How are you handling retries in agent tool calls?

Keep hitting the same wall: an agent fires a tool call, gets a 502 or a timeout, and then the model either gives up or "helpfully" rewrites the arguments on the second attempt. Wrapping the tool in a deterministic retry layer with idempotency keys helped, but now the agent sometimes retries semantic failures (400s) that should surface to the user instead. Curious how others draw the line between transient and terminal errors. Do you give the model a retry budget, or handle it entirely outside the loop? Also interested in how you log tool args across attempts without the trace becoming unreadable.