Factory docs, home
Page navigation

Pre-release documentation. Proposed designs are labelled in the page; acceptance evidence is maintained in the build review. Local paths appear as placeholders. Current state: Build review.

Proposed; unimplemented and uncommissioned. This page preserves the 2 October 2026 revision-4 design by Opus and Fable. Astra's round-5 PASS accepted that design against its historical repository baseline, with one Low finding corrected below. It established no runtime acceptance. Claude subscriptions remain unavailable under the current commission; saving this proposal starts no provider, account setup or scheduled task.

What the design would do

The Factory would send bounded production work to a separately commissioned owner worker or vendor environment. A worker returns an untrusted Proposal. Existing sealing, independent verification, admission and release boundaries still decide what may happen next. Remote production alone proves no release or customer value.

Each order retains its original identity, fixed allowance, route configuration and current authority. An epoch owns a bounded attempt. Reservations precede contact; an uncertain submission cannot be sent again under a new identity. Cancellation and fencing settle eligibility separately from knowledge of remote compute. A timeout alone proves neither process death nor a failed remote effect. Restore must retire the previous authority, reconcile original admissions and hold unresolved capacity before admitting replacement work.

The initial form keeps the control plane on owner hardware, with a cloud relay carrying only a reviewed metadata class. A later hosted journal or tenant service needs a separate placement decision and proof that its records, backups and logs obey that class. Scope paths, private material and credentials stay outside hosted metadata. An export to a work mirror needs an approved exact manifest; returned changes remain untrusted. A vendor session cannot be assumed isolated from other sessions' result branches.

Accounts and authority remain separate gates

The possible routes are an owner worker, Codex cloud, a dedicated Codex agent and a Claude routine. They remain unqualified proposals. Naming the routine route does not reopen Claude CLI use or establish that a preview endpoint is available or permitted. The design proposes subscription-backed work only; it permits no Platform-key fallback, exported sign-in, copied session, paid overage or automatic provider substitution.

Before any implementation can operate, the owner and independent reviewer must resolve:

  • Dedicated operating-system identities, each tool's own sign-in, vendor terms and whether the subscription permits the proposed use.
  • Broker custody and scope for any dispatch credential; no agent reads an owner's authentication store.
  • Dedicated cloud-account and mirror boundaries, repository rules, installation permissions and any purchase or billing decision.
  • Exact export approval, result attribution, collector isolation, remote cancellation and independently observed fencing.
  • Restore authority outside the restored journal, original admission retirement, capacity reconciliation and usage windows.
  • Hosted-data placement, future tenant isolation and the later phone-signing dependencies described in the owner-presence proposal.

These grouped gates summarise the original 23 unresolved owner decisions. The archived design remains their full historical wording. Defaults were recommendations, not consent. Fixture increments CO0–CO5 would establish local contracts only; live routes require their own commissioning and evidence.

CO1 correction: validate fields according to their meaning

The original CO1 required the literal canary FIRE_TOKEN to be rejected in every string field. That contradicts the proposed vendorRef grammar, ^[A-Za-z0-9][A-Za-z0-9._:-]{0,127}$, which permits it.

The corrected acceptance criterion is field-specific: test each field's grammar and reject wrong types, paths or free text where that field excludes them. A syntactically valid vendor reference must still be established as the vendor's actual identifier, rather than a label or local credential name, before commissioning. Separately seed credential names and other local data at their real input boundaries and prove that they cannot reach hosted records, backups or logs. Identifier syntax alone cannot detect a secret name. Refusals name the field and rule without echoing the rejected value.

Review and provenance

The source manifest records exact document hashes under cloud-design, cloud-astra-r5 and cloud-dispositions, alongside the original convergence and status notes. Private source locations remain in the local preservation record. This condensed page corrects CO1 and stale status wording; it does not replace the historical review or claim that all proposed contracts have been requalified against current code.

The current outcome map, including separately registered Activity scope E42, has 42 epics: 27 customer-value and 15 adoption epics. This proposal creates or closes none. The historical GO decision did not activate a route, and the historical pending note did not prove a canonical status update. No automation is resumed by this document.

Source: docs/explanation/cloud-orchestration-proposal.md