Fix #5510: Make crew chat description generation resilient to LLM failures

Co-Authored-By: João <joao@crewai.com>

lib/crewai/src/crewai/cli/crew_chat.py

@@ -25,6 +25,13 @@ from crewai.utilities.version import get_crewai_version
 
 MIN_REQUIRED_VERSION: Final[Literal["0.98.0"]] = "0.98.0"
 
+# Static fallbacks used when an LLM call fails while generating descriptions
+# for chat inputs or the crew itself. Returning a generic description is
+# preferable to crashing the process (see issue #5510), since these strings are
+# only surfaced in the CrewAI chat UI.
+DEFAULT_INPUT_DESCRIPTION: Final[str] = "Input value for the crew's tasks and agents."
+DEFAULT_CREW_DESCRIPTION: Final[str] = "A CrewAI crew."
+
 
 def check_conversational_crews_version(
     crewai_version: str, pyproject_data: dict[str, Any]
@@ -482,7 +489,22 @@ def generate_input_description_with_ai(
         "Context:\n"
         f"{context}"
     )
-    response = chat_llm.call(messages=[{"role": "user", "content": prompt}])
+    try:
+        response = chat_llm.call(messages=[{"role": "user", "content": prompt}])
+    except Exception as e:
+        # The LLM call can fail for many transient reasons (network, rate
+        # limits, provider outages, misconfigured credentials, etc.). This
+        # function is called at import time by downstream consumers such as
+        # `ag_ui_crewai.crews.ChatWithCrewFlow`, so letting the exception
+        # propagate would crash the containing process before any HTTP server
+        # has a chance to bind to its port. Fall back to a generic description
+        # rather than taking down the whole process (see issue #5510).
+        click.secho(
+            f"Warning: Failed to generate AI description for input '{input_name}' "
+            f"({type(e).__name__}: {e}). Falling back to a generic description.",
+            fg="yellow",
+        )
+        return DEFAULT_INPUT_DESCRIPTION
     return str(response).strip()
 
 
@@ -532,5 +554,16 @@ def generate_crew_description_with_ai(crew: Crew, chat_llm: LLM | BaseLLM) -> st
         "Context:\n"
         f"{context}"
     )
-    response = chat_llm.call(messages=[{"role": "user", "content": prompt}])
+    try:
+        response = chat_llm.call(messages=[{"role": "user", "content": prompt}])
+    except Exception as e:
+        # See comment in `generate_input_description_with_ai`: falling back to
+        # a generic description is preferable to crashing the process when the
+        # LLM provider is temporarily unavailable (see issue #5510).
+        click.secho(
+            f"Warning: Failed to generate AI description for crew "
+            f"({type(e).__name__}: {e}). Falling back to a generic description.",
+            fg="yellow",
+        )
+        return DEFAULT_CREW_DESCRIPTION
     return str(response).strip()

