Google Models - 'role' Key Error on LLM Response

[RaspberryPicardBox]

Describe the bug
Seemingly at random points when a heartbeat message is issued, the LLM response is misconfigured and does not include a 'role' key.

I suspect this is due to the rate limits of Gemini returning an unexpected message after the rate limit is exceeded on the free plan.

Please describe your setup

• How did you install letta?
  • Official Docker image.
• Describe your setup
  • MacOS Sequoia 15.3.1
  • Letta is ran as a Docker image, with a Google Gemini API key, and Postgresql persistence.

Screenshots

Additional context
Unfortunately, I am unable to see the generated message. It could be that the message was not generated at all due to rate limits, and therefore the 'role' key is not there as the message is blank.
This is likely the cause, but I am unable to debug further currently.

LLM details

Running with Google's gemini-2.0-flash-001 model, and the letta-free embedding model.

[sarahwooders]

Thanks for reporting this - looking into it!