Python: Bug: Handling Warnings and Errors from the Code Interpreter Tool in SK Agents

[CristinaStn]

When running complex Python code inside the Code Interpreter, we’ve observed that warnings and errors from the Python runtime are streamed directly back to the agent—and consequently shown to the end-user.

Example
Query to the agent:
"Generate a sunburst chart based on job titles in the report."

Expected behaviour:
The agent explains what a sunburst chart is and the steps it will take to generate it.

Actual behavior:
While streaming the response, the agent emits chunks of messages that include internal warnings from the Code Interpreter, such as:
/home/sandbox/.local/lib/python3.11/site-packages/plotly/express/_core.py:1637: FutureWarning: The frame.append method is deprecated and will be removed from pandas in a future version. Use pandas.concat instead. df_all_trees = df_all_trees.append(df_tree, ignore_index=True) /home/sandbox/.local/lib/python3.11/site-packages/plotly/express/_core.py:1637: FutureWarning: The frame.append method is deprecated and will be removed from pandas in a future version. Use pandas.concat instead. df_all_trees = df_all_trees.append(df_tree, ignore_index=True)
(same warning message, duplicated)

Finally, the run fails with a non-intuitive error "Run failed with status: failed for agent GeneralAssistant and thread thread_PTCvyjhsL3dzqtASi7bH3TRp" with Last Error: Sorry, something went wrong.

Questions:

What is your recommendation to handle scenarios like this?
How we can avoid sending this kind of code interpreter error&warnings messages to the end-user? Ideally, such logs should be available for debugging but hidden from the user-facing interface.
Is there any mechanism available to intercept or filter the input/output of the Code Interpreter tool?
(e.g., something similar to the Filter concept used in Plugins?)

Semantic-kernel version: 1.29.0

[moonbox3]

Hi @CristinaStn, thanks for filing this. Which agent type is this for? It is a good point. Up until now, we've assumed whatever output comes back from the code interpreter tool is something we can show to the end user. Honestly, I haven't seen warnings like that before from the code interpreter, and I agree that's not something we'd want to expose.

Right now, our approach for agent invocations is that the caller gets the final agent message, and any intermediate tool messages go through the on_intermediate_message callback. The code interpreter tool is a bit of an outlier here, since we return those code outputs as direct messages, rather than just as interim content. I've actually thought for a while that we should handle the code interpreter the same way as other tools, surfacing those interim results through the callback so devs can filter or process them before sending anything to the user.