English source · 18 min

Private customer scenarios — 22 October 2026

Local acceptance

The compiled Rust Core, real PostgreSQL and Chrome were used with a disposable synthetic shop. The sandbox was stage-698dee6ea85a41a1aa69; all buyers use example.test addresses. This is local evidence, not a public deployment claim.

The browser goal was: find a desk lamp, inspect its product information, stay within the customer's budget and purchase only if the native total fits. Three editable profiles used German content and Germany: a budget buyer with EUR 1.00, a first-time buyer with EUR 300.00 and a focused buyer with EUR 300.00.

The actual local Ollama model qwen3.6:35b chose the actions. In the final run:

• The EUR 1.00 buyer finished without buying: even the cheapest observed product exceeded its budget.
• The first-time buyer inspected the Arc desk lamp, added it and completed native simulated order eaf1dedb714242238f8f3ffdff0be0cb for EUR 79.90.
• The focused buyer inspected the lamp, added it and completed native simulated order 4673d92ab8de496088f320136dbc58ed for EUR 79.90.

Both orders appeared in the ordinary sandbox order grid. Full browser reload retained the saved test history. Reusing a test restored its exact profiles, instruction and model selection. Baseline and selected results appeared together; the interface warned that changed starting stock and model decisions affect the comparison. Desktop and 390 × 844 mobile controls were inspected.

Defects found and fixed during acceptance

The first real-model run interpreted budgetMinor: 100 as EUR 100. The native checkout nevertheless rejected that attempted purchase because the true budget was EUR 1. The model input now receives the exact decimal budgetAmount and an explicit currency, without the ambiguous minor-unit field. The repeated run above correctly stopped before checkout. This preserves the failed observation instead of treating the first run as successful model reasoning.

Independent review and native races also found and repaired nested database-pool acquisition during checkout, cancellation while waiting for a cart lock, and a lease expiring while waiting for the task lock. The final transaction fence checks current authority and the lease after those waits. UI regressions cover stale scope/selection responses, resumed polling and incomplete comparison snapshots.

Automated verification

• 796 frontend tests across 134 files passed; the seven simulator tests include private admission, retained edits, shop changes, exact repetition, continuation, late-control responses and pending comparisons.
• Rust unit tests, Clippy, formatting, source/permission/localization boundaries, generated operation contracts and the production formal checks passed. The formal checks concern extracted policies, not a proof of the entire simulator.
• The registered agentic_scenarios suite used actual native HTTP/PostgreSQL transactions: idempotent cohorts, exact budgets, two-connection concurrency, injected catalogue text, lost checkout acknowledgement/cold restart, multiple products, stock failure then repair, cancellation, lease expiry, foreign access, revoked rights and emergency stop.
• Restricted signed system/tenant roles were tested with two replicas and the workspace-role, durable-agent and scenario paths. This was a focused signed-role run, not a repetition of every HTTP suite under that configuration.
• Connected suites passed: checkout review, staging, developer documents, automation lifecycle, catalogue management, configured business capability, agentic commerce and the five-source migration assistant. Qdrant was started in an isolated local store after an earlier dependency-unavailable failure.

The agent journey uses native APIs. The screenshots prove the merchant workbench was rendered in Chrome; they do not imply customer agents navigated a browser or that synthetic purchase rates predict real conversion. No real payment or mail delivery was performed. Public rollout and acceptance are recorded separately.

Connected catalogue and app follow-through

The follow-up adds Products → Prepare catalogue over the same paged native product read, with explicit enabled-language name/description coverage, latest image-job state and direct native text/media editor links. Chrome verified the private Workshop catalogue, including an explicitly injected failed-provider fixture for chair, reopening the media tab and its retained German failure message. Other saved product images remained present. No paid image call was made. Legacy main-language source columns are counted exactly as in the native editor.

The catalogue HTTP/PostgreSQL/MCP suite passed with failure persistence, foreign-shop absence, unauthorized denial and selected publication. Image UI regressions cover locale refresh during creation, delayed post-commit reconciliation, keeping writes blocked through that reconciliation, and discarded results from a previous scope. The app diagnostics now identify missing view/model/action references, nested calls, API routes and Intelligence references, with direct links to the existing editors. They remain draft feedback; native authorization and installation validation apply.

The native Product Advisor template was saved and installed privately in Chrome. Its app review panel and Customer simulator button opened the same private scenario workbench. A 390 × 844 check found and repaired intrinsic grid overflow; the repeated check measured a 390 px page and a 358 px preparation area. The shared table keeps its own horizontal scroll.

