Vercel AI SDK chunkMs timeouts and a 55 s Sume jobs_wait

Vercel's ai@7.0.136 stops the chunkMs and firstChunkMs timers when the model response ends, so a slow local tool no longer trips an output timeout. stepMs still covers the whole step, so if a tool in the step calls Sume's jobs_wait, which holds up to 55 seconds, set stepMs above that.

What does the release say?

The release note states the timers stop when the model response ends so long-running local tools do not trigger output timeouts, that stepMs still covers the whole step, and that each retry gets fresh budgets.

How long is a Sume wait?

On hosted MCP, jobs_wait defaults to 50 seconds and is capped at 55. A larger timeout_seconds is clamped and reported in wait_slice_clamped. When the slice ends with wait_slice_expired, the job is still running, and you call the wait again.

What should you set?

Work from the slice, not from the render:

• stepMs above the longest slice you plan, 55 seconds, plus model time.
• A loop or several steps for the full render. Do not raise one step to many minutes.
• The same idempotency_key on any retried create. A retry from the SDK gets fresh time budgets, not a fresh job.

What about the overall deadline?

Keep a client-side deadline for the whole job, 20 minutes for video is a reasonable ceiling in the docs example. A client timeout does not cancel the job; it keeps running and billing. See Jobs and results.