Public: anyone can read this; the crew can change it.

Elastic compute: with a cloud account connected, a forge that does not fit on owned machines runs on a rented one, steered by the profit slider #11656

open deferred: behind the Later gate opened this

Picked up by scout; no pull came of it (failed) · 2 runs · run

Held by the Later gate, set by trevato . The crew picks nothing here until it is lifted.
opened this

Why. "i want to connect to a cloud service provider to add elastic compute and use the profit slider." Today Settings → Economics shows the slider at 50 ("cloud takes work above 85% of owned capacity"), but placement reads "no cloud body: bursts queue", no cloud account is connected, and every price is 0. Measured this week: a building forge uses about 3 cores for under a minute and idles near zero; the owned machine has 72 CPUs and at most 8 forges awake.

What.

  • Connect once. Settings → Cloud takes a provider token (start with DigitalOcean, which the forge's cloud tooling already names and the owner already uses; the seam stays provider-neutral) and proves it: lists regions and sizes, the hourly price of each, and what the token may do. Sealed, never shown again.
  • A cloud body. When owned capacity is past the slider's threshold, or a forge is larger than what is free, the forge is placed on a rented machine: created from the forge's image, its data on a volume that outlives the machine, reachable at the same <name>--forge address through the existing edge. One forge per machine for isolation; the machine is sized to the forge's plan.
  • Scale to zero. A cloud forge that sleeps has its machine destroyed and its volume kept (or snapshotted, whichever is cheaper for its size; the page shows the choice and the cost). A request or a due timer (#11595) wakes it: a machine is made, the volume attached, the request held and answered. Wake time is measured and shown; a schedule is never missed because a forge was asleep.
  • Moving home. When owned capacity frees up, a cloud forge moves back at its next sleep. Nothing moves while it is in use.
  • The slider is real. Settings → Economics takes the true numbers (what the owned machine costs a month and how many forges it holds; the provider's hourly prices are read from the account) and shows, for the position chosen: what a forge costs awake and asleep on each side, the margin at each plan's price, and this month's bill split owned / cloud / storage / crew. The price fields that are 0 today are filled or flagged.
  • Guards. A monthly cloud spend ceiling (required before the first machine is made), a cap on machines, an ask to the owner at 80% of either, and an orphan sweep that finds and removes any machine or volume the forge no longer knows.

Order. First a spike that proves, on one real machine: the crew's sandbox works there (bwrap --unshare-user), cold wake time with a volume attached, and the real cost of an idle and a busy forge of each size. Write the numbers in docs/CLOUD.md with a go or no-go before building the rest. (This replaces the parked spike #11145.)

Acceptance. With the owned side forced full in a test, a new forge lands on a (fake) cloud body, sleeps to zero machines, wakes on a request with its data intact, and moves home when room returns; the spend ceiling stops creation at its limit; the orphan sweep removes a planted stray. On production: one real forge runs on a rented machine through a sleep and a wake, and Settings → Economics shows its true cost.

2 earlier comments
commented

Parked by the crew: the run ended without a pull: declined: this issue orders a spike on one real rented machine before anything else is built, and I can't run it. It must show that bwrap --unshare-user works there, give the cold wake time with a volume atta

commented

Parked by the crew: the run ended without a pull: declined: this issue orders a spike on one real rented machine before anything else is built, and I can't run it from here. The sandbox reaches only the forge, the package registry and this app's own deployment

Parked twice since a person last spoke here, so it is the planner's now. Every reason:

  1. the run ended without a pull: declined: this issue orders a spike on one real rented machine before anything else is built, and I can't run it. It must show that bwrap --unshare-user works there, give the cold wake time with a volume atta
  2. the run ended without a pull: declined: this issue orders a spike on one real rented machine before anything else is built, and I can't run it from here. The sandbox reaches only the forge, the package registry and this app's own deployment

The planner splits it, replaces it with a fresh issue, or defers it; a person can also say what to try. Nothing starts here until then, and the board lists it.

commented

Deferred behind the Later gate (#595). Why: the owner's direction of 2026-11-12 puts the v1.0.0 list first and elastic compute is not on it (the cloud host #11145 comes right after v1.0.0); and both tries stopped on the same wall — the first step is a spike on one real rented machine (bwrap --unshare-user there, cold wake with a volume attached), which a builder's sandbox cannot reach. When the gate opens, that spike is a person's or the steward's step with a cloud token, and the build splits from its findings.

Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs