English source · 9 min

Completing the connected merchant workspace

Original implementation sequence approved on 21 October 2026, following the 20 October workspace. The sequence below preserves the original acceptance criteria; dated sections record the work delivered since then. Do not read every row as still open or as already publicly accepted. Current public status remains the release entry point.

Comparing recorded orders: hosted acceptance

The native rule editor, API and MCP now share a bounded read-only condition comparison on existing frozen order.placed Flow inputs. It compares a revision-bound saved condition with the unsaved draft, preserving frozen references and reporting missing facts as unavailable. Actual checkout/Flow PostgreSQL tests cover both transports, current/withdrawn rights, foreign shops and unchanged orders, receipts and jobs. This is a selected Flow sample, not complete checkout repricing or a revenue forecast. Usage, limits and acceptance. Public Core/Studio 6a71d072 now compares the existing EUR 49.90 test order in the original Studio: saved threshold EUR 100 is false, draft threshold EUR 1 is true. Editing clears the answer, an empty Flow is explained, and reload retains the saved rule at revision one and the single original order. See the linked guide for screenshots and the exact selected-sample boundary.

Reusing work: hosted acceptance

Direct and graphical Flow AI proposals now join the existing ready-task and personal-approval path. Actual local product-event tests cover one private native write, replay protection, unchanged live data, emergency stop, rights withdrawn during inference and late results after worker lease expiry. Behavior and acceptance scope. Public Core/Studio 09b1aa35 now completes the original configured-model proposal through the personal workspace: exactly one hidden native category, revision one, retained completion after reload and no category in the parent live shop. Public browser evidence. Public Core/Studio 4c0392ae also restores that original completed task directly in Flow history after reload, retaining the private workspace without asking for another approval. Direct history and the personal workspace now share the native progress/continuation card and current control checks. The linked evidence includes the public direct-history screenshot and 761 frontend, 27 durable-agent and 43 tenant-isolation checks. Public Core/Studio 650529c7 adds Reuse as a Flow for completed native work: reference-only source tasks, explicit event-field mappings, inactive drafts, per-Flow/workspace daily limits and the original personal review/execution path. See the usage and permission boundary. Its public browser acceptance now covers the completed-task handoff, an inactive draft, explicit activation, one real product event, a new personal approval, one hidden category at revision one, retained completion after reload and absence from the parent live shop. The test Flow is inactive again. The evidence page includes separate screenshots for this extension; generated per-run references are observed publicly, while other event-field mappings and competing daily-limit workers are covered by local native tests.

Collection preparation: implementation and public acceptance

The first extension uses the existing Rust task worker, translation queue and site-builder app; it adds no independent orchestration service or app tables.

An explicit productIds array on merchant.translations.create restricts a job to 1–1,000 distinct current-shop products. Missing selection preserves the historical whole-catalogue API behavior. Empty, duplicate, malformed and unavailable IDs are rejected. A collection operation inside an agent work plan always requires an explicit selection. Selections may also reference a product created by an earlier reviewed native step. Unknown IDs and references to later creates are rejected before approval. The selected IDs survive a worker restart; product contents are read when the draft is generated and applying that draft checks its original product revision. This is not a snapshot of catalogue text at queue time.

A private reviewed plan can combine native category/product operations, scoped translation preparation and managed page/app saves. Translation preparation yields the task worker while its existing translation worker runs. The task advances only after the drafts are ready. A failed child stops subsequent steps. Explicit task continuation resumes failed child work and retains completed operation receipts. Child creation and its task link commit together. A dead translation worker stops with an explicit-continuation error instead of automatically repeating a possibly charged model call. Terminal child results wake the existing task scheduler; other waits yield for five seconds and keep no database backend. The child retains the parent trace. Each generated draft stores bounded provider/model, token-usage and registered prompt-version evidence, without copying prompt text or credentials into the receipt. Native admission cases include null/empty/duplicate, forged-object, SQL-looking and instruction-like selections. This corpus tests actual admission, not semantic translation quality. Cancellation invalidates the child lease; the original session, current membership, private preview admission and shop emergency stop are checked before inference and before storing its result. Completed translation drafts still require their own review/apply action. Completing a preparation task never publishes pages or product text automatically.

The conversation links to the actual translation job. Use the existing native product/page editors for manual adjustments, the private storefront preview to see them, and the selected-release review to publish only the chosen configuration.

