fix: prioritize tool_calls over text when available_functions is None

[education-01]

Fixes #4788

Problem

When LLM returns both text and tool_calls simultaneously, and executor passes available_functions=None to get_llm_response, the current logic returns text instead of tool_calls. This causes native tool calls to be discarded.

Root Cause

In llm.py, the branch order was:

1. If (no tool_calls OR no available_functions) AND text_response → return text
2. If tool_calls AND no available_functions → return tool_calls

When both text and tool_calls exist with available_functions=None, branch 1 fires first, and branch 2 is never reached.

Fix

Reorder to prioritize tool_calls:

1. If tool_calls AND no available_functions → return tool_calls
2. If (no tool_calls OR no available_functions) AND text_response → return text

Changes

• lib/crewai/src/crewai/llm.py: Reorder return logic in both sync and async paths
• tests/test_llm.py: Add regression tests for this exact scenario

Testing

Added two unit tests:

• test_non_streaming_returns_tool_calls_when_text_and_tool_calls_exist_without_available_functions
• test_async_non_streaming_returns_tool_calls_when_text_and_tool_calls_exist_without_available_functions

Both verify that tool_calls are returned when available_functions=None, even when text content is present.

---

Note

Low Risk
Low risk logic reordering in LiteLLM non-streaming sync/async paths; behavior changes only when responses include both content and tool_calls and available_functions is None/empty.

Overview
Fixes LiteLLM non-streaming response handling so that when an LLM returns both content and tool_calls and available_functions is not provided, the LLM returns the raw tool_calls instead of discarding them by returning text.

Adds sync and async regression tests (mocking litellm.completion/litellm.acompletion) to assert LLM.call/LLM.acall prefer tool_calls over text in this scenario.

^{Written by Cursor Bugbot for commit 684de7e. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Missing LLMCallCompletedEvent when returning tool_calls with text

Low Severity

When both tool_calls and text_response are present with available_functions=None, the new early-return path at return tool_calls skips emitting LLMCallCompletedEvent. Previously this scenario took the text-response branch, which did emit the event. Since LLMCallStartedEvent is always emitted by the caller, this creates an unpaired start event. The event listener uses LLMCallCompletedEvent to call handle_llm_stream_completed(), which resets formatter state like _is_streaming and stops the live display — so skipping it could leave the console formatter in an inconsistent state.

Additional Locations (1)

• lib/crewai/src/crewai/llm.py#L1368-L1370