Claude Code routine limit of 30 fires an hour: use a bulk run

Claude Code routines count Run now, API fires and one-off re-runs together at 30 an hour per routine, and the action fails until the window resets. If your plan is to fire a routine once per video, the 31st video in an hour is the one that fails. Fire the routine once instead, and let that one run start a single Sume bulk run: one POST queues up to 100 Format runs with a concurrency window of 1 to 16.

The limits are from the routines page, read on 2026-10-03; the bulk-run contract is Sume's Format bulk runs page.

What are the routine limits?

Anthropic documents separate hourly limits by how a run starts, and says none of them has overage.

How does a bulk run fit?

A Sume bulk run takes concurrency (an integer from 1 to 16) and items, 1 to 100 entries, each the same body as a single Format run. Every item can carry its own generation_spend_cap_usd, and the create call returns 202 with a queue you poll at GET /v1/format-run-queues/{id}. A bad item fails the whole create with 400 and details.index before any queue exists.

Workspace generation concurrency still applies to the children: Free 1, Pro 4, Startup 8, Scale 20. Extra accepted jobs wait as queued until a slot opens, and a full queue returns 429 queue_full, so a bulk run of 100 on a small plan is paced, not rejected.

What does the routine's single call look like?

This script builds the body for a 40-item batch and prints the envelope. It runs offline, so you can check the shape before the routine sends it with an Idempotency-Key minted for the batch.

import json, uuid
products = [f'sku-{n:03d}' for n in range(1, 41)]
body = {
    'concurrency': 4,
    'items': [
        {
            'instruction': f'Make the promo video for {sku}.',
            'input': {'sku': sku},
            'generation_spend_cap_usd': 5,
        }
        for sku in products
    ],
}
assert 1 <= len(body['items']) <= 100
print(json.dumps({k: body[k] for k in ('concurrency',)}))
print(len(body['items']), 'items, header:', str(uuid.uuid4()))

What should the routine do after it queues the batch?

Before you size a batch, remember that a bulk run create fails whole with a 400 and details.index if any item is invalid, so validate items in your own code first. The queue's completed state means every item is terminal, not that every item succeeded; read each item's status before reporting. Use a fresh Idempotency-Key for each batch, and the same key only when you are retrying the identical batch.

• Read counts on the queue: completed means every item is terminal, not that every item succeeded, so branch on counts.failed.
• Put communication.webhook_url on each item; the queue itself has no webhook.
• Mint a fresh Idempotency-Key per batch, because replaying a spent key returns the old queue.
• Keep the routine's own prompt short: one fire, one batch, one summary.