What
Some MCP servers, add-ons that give Claude Code extra tools, can run long tasks in the background. Claude Code used to check on such a task by asking the server directly, again and again (this is called polling). Task records now note who does that checking:
- Whether a service or Claude Code itself (the client) is polling.
- Offers to take over polling, when polling started and whether it stopped.
- Times a poller gave up.
- A "not taken" result when Claude Code keeps polling itself.
The polling loop can now wait on another poller and send extra request headers. When a task ends, follow-up steps can be queued to run afterwards. A check for servers that are not configured is skipped while another poller is in charge.
Why
This looks like a way to hand polling of long MCP tasks to a service, so Claude Code would no longer need to keep checking on them itself.