Native intent-to-purchase browser journey

After the connected changes, Chrome completed an ordinary shopper journey on the same private local shop. The actual Ollama qwen3.6:35b model compared the Field notebook (EUR 12.50) and Terra cup (EUR 24.90) as gifts within EUR 40. The shopper opened the notebook's native product detail, added it to the native cart, entered synthetic example.test contact data, selected free pickup and completed the explicitly simulated payment.

Order RAC-c29bc2b9 retained the German notebook name, EUR 12.50 total and EUR 0.82 VAT. Returning to the shop and fully reloading both showed cart 0. The earlier optional-privacy rejection remained in effect without a repeated prompt. The privately installed Product Advisor view was present on the native detail page. This acceptance concerns the native shopper route; it is not a fresh certification of the separate Storyfront iframe or of real settlement/mail delivery.

The final connected frontend run passed 805 tests across 137 files, including the image-job reconciliation and app-reference diagnostics regressions. Rust unit checks, Clippy, operation contracts, source inventory, structure checks and the extracted formal-policy/conformance checks passed again for that branch.

Public rollout and customer-scenario acceptance

After the complete required GitHub verification passed, PR 195 merged as bbb8b5ab. Northflank Core was updated at 15:43 CEST on 22 October from sunset-spoon-5707 (defac3dc) to the already-built distinct-giants-7353 (2253a627). The new instance passed its health check with zero restarts. There was no schema migration or change of database credentials for this release.

Chrome resumed the existing merchant session on app.veyndo.ai. All acceptance below uses the separate full private sandbox stage-62ddd2bbd80642e38c4f, named Kundensimulator Abnahme 22 Oktober. No catalogue/stock change was published. The actual configured qwen3.6:35b model chose the shopper actions.

The shared goal was to buy only Notizbuch »Stille Seiten«, inspect details, check stock and the complete price, and choose no substitute. Profiles used German/Germany, with budgets of EUR 1.00, EUR 100.00 and EUR 100.00.

1. Set the notebook's private stock to zero. All three customers stopped without buying. The budget buyer also correctly identified EUR 28.90 as over EUR 1.00.
2. Restore only that private stock to 20. Use these customers again restored the exact goal, profiles, budgets, languages and inference selection.
3. In the repeated run the budget buyer still stopped. The first-time and focused buyers inspected the notebook, created separate native carts and completed simulated orders da90888b26e0456ca681b37dbcd6b265 and c6b82f0fe7274034af388e9977cafbcc, each for EUR 28.90.
4. The comparison identified matching inputs and a changed observed catalogue. The ordinary private order grid contained exactly those two orders. Opening RAC-c6b82f0f showed the German product name, quantity one, Demo payment, pickup and the native order.placed Flow history.

Full Studio reload, reselecting the same sandbox and reopening saved groups b0055539-5fd2-4439-9d50-5a8665384d5a and 37ab55ad-c97a-4ca0-b0a6-3a1303aacda5 restored both action journals and the original order IDs without starting another customer. The public 390 × 844 viewport had a 390 px document and 358 px scenario area; editable cards and selectors remained inside that area. Temporary viewport emulation was cleared after the check.

Public native shopper regression after the rollout

A separate Chrome tab rejected optional tracking, asked the real shopper assistant to compare the notebook and Stiftehalter »Tonruhe« within EUR 40, followed the notebook's returned native detail link and bought one notebook with synthetic contact/address data. The assistant correctly quoted EUR 28.90 and EUR 34.50, and rejected the pair's EUR 63.40 as over the combined budget.

The normal checkout completed RAC-90aa93a4, EUR 28.90 including EUR 4.61 VAT, with the German product name and explicitly simulated payment. Return and full browser reload both retained cart 0, optional personalization remained off, and no repeated privacy prompt appeared. The admin then contained exactly three orders: the two simulator orders plus this separate browser purchase.

These are public native API/checkout observations in a private sandbox. They do not constitute visual customer-agent navigation, a controlled conversion study, real provider settlement or mail delivery. A separate original Storyfront check is recorded below.

Original Storyfront iframe regression on the simulator Core

The separate original Lumen Experience was tested against Core 2253a627 as well. A fresh extension-controlled tab timed out with an empty iframe on its first open and ordinary retry; no console error was captured. Its unpurchased diagnostic basket item was removed afterward. This observation is retained as a browser-control reliability limit rather than labelled a successful first load.

A separately reloaded native Chrome incognito session then added the synthetic Kleine Tischleuchte Mattblau. Its first ordinary Checkout action loaded the embedded native privacy/contact/address form. Native browser input completed RAC-cf96e844, quantity one, German product name, EUR 59.90 total and EUR 9.56 VAT, using pickup and explicitly simulated payment. The receipt return closed the iframe and the original Storyfront bag showed the order-confirmed message and no purchased item. A full reload, hydration and reopening the original Bag retained Nothing yet. No manual user test, separate-checkout purchase, interception or browser security-setting change was used for this accepted journey.

Editable visual shopper selection

The follow-up replaces text-only product links with native product cards and a reversible shortlist. Local Chrome and the real qwen3.6:35b model compared Field Notizbuch (EUR 12.50) and Terra Becher (EUR 24.90). The cards displayed native images, descriptions, stock and prices. Removing the cup changed only the shortlist; opening the notebook and returning preserved that removal. Restore returned both cards without changing the basket.

The selected notebook used its ordinary detail page, including the installed Product Advisor slot, then native checkout. Simulated order RAC-fcbe27c7 completed for EUR 12.50 including EUR 0.82 VAT with the German product name. Return and full browser reload retained cart 0. A 390 × 844 check measured a 390 px document and 312 px selection. Switching to Spanish removed the preceding German answer immediately. Temporary device emulation was cleared.

All 810 frontend tests across 138 files, TypeScript, production build, localization and structure checks passed. Targeted regressions verify native product facts, duplicate/missing IDs, reversible removal, detail/back retention and rejection of late provider results and errors after a scope change. Independent review confirmed the navigation and stale-error repairs. This section records local browser acceptance; public delivery is recorded separately.

The complete catalogue/app follow-up CI also passed on 6b314459 (run 39789784144), including its actual database/provider suites and coverage gates. Its delivery is consolidated with visual shopping in PR 199 rather than landing intermediate release branches separately. The shared runtime tree at db6acdb4 is identical to the built source aa1dd115; only merge ancestry changed.

Readable profile destinations

A presentation follow-up uses localized locale names and the shared searchable country picker, while retaining the native channel's exact allowed codes. Its 811 frontend tests across 139 files passed, including destination search, original payloads, exclusion of unavailable countries and prior-scope dismissal. Local Chrome selected the German profile language and searched Germany by name and by ISO alias. At 390 × 844, the document and cards stayed at 390 px / 358 px; the expanded picker stayed between x=37 and x=353. Device emulation was cleared.

Final combined public rollout

The complete PR 199 verification passed. It merged as c948bccb. The profile-label follow-up passed its complete frontend/coverage lane again after the strict branch update (run 39802944161) and PR 200 merged as 07812240. The full runtime tree of 07812240 is identical to the built source f74d8ff9. The intermediate PR 196 and 197 implementations are included in PR 199; their separate delivery branches were closed as superseded.

Northflank activated affluent-week-4541 / f74d8ff9 at 17:20:12 CEST on 22 October, deployment vendune-core-66fcd94d4c. Immediately before this rollout, the service showed teal-farm-4382 / bbb8b5ab, the unchanged merged PR 195 tree, activated at 15:54:30. The first simulator deployment at 15:43 and the earlier observations above retain their dates. No schema, credential, permission or resource-size change was needed for this final rollout.

One automatically triggered post-merge main verification was explicitly cancelled as a duplicate: c948bccb and the fully verified db6acdb4 have identical complete Git trees. The required PR check was successful, not bypassed. The separate UI and documentation changes retain their own required checks.

Public acceptance after the final rollout

The new Core passed its health check with PostgreSQL, knowledge retrieval and explicitly simulated payment; its pod was running with 1/1 checks passing and zero restarts. Full Studio reload retained the authenticated merchant and selected private environment. Saved baseline b0055539 and comparison 37ab55ad retained their original observations and order IDs across that restart.

The public profile editor now shows readable enabled language names and the shared country search, including Deutschland. The catalogue-preparation grid shows all five enabled content languages and actual saved-image state. For example, the notebook has four completed languages and missing Italian; the test chair has only English and no saved image. Bild ergänzen on that chair opened its correct native product's Medien editor, retaining the private environment. No image-generation request or paid retry was submitted for this check.

The existing Furniture Configurator's release view displayed its new connection diagnostics: one view, one model, four agent tools, seven declared examples and no unresolved references. Those counts describe the package; they do not by themselves prove successful execution. Its existing unsaved draft kept the scenario-launch button disabled, preserving the save-before-testing control. The draft was neither overwritten nor discarded during this read-only check.

