world.queue passes delay_seconds as float; queue server rejects with "must be an integer between 0 and 86400"

[kalyanaramansanthanam]

Environment

• vercel==0.5.8
• Python 3.12 (Vercel Lambda runtime)

The bug

The workflow runtime computes the delay for a vercel.workflow.sleep wakeup as:

# src/vercel/_internal/workflow/runtime.py:441
seconds = (sus.resume_at - now).total_seconds()

That's a float. worlds/vercel.py:237 clamps it (also float math) and passes it through unchanged to vqs_client.send_async(... delay_seconds=...). The Vercel queue server then rejects:

{"error":"Invalid Vqs-Delay-Seconds. Must be an integer between 0 and 86400 seconds (cannot exceed retention time)"}

So the very first sleep() wakeup queue-send 400s.

Repro

from vercel.workflow import sleep
from vercel._internal.workflow import Workflows

wf = Workflows()

@wf.step
async def noop_step(*, payload: dict) -> None:
    return

@wf.workflow
async def my_workflow(*, payload: dict) -> None:
    await noop_step(payload=payload)
    await sleep("60 seconds")  # → world.queue(..., delay_seconds=60.0) → 400

The chain self-heals eventually (after resume_at is in the past, min_timeout_seconds flips negative and no delay_seconds is sent), but each cycle takes ~5 minutes of queue redeliveries and floods the worker logs with 500s/400s. Compounded by #(other-issue-id, the DuplicateIdempotencyKeyError AttributeError) which masks the original BadRequestError.

Suggested fix

int(delay_seconds) (or math.ceil if you want to err on the side of "wake no earlier than requested") before passing it to the queue client. Fractional seconds are meaningless at the queue's per-second granularity.