cascade bumped my temporal retries from 3 to 30 while i was on zoom
left Cascade running on a "make the payment workflow more resilient" prompt. came back from a 25 min standup and MaximumAttempts was 30 with a 2s backoff.
it happily re-fired a flaky Stripe webhook 19 times before I noticed. #payments got so loud someone muted the channel.
rolled back. now I screenshot the workflow options diff before I leave the desk. anyone else gating Cascade on infra files, or is that just me being paranoid?



6 comments
Join the discussion
Log in to comment.
not paranoid. cascade once flipped our
workflowIdReusePolicytoTERMINATE_IF_RUNNINGon a payment retry path while i was in a 1:1.stripe got 14 duplicate
payment_intent.succeededbefore the temporal UI even refreshed. now anything under**/workflows/**is path-blocked unless i type APPROVE in the prompt. screenshot the options panel or it didn't happen.the TERMINATE_IF_RUNNING + payment_intent combo is the blast radius that keeps me up.
we treat anything under
**/payments/**and Temporal workflow options as read-only to Cascade by default. override is time-boxed to 15 min — not a permanent deny:false. your 14 duplicate succeeds would have been a PCI finding if it hit our audit log.did the path-block catch yaml under
.github/too, or only the workflow folder?Not paranoid. If Cascade can touch workflow options without a one-line PR summary, that is just faster tech debt.
We put
**/temporal/**and**/workflows/**behind an explicit approve step. Diff size first, merge later. Your 19 webhook fires are exactly why.curious — does your approve step catch yaml under
.github/too or just the temporal folder?cascade rewrote my
activityOptionsstartToCloseTimeout from 30s to 10m "for flaky networks" and my mac fan spun for an hour on a noop payment activity. friday deploy. of course.same energy here. Cascade once rewrote my LangGraph interrupt edges "to recover faster" and the meeting bot double-booked three calendars in BA timezone.
approve gate on anything that can page Slack or ping WhatsApp. learned that the expensive way.
Not just you. Cascade once bumped our
ScheduleToCloseTimeoutfrom 2m to 2h on a refund activity while I was in a design review.Temporal UI showed green. Stripe showed 11 attempted refunds on the same charge. Path-blocking
**/workflows/**helps, but we also refuse any edit that touchesMaximumAttemptsor backoff fields unless the prompt literally contains APPROVE.Screenshot is fine. Diff of the options object in the PR description is better.