The actual configured qwen3.6:35b model answered the German request for a notebook and pen holder under EUR 70 with Notizbuch »Stille Seiten« (EUR 28.90) and Stiftehalter »Tonruhe« (EUR 34.50): EUR 63.40 before shipping. Native cards displayed their stored images, descriptions, stock and prices. Removing the holder, opening the notebook and returning retained the one-item selection; Restore returned both cards without adding anything to the cart.

Opening the notebook card, adding quantity one and completing the ordinary native checkout produced RAC-eb78b7d7, EUR 28.90 including EUR 4.61 VAT, German product name, pickup and simulated payment. Return and full reload retained Warenkorb 0, optional personalization off and no repeated privacy prompt. Temporary device emulation was cleared before the purchase.

The original Lumen Storyfront was then tested again against this final Core, using ordinary native Chrome input in the existing incognito session. Its first Checkout action loaded the embedded privacy and contact/address steps. After rejecting optional processing, RAC-3b44ea85 completed for one German-named demo lamp, EUR 59.90 including EUR 9.56 VAT, pickup and simulated payment. Receipt return closed the iframe and showed the original bag's order-confirmed message with no item. Full reload and reopening Bag retained Nothing yet. There was no separate-checkout navigation, user manual test or security-setting change. The earlier controlled-tab blank-load observation remains recorded above.

Public homepage follow-through

A final apex check found that the React homepage is served by the separate private Experience application, not the public Core static index. Its shared four-language homepage was therefore updated in Experience PR 77, merged as e53ae44a. It explains private customer scenarios, catalogue readiness and editable native shopping, with the actual saved public comparison screenshot and links to these published instructions. The original Storyfront submodule pin, identity, creation, provider and checkout code are unchanged.

The exact committed tree passed bun run check in an isolated temporary export: TypeScript, 255 module boundaries, 951 localization keys, 249 tests in 58 files with 1,254 assertions, and the production frontend/team build. GitHub's separate private application jobs did not start: the account annotation reports failed recent payments or an insufficient spending limit. No test step ran there, no workflow control or billing setting was changed, and that failed admission is not labelled a successful CI run. Public Core CI and documentation deployment passed independently.

The existing Northflank native build adorable-mist-5131 from e53ae44a succeeded in 1 minute 21 seconds, including both original Astro applications and the shared homepage. It used the existing Dockerfile, runtime configuration and persistent volume; no service, provider or credential was added.

A later Core observation showed golden-horse-2498 from main 07812240, activated at 17:40:48 CEST, with 1/1 health checks and zero restarts. That source has the same complete runtime tree as the originally accepted f74d8ff9 artifact above.

The Experience artifact activated at 17:50:35 CEST, deployment vendune-experience-675fc75c96. After its ordinary startup the pod was Running, 1/1 checks passing with zero restarts. A request during the replacement briefly returned no healthy upstream; the healthy-state public reload then displayed the new section and the real comparison image. English and German rendered copy was inspected. At 390 × 844, the document stayed at 390 px and the section at 350 px; the image loaded at its native width of 2,060 px. Device emulation was cleared. Its three cards link to the simulator, native catalogue preparation and shopper UI guides, while the evidence link opens this acceptance.

After the Experience restart, the original Lumen page was fully reloaded and its retained one-item bag opened the embedded native checkout on the first action. The new cart displayed its optional privacy choice; optional processing was rejected for this purchase. RAC-b39c3fe0 completed for one German-named demo lamp, EUR 59.90 including EUR 9.56 VAT, pickup and simulated payment. The receipt return closed the iframe and emptied the original bag. A second full page reload and reopening Bag retained Nothing yet. This final repetition used ordinary native input with synthetic contacts and no user manual test.

Catalogue-based profile drafts: local follow-up

The follow-up adds a reviewed profile preview to the existing customer workbench. Native HTTP/PostgreSQL verification checks the bounded product sample against the actual public catalogue projection, restricts a second channel to one lamp, verifies HTTP/MCP equality and confirms zero model requests during suggestion preparation. The existing suite also repeats simulated purchases, exact-budget refusal, stock failure/repair, cold restart after a lost checkout acknowledgement, permission withdrawal and cancellation/lease expiry during native lock waits.

The new field projection has separate Rust checks for allowed fields, bounded size, wrong currency, invalid prices and zero/two/three-digit scales. UI regressions cover preview-before-adoption, retained merchant brief, explicit start, channel replacement, empty samples and enabled-language fallback. A German-interface request also prepares the French-only channel and completes its French-profile native purchase; English profiles are rejected before queueing. Public deployment of this follow-up remains separate from the earlier accepted release above until recorded here.

