Claude Code remote MCP backoff up to 30 s: retrying Sume safely

Claude Code 2.1.295 (Oct 8) makes remote MCP connections in headless and SDK mode back off for up to 30 seconds after drops. On Sume, a drop after a paid create is not proof the create failed, so the retry must carry the same idempotency_key.

What changed?

The changelog says remote MCP in headless and SDK sessions now backs off up to 30 seconds after connection drops. It describes connection handling, not tool-call replay.

What happens to the call that was in flight?

Unknown from the client side. The request may have reached Sume and created a job before the connection dropped. Sume's guidance for a timed-out submit is to retry with the same idempotency key so the retry returns the original job, and to poll a job you already have rather than submit again.

A retry checklist

Put these rules in the agent prompt:

• Generate one idempotency_key per intended render and reuse it on every retry of that create.
• After a reconnect, call jobs_list or jobs_get before any new create.
• Treat wait_slice_expired from jobs_wait as a scheduling fact, not a failed job.
• Send max_spend_usd each time, because the cap is enforced only when sent.

How long can a wait outlast a backoff?

A jobs_wait slice is 50 seconds by default and at most 55, so a 30-second backoff plus a slice fits inside most CI step limits. A full render still takes many slices. See Jobs and results.