Connected agentic commerce implementation plan

For agentic workers: Use superpowers:executing-plans in this existing isolated checkout. Execute continuously; the user already authorized implementation.

Goal: Connect merchant setup, app changes, shopper intent, testing and migration with a persistent, bounded customer-agent scenario simulator.

Architecture: Existing Rust domain operations remain authoritative. Simulator runs extend agent_runs/tasks; UI consumes native capability contracts. App-owned implementation stays in its package, and publication stays with staging.

Tech Stack: Rust, PostgreSQL, React/TypeScript, existing inference and MCP registry.

Spec: 2026-10-22-agentic-commerce-spec.md.

Global constraints

No parallel queue, commerce engine, secrets store or per-app database tables.
Current actor/domain rights, tenant isolation, private stages and native revisions.
Interface vocabulary EN/DE/ES/FR; content uses enabled shop languages/fallbacks.
Simulated payments only; no outbound mail or paid booking from simulation.
No claimed conversion uplift or browser coverage inferred from model prose.
Read scoped AGENTS.md; preserve Rust production runtime and structure checks.

Review focus

An injected product description must never cause a privileged tool or live write.
Partial cohort creation and a lost checkout response must resume without duplicates.
Changed channel/app configuration must invalidate stale scenario evidence.
Switching shop/account/environment must fence late UI responses and saved reports.
Failure, uncertainty, cancellation and successful completion must remain distinct.

Task 1: Durable customer-agent scenarios

Files: src/agent_runs/scenarios/, src/agent_runs.rs, src/agent_runs/{worker,contract}.rs, native capability registry, prompts/, scripts/agentic_scenarios.py.

Interfaces: bounded scenario creation/inspection via native operations; existing run status/resume/cancel handles individual workers. Native catalogue/cart/checkout owners execute approved simulation actions. Reports contain observations, not tokens.

[x] Add failing real HTTP tests for private-only admission, cohort, native cart effects, comparison, idempotency, stop/restart and tenant/role denial.
[x] Implement profile/step schemas, native action adapter and durable task execution.
[x] Reuse current model accounting and task fencing; run the affected native suites.
[x] Commit the tested backend slice (PR 195).

Task 2: Visual simulator workspace and entry points

Files: frontend/src/admin/scenarios/, knowledge/workspace/designer entry owners, frontend/tests/unit/.

Interfaces: use Task 1 operations; shared environment, channel, locale and API client.

[x] Add tests for editable profiles, progress, stop/continue, comparison and scope switch.
[x] Implement compact cohort cards, action timeline, baseline comparison and direct repair links. Persist selection via saved server runs, not secrets in local storage.
[x] Verify localized responsive UI and native browser flow; commit (PR 195).

Task 3: Connected catalogue and personal daily work

Files: catalogue intake, shared task progress, frontend/src/admin/workspace/, src/agent_runs/, native work-template UI/tests.

Interfaces: existing intake, translation/media job IDs, native products and staging; Task 2 receives selected private channel/work context.

[x] Test incomplete products and failed media jobs through reload and selected release.
[x] Connect source/text/image/review readiness and scoped follow-up actions in one view.
[x] Connect completed work to existing Flow templates and contextual task results.
[x] Run native/UI regressions and commit.

Task 4: Complete app connection diagnostics and guided functionality

Files: existing app discovery/developer and package owners; native App Studio UI.

Interfaces: installed immutable manifest, native operation registry, existing configuration/order history and scenario entry point; no alternative app schema.

[x] Test one material/configuration app through native edit, cart, order and upgrade.
[x] Surface missing connections with repair links and a guided test-before-release path.
[x] Preserve package-owned code/settings, current rights and historic order snapshots.
[x] Run app and private publication regressions; commit.

Task 5: Intent-driven