Actual local model and browser result

Chrome used the real local Rust application and a new disposable PostgreSQL shop. The suggestion preview showed six sampled channel products; adoption retained the merchant's original task and populated editable goals, budgets, native languages and destinations. Explicit start used the existing agent-worker with actual Ollama qwen3.6:35b, not the synthetic provider fixture.

• The price-conscious profile stopped: Field Notes costs EUR 12.50 and its suggested budget was EUR 12.49. No order was created for that profile.
• The first-time German profile purchased the notebook for EUR 12.50, native private order ID 69f57816331f40579b6e2e1e14d789e2.
• The focused Austria profile purchased the desk for EUR 452.77, native private order ID e996ee766d6b432bb774df786791a5cc. The engine's total differs from the model's earlier EUR 449 explanation; the native order is authoritative.

The private database contains exactly two orders and the parent live shop zero. The 390 × 844 browser view has document width 390 px and suggestion width 358 px; the temporary viewport override was cleared. These are local acceptance captures, not a public rollout or real-customer conversion estimate. The full frontend suite passes 816 tests across 140 files.

Public catalogue-profile acceptance

The follow-up from PR #203 was tested on the actual public Core and Studio. The exact feature image was activated while the final required GitHub run continued. That complete run passed on 2bb5a99f at 20:27:09 CEST; regular protected merge followed at 20:27:28 CEST as 35755608472bff2690d77f76f7317bf146100084. Git comparison confirmed that the complete feature and merge trees are identical (fb25173baf3a24e49c4f009a7ef6c0a122178018).

The existing authenticated Studio session survived the rollout and full reload. Inside private sandbox stage-62ddd2bbd80642e38c4f, the native channel wizard created FR Customer Scenarios 22 October, ID shop_72a18e285109, with only French enabled and only the notebook product 624eff3d-b57a-4d40-aa11-24c73f4e42e0. This channel was not published to the live shop. The Studio interface stayed German throughout.

Testkunden aus meinem Sortiment showed three fictional profiles from the one native product, Carnet « Pages silencieuses ». The preview and explicit adoption retained the merchant's original instruction. All shopper profiles used French; Germany and Austria came from the enabled destinations. Only the later Kunden testen lassen action started the actual model tasks.

Saved cohort 4b450a8a-31bd-4f04-9284-f51a9304cb7e produced these outcomes:

The native private order grid increased from four existing orders to exactly six; both new IDs appeared alongside the original four. Opening the Austria order showed quantity one, the French product name, the Austrian destination, EUR 29.14, simulated payment, pickup and the native order.placed Flow history. The model's earlier prose mentioned EUR 28.90; the engine's country-specific EUR 29.14 total is authoritative. These fictional addresses do not establish physical shipping-provider validation or real settlement.

After a complete browser reload, selecting the saved cohort retained its original two order IDs, journals and outcomes. Diese Kunden erneut testen restored the original instruction, all three profiles and budgets, French, Germany/Austria, without starting another cohort. No additional purchase was requested.

At 390 × 844, document width stayed 390 px and the suggestion panel measured 358 px. The temporary device override was cleared afterwards.

Local verification on this exact source includes the full 816-test frontend suite, TypeScript, production build, four-language localization, Rust format and clippy, native projection unit tests, formal contracts and mutation negative controls. Actual PostgreSQL integration repeated scenario cancellation, lease/ rights guards, cold restart, HTTP/MCP parity, native channel management, international commerce and tenant isolation. Its German-interface request prepared and purchased from the French-only channel, while English profiles were rejected before queueing. The latest-head full CI run independently passed Rust, formal, differential, frontend, native PostgreSQL/provider, signed runtime, transaction-pooler and coverage gates. It took 47 min 37 sec; native integration accounted for 34 min 40 sec. The earlier successful run on 1a60d6b8 is retained as prior evidence, separate from this final result.

Regular main rollout

The existing main-triggered workflow a9bbbf4b-8d1e-4c4a-aef7-e77e12130d94 then rebuilt exact main 35755608 as jovial-steel-7373 (8 min 56 sec), completed disk snapshot 26-10-22-183633-backup (1 min 12 sec) and ran the same image in the existing migration-only job acf281e1-917f-41a2-a931-27cbbf63bd9c successfully (25 sec). Core deployment vendune-core-76ff6f475d, created at 20:38:11 CEST, reached Running with 1/1 health checks passing and zero restarts. The preceding pod terminated normally. The full workflow completed successfully at 20:39:20 CEST. No resource, provider, credential or permission setting was added.

A complete public Studio reload retained the authenticated private workspace, all three saved cohorts and the French-only channel's language restriction. Reopening the accepted cohort retained the budget refusal and the same two order IDs. No customer test was started again after the main rollout.