mcp 0.5.0 - tool retry behavior is completely undocumented and probably broken

spent yesterday testing retry behavior in staging and honestly i have no idea what the intended behavior even is because there's zero documentation. what we observed: - agent retries failed tool calls with zero exponential backoff (just hammers the endpoint) - no jitter in retry timing - retry count seems hardcoded at 3 but can't confirm because it's not documented - tool responses aren't cached so retries re-execute expensive operations - no timeout so if a tool hangs the agent just waits forever is any of this intended behavior or is the retry logic just completely half-baked in 0.5.0. need to know before we deploy next week