The test flaked on Linux x64 when the model responded with text instead
of calling the bash tool. Without a tool call, no approval is generated
and the test fails. Now retries the prompt up to 3 times (same pattern
as the prestream approval recovery test).
👾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
