The detail that points away from auth/connection is your Update 1: fresh chats work, then it dies after 2-3 invocations, and Update 2 says responses are getting truncated. If OAuth or the connector registration were actually broken it would fail from the very first call, not degrade. That degradation curve plus truncation is what context pressure looks like: the runtime keeps the cheap tool descriptions in the prompt (so the model still “sees” them) but stops registering the callable functions once the conversation is near the budget. From the model’s side that reads exactly as “the definition is shown but it isn’t a callable tool.”
The usual driver is large MCP tool results. A couple of fat tool responses fill the window, and after that the runtime drops executable tools to make room. So before touching the connector again, I’d cap and paginate the server’s outputs: return a short structured summary plus an id/cursor to fetch detail on request, instead of the full payload.
Quick way to confirm it’s context and not a per-connector bug: new chat, make the very first tool call return a deliberately huge result. If invocation breaks immediately instead of after 2-3 calls, it’s window pressure, not your OAuth or tool schema.