[BUG] LiteLLM path drops token usage for tool-call turns and structured output (response_model) #8679

🤖 [agent on alexmarchenko-crewai's behalf]

Description

When an LLM uses the LiteLLM fallback path, two kinds of call never report token usage. That covers LLM(model=..., is_litellm=True) and any model prefix that isn't a native provider, for example litellm_proxy/....

1. Native tool-call turns. These are calls with tools but no available_functions, which is how the agent executor makes them. They return the tool calls without emitting LLMCallCompletedEvent. This happens in sync and async handlers, streaming and non-streaming.
2. Structured output through Instructor (response_model, e.g. a task with output_pydantic). LLMCallCompletedEvent is emitted with usage=None, and the tokens are never tracked.

Tracing still records these calls, so traces show requests with 0 tokens and $0 cost. The native OpenAI, Azure and Anthropic providers report usage for both cases.

Steps to Reproduce

1. Create LLM(model="anthropic/claude-sonnet-4-5", is_litellm=True). Any LiteLLM-routed model shows the same behavior.
2. Run a crew with a task that has output_pydantic=SomeModel, and a crew whose agent has one tool.
3. Listen on crewai_event_bus for LLMCallStartedEvent and LLMCallCompletedEvent, and inspect event.usage.

Expected behavior

Every LLM call emits LLMCallCompletedEvent with usage populated, the same as on the native providers:

• Tool-call turns should emit a tool_call completion.
• Structured-output calls should carry the completion's token usage.

Screenshots/Code snippets

Observed on 1.15.29 with the LiteLLM path:

Operating System

macOS

Python Version

3.12

crewAI Version

1.15.29 (also present on main / 1.15.30)

crewAI Tools Version

N/A

Virtual Environment

uv

Evidence

The drop points are in lib/crewai/src/crewai/llm.py:

• if tool_calls and not available_functions: return tool_calls in the four response handlers.
• The response_model and self.is_litellm branch, which emits usage=None.

Possible Solution

• Emit a TOOL_CALL completion with usage before returning tool calls, as the native providers do.
• For structured output, get the raw completion from Instructor's create_with_completion and report its usage.

Implemented in #8678.

Additional context

None.