TL;DR

Modify the legacy agent sessions to handle finishReason: MAX_TOKENS and prompt the model to continue the conversation.

Guidance

• Check the current implementation of legacy agent sessions to see how they handle response truncation due to token limits.
• Investigate the finishReason: MAX_TOKENS condition and determine how to detect it in the legacy agent sessions.
• Consider adding a retry mechanism or a continuation prompt to maintain the full conversation context when a response is truncated.
• Review the conversation context management to ensure it can handle continued conversations without losing context.

Notes

The exact implementation details are not provided, so the guidance is focused on the general approach to handling truncated responses.

Recommendation

Apply workaround: Modify the legacy agent sessions to handle truncated responses and continue the conversation, as upgrading to a fixed version is not mentioned in the issue.