So we had a few goals here
1. Be opinionated on best practises, tools and libraries
2. Not get in the way of what the developer wants to do
To that end the core is built on top of Temporal, and our llm package is a thin wrapper around ai-sdk that provides QoL enhancements (Prompt files, tracing, cost tracking etc..)
So for failures in general, and tool calling specifically there are two levels of retries.
1. ai-sdk level tool retries: The library by default handles tool call failures and will retry if the LLM deems it a transient issue, and will never hard fail if one of its tool calls in unsuccessful (unless perhaps you instruct it to).
2. Temporal level activity failures: Our workflows and steps are all configured with a base line affordance to reattempt steps that have failed. You as the developer are able to change this, you can make it so a step is never retried, or retried say 100 times with exponential backoff.
Hope that helps!