaded
+ assert loaded_context.llm == "gemini/gemini-2.0-flash"
+
+ # Phase 2: Resume with positive feedback - should use LLM to classify
+ flow2 = TestFlow.from_pending(flow_id, persistence)
+ assert flow2._pending_feedback_context is not None
+ assert flow2._pending_feedback_context.llm == "gemini/gemini-2.0-flash"
+
+ # Mock _collapse_to_outcome to verify it gets called (not skipped)
+ with patch.object(flow2, "_collapse_to_outcome", return_value="approved") as mock_collapse:
+ flow2.resume("this looks good, proceed!")
+
+ # The key assertion: _collapse_to_outcome was called (not skipped due to llm=None)
+ mock_collapse.assert_called_once_with(
+ feedback="this looks good, proceed!",
+ outcomes=["needs_changes", "approved"],
+ llm="gemini/gemini-2.0-flash",
+ )
+ assert flow2.last_human_feedback.outcome == "approved"
+ assert flow2.result_path == "approved"
+
+ def test_string_llm_still_works(self) -> None:
+ """Test that passing llm as a string still works correctly."""
+ context = PendingFeedbackContext(
+ flow_id="str-llm-test",
+ flow_class="test.Flow",
+ method_name="review",
+ method_output="output",
+ message="Review:",
+ emit=["approved", "rejected"],
+ llm="gpt-4o-mini",
+ )
+
+ serialized = context.to_dict()
+ restored = PendingFeedbackContext.from_dict(serialized)
+ assert restored.llm == "gpt-4o-mini"
+
+ 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
+
+ # 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
+
+
class TestAsyncHumanFeedbackEdgeCases:
"""Edge case tests for async human feedback."""