Assistants code interpreter output bug with a tuple of float, image


I believe there is a bug in how the OpenAI system generates outputs returned in the call to get the run’s steps. When the output of the code the assistant writes has a final line like


The only output returned in the outputs key in the run steps is the image. The assistant itself can see the output of some_variable in this case and comment on it, but the value is not returned in the outputs via the API, which I think it should be.

Steps to reproduce

Use a vanilla assistant in the assistants playground with the code interpreter tool enabled, and try the prompt

write python code to create a normal distribution, plot it as a histogram, then define a variable names s_variable as the average of the values going into the histogram and in your final output line write s_variable,

Current result

Let me know if anything is unclear, or more detail is helpful!