Claude Code 2.1.286 prompt count '2 of 5': parallel Sume calls

Claude Code 2.1.286, published on 2026-09-30, adds a count such as "2 of 5" to the permission prompt when several requests stack up, and shows waiting prompts oldest first (read 2026-10-02). If an agent fans out five parallel Sume generations, you will now see five prompts counted off, and each one is a separate paid call. A single script_run call is one prompt, but approving it approves everything the script is allowed to call.

What the changelog says

Two entries matter here, and one from 2.1.287 on 2026-10-01 completes the picture.

Why parallel Sume calls stack

Each paid create, such as generate_video or tts_create, is its own tool call and needs its own idempotency_key (MCP tools and gates). Five scenes become five prompts. Read each prompt for the model, duration and max_spend_usd before you approve; the count tells you how many are left, not what they cost.

For the wait, Sume's jobs_wait accepts up to 20 job_ids in one call, so one batch wait covers the jobs instead of five separate waits (Jobs and results).

What script_run changes, and what it does not

Sume documents script_run as a short program that calls tools in a loop or in parallel and returns one value, for use when a turn needs three or more independent calls of the same shape. Inside the script, each call runs under the same gates as a direct call, paid creates still need their own idempotency_key, and the run is bounded by timeout_seconds (5 to 55), max_calls and max_paid_calls.

From Claude Code's side that is one tool call, so one prompt. The trade is that one approval now covers up to max_paid_calls spends. Set that bound to the number you actually intend, and ask for dry_run=true in the first pass. I have not verified how Claude Code words the prompt for script_run, so read the arguments before you approve.

A practical rule

Sume's OAuth page covers how to leave Write off when you only need discovery.

• Under about three parallel calls, approve them one by one and read the counted prompts.
• At three or more, use script_run with a tight max_paid_calls and a max_spend_usd where the tool supports it.
• With OAuth mcp:read only, paid calls return insufficient_scope, so no prompt can lead to a spend.