lib/crewai/tests/cli/test_crew_chat.py

@@ -0,0 +1,139 @@
+"""Tests for the crewai.cli.crew_chat description generators.
+
+These tests focus on the defensive behaviour introduced for issue #5510:
+``generate_input_description_with_ai`` and ``generate_crew_description_with_ai``
+must never propagate LLM call failures to their callers, since they are
+commonly invoked at container / module import time via downstream
+integrations such as ``ag_ui_crewai.crews.ChatWithCrewFlow``. A transient LLM
+provider hiccup should not crash the containing process before it has a chance
+to bind to its HTTP port.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+from crewai.agent import Agent
+from crewai.cli.crew_chat import (
+    DEFAULT_CREW_DESCRIPTION,
+    DEFAULT_INPUT_DESCRIPTION,
+    generate_crew_chat_inputs,
+    generate_crew_description_with_ai,
+    generate_input_description_with_ai,
+)
+from crewai.crew import Crew
+from crewai.task import Task
+
+
+def _make_crew_with_topic_input() -> Crew:
+    """Build a minimal Crew whose task/agent reference a ``{topic}`` input."""
+    agent = Agent(
+        role="Researcher on {topic}",
+        goal="Investigate the latest developments about {topic}",
+        backstory="An expert analyst focused on {topic}",
+        allow_delegation=False,
+    )
+    task = Task(
+        description="Write a short report about {topic}",
+        expected_output="A concise summary about {topic}",
+        agent=agent,
+    )
+    return Crew(agents=[agent], tasks=[task])
+
+
+def test_generate_input_description_returns_llm_response_on_success() -> None:
+    """Happy path: the LLM response is stripped and returned verbatim."""
+    crew = _make_crew_with_topic_input()
+    chat_llm = MagicMock()
+    chat_llm.call.return_value = "  The topic to research.  "
+
+    result = generate_input_description_with_ai("topic", crew, chat_llm)
+
+    assert result == "The topic to research."
+    chat_llm.call.assert_called_once()
+
+
+@pytest.mark.parametrize(
+    "exc",
+    [
+        ConnectionError("connection refused"),
+        TimeoutError("llm timed out"),
+        RuntimeError("litellm APIError: 500"),
+    ],
+)
+def test_generate_input_description_falls_back_on_llm_failure(exc: Exception) -> None:
+    """If the LLM call raises, we must return the static fallback instead of
+    propagating the exception. This is the core fix for issue #5510.
+    """
+    crew = _make_crew_with_topic_input()
+    chat_llm = MagicMock()
+    chat_llm.call.side_effect = exc
+
+    result = generate_input_description_with_ai("topic", crew, chat_llm)
+
+    assert result == DEFAULT_INPUT_DESCRIPTION
+
+
+def test_generate_input_description_still_raises_when_no_context() -> None:
+    """The fallback only applies to LLM call failures. When there is no
+    context at all for the given input, we still raise ``ValueError`` so that
+    callers can detect a truly malformed crew definition.
+    """
+    crew = _make_crew_with_topic_input()
+    chat_llm = MagicMock()
+
+    with pytest.raises(ValueError, match="No context found for input"):
+        generate_input_description_with_ai("does_not_exist", crew, chat_llm)
+
+    chat_llm.call.assert_not_called()
+
+
+def test_generate_crew_description_returns_llm_response_on_success() -> None:
+    crew = _make_crew_with_topic_input()
+    chat_llm = MagicMock()
+    chat_llm.call.return_value = "  Research topics and produce reports.  "
+
+    result = generate_crew_description_with_ai(crew, chat_llm)
+
+    assert result == "Research topics and produce reports."
+    chat_llm.call.assert_called_once()
+
+
+@pytest.mark.parametrize(
+    "exc",
+    [
+        ConnectionError("connection refused"),
+        TimeoutError("llm timed out"),
+        RuntimeError("litellm APIError: 500"),
+    ],
+)
+def test_generate_crew_description_falls_back_on_llm_failure(exc: Exception) -> None:
+    crew = _make_crew_with_topic_input()
+    chat_llm = MagicMock()
+    chat_llm.call.side_effect = exc
+
+    result = generate_crew_description_with_ai(crew, chat_llm)
+
+    assert result == DEFAULT_CREW_DESCRIPTION
+
+
+def test_generate_crew_chat_inputs_never_crashes_on_llm_failure() -> None:
+    """End-to-end: a crew with at least one required input placeholder and a
+    chat LLM whose ``.call`` always raises should still yield a valid
+    ``ChatInputs`` object populated with the static fallbacks, rather than
+    bubbling up the exception. This is the exact scenario described in
+    issue #5510 for ``ChatWithCrewFlow.__init__``.
+    """
+    crew = _make_crew_with_topic_input()
+    chat_llm = MagicMock()
+    chat_llm.call.side_effect = ConnectionError("transient outage")
+
+    chat_inputs = generate_crew_chat_inputs(crew, "MyCrew", chat_llm)
+
+    assert chat_inputs.crew_name == "MyCrew"
+    assert chat_inputs.crew_description == DEFAULT_CREW_DESCRIPTION
+    assert len(chat_inputs.inputs) == 1
+    assert chat_inputs.inputs[0].name == "topic"
+    assert chat_inputs.inputs[0].description == DEFAULT_INPUT_DESCRIPTION