Stopped a Sume MCP call in Claude Code? The job still runs

Stopping a long MCP call in Claude Code does not stop the Sume job behind it. The render continues and still bills until it finishes or you cancel it, so recover the job by id instead of creating it again.

What does stopping actually stop?

Your wait, not the job. The client stops listening; the server-side render was already accepted. This post does not rely on any Claude Code release note: the behavior follows from how Sume jobs work, as an accepted job is a separate resource from the call that read it.

Which Sume call is the long one?

The long call is almost always jobs_wait. On hosted MCP it holds at most 55 seconds per call, and 50 if you omit timeout_seconds; a larger value is clamped, not rejected. When a slice ends with wait_slice_expired, the job is not failed. You call jobs_wait again with the same ids.

The paid create that started the job returns quickly with a job id. The stop affects your wait, not the job.

How do you recover?

Ask Claude for the job id from the create result, then pick one:

• Wait again with jobs_wait using the same ids. Never repeat the paid create.
• Read the job with jobs_get or jobs_status.
• Cancel with jobs_cancel. It needs mcp:write and an idempotency_key, and it works only before generation starts. After that the API answers 409 job_generation_already_started and the job runs to completion.

What if you only had OAuth read?

Then jobs_cancel returns insufficient_scope. You can still read the job. Add mcp:write or use an API key if you want the agent to cancel. See Jobs and results.