[bug] _llm_input_messages fails to parse ChatMessage objects in smolagents >= 1.x, resulting in empty LLM input spans

[zhihuat]

Describe the bug
Since smolagents changed its internal message format from plain dict to ChatMessage objects, _llm_input_messages in _wrappers.py silently drops all messages and LLM input spans appear empty in the trace.

Root Cause
In _wrappers.py, _llm_input_messages checks:

for i, message in enumerate(messages):
    if not isinstance(message, dict):
        continue   # ← ChatMessage objects are skipped here
smolagents generate() now passes a list[ChatMessage] instead of list[dict]. Because ChatMessage is not a dict, every message is skipped.

smolagents generate() now passes a list[ChatMessage] instead of list[dict]. Because ChatMessage is not a dict, every message is skipped.

Environment
openinference-instrumentation-smolagents: 0.1.25
smolagents: 1.24.0

Expected Behavior
llm.input_messages.* span attributes are populated with the full conversation history.

Actual Behavior
llm.input_messages.* attributes are empty. input.value falls back to str(messages), producing Python object repr strings like:

ChatMessage(role=<MessageRole.USER: 'user'>, content=[{'type': 'text', 'text': '...'}])

instead of JSON.

[dosubot]

Hi @zhihuat! I'm Dosu and I’m helping the openinference team.

Thanks for the detailed bug report. This is a confirmed issue — the isinstance(message, dict) check on line 182 of _wrappers.py silently skips ChatMessage objects since they aren't plain dicts.

There's an existing open PR that targets this exact problem: PR #2059 — "Instrumentation of smolagents are not handling the LLM input messages properly anymore". It hasn't been merged yet, though, so there's no fix in the current release.

As a quick workaround, ChatMessage objects support attribute access (.role, .content) similarly to dicts. A minimal fix in _llm_input_messages would be to convert ChatMessage to a dict (e.g., via dict(message) or message.model_dump() depending on the class) or to use getattr instead of .get() — similar to how _llm_output_messages already handles objects with getattr just a few lines below.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}