Local HTTP/PostgreSQL acceptance exercises one category, an existing product, a newly created selected product and a native landing page, deliberately invalid translation output, a fresh process, explicit continuation and selected publication of the page alone. Native category revision remains one; other product changes stay private. Cancellation during inference, original-session revocation and a killed translation worker also pass locally. PR #177 passed its full verification and is publicly deployed with migration 107. The public configured-model task exposed a provider schema error (fixed by deployed 693dd33) and then an unknown app field. Deployed 40b4a97a adds current app contracts and one bounded model repair. The continued public task reached page block validation. The app-owned contract now describes block IDs, localized text, product selection and form/extension limits explicitly, with a native-validated example. Rejections identify the block index and property without echoing content. The continued public plan now reaches review. Execution exposed an incomplete product edit after its category was saved. Native category-only edits and pre-review product-shape validation are deployed as 9d70e60. Continuing that exact saved public task completed all five steps; both Italian translation drafts were reviewed and applied in the private environment. The category was retained instead of being created again. Browser verification found two additional editor handoff issues: translation-review links reloaded into the live shop, and ordinary page editing required the optional form receiver. Repairs in PR #182 are publicly running as 498ced45: translation review retains the private environment and ordinary page editing no longer requires the optional form receiver. The private Italian preview shows exactly both selected products. A second configured-model request shortened the headings and moved the existing descriptions unchanged into paragraph blocks; explicit approval applied that correction, and the native preview reflects it. Selected publication exposed a dependency-order defect, repaired and publicly deployed in be620bfd. All six selected units published together, and the public simulated purchase retained the Italian product name and left an empty cart after return. Published page text still used the interface locale; the app-owned content-locale repair is publicly running as f3255dab, with headings, paragraphs, exact products and title observed in the browser. See the current status page.

The one-backend transaction-pool checks additionally exposed nested database acquisition in publication diffs, reviewed app staging and app reactivation. These paths now reuse the locked package snapshot or existing transaction, retaining current model/action/PII rights, environment ownership and digest/CAS checks. Activation uses the same tenant/app package fence as installation and record writes. The required pooler suite retains staging and collection tests.

Material and dimensions: implementation and public acceptance

The app-owned 1.0.1 pattern adds the existing material records to the native AI context and namespaced graph, shared by authorized API/MCP/public reads. Local acceptance now completes a second independently approved purchase after saved-input reorder and material/rule changes, checks idempotent order/stock effects and keeps the first order immutable. The original 1.0.0 definition and workflow tests stay inside the plugin directory. The hosted private browser journey now completes material/dimensions configuration, simulated EUR 417 purchase, empty cart return, native order/Flow inspection and immutable order display after a material update. Independent B2B approval and saved-input reorder now also pass in the local browser: the buyer restores a saved configuration, a separately signed-in approver reviews the exact quote, and simulated order RAC-b30bec87 retains the updated material, dimensions and surcharge. Return and reload retain an empty cart. This exposed missing configuration details in the approval inbox and reopened customer orders; both now use the shared frozen-order summary. Reopening the original order still shows its original material and EUR 68 surcharge after the app update. The combined journey now also passes on public ef093e7a, with browser purchases/list restoration and a separate native API approver: EUR 427 for the first order, EUR 447 after updating the same material, fresh approval required and empty carts after return and reload. The first approval retains its original values. Merged main 7f0d4266 also passes the complete two-account browser repetition: the approver sees material, dimensions and surcharge, approves, and the buyer completes RAC-34c09fc1 for EUR 447. Reopened buyer/company orders retain accepted values; the older order still shows EUR 78 surcharge and EUR 427 total. See the local and public sections of the evidence page below. The public fb8d54e0 follow-up restores saved material and dimensions after a full reload even before any line item exists. Simulated order RAC-2246a373 then preserved Walnut Select, 250 x 80 cm, EUR 78 surcharge and EUR 427 total; returning to the shop showed cart zero. The earlier b0b8cba7 repair retains configuration controls when a subsequent purchase starts a fresh cart. Evidence and continuation repairs.

The expanded one-backend check also caught a nested settings read when saving roles. Role writes now use the existing locked commerce configuration reader in their current transaction. Current delegation and revision checks remain in the same owner. Required transaction-pool acceptance includes roles, native object extensions, material graph/upgrade/reorder and UCP purchasing.

Product gaps: current hosted acceptance

The public assistant inspected one selected private product, identified missing German text/media and prepared a translation job for exactly that product. Separate review/apply produced visible German storefront text while preserving price and stock. Observed scope and remaining limits. This establishes a bounded correction path; it does not claim whole-catalogue coverage or completed image generation.