Fixed retry loop on tool timeouts

Some agents that called slow tools could get stuck retrying the same call. When a tool returned a partial result right before its timeout, the SDK treated it as a new attempt and reset the retry counter.

What changed

• Retries are now counted per tool call ID, not per response, so partial results can’t reset the counter.

• The default maxRetries for tool calls is 3, with exponential backoff starting at 500 ms.

• Exhausted retries now emit a tool.retry_exhausted event you can alert on.

Do I need to do anything?

Upgrade to @sigil/sdk@2.3.1. If you had set a custom retry handler to work around this bug, you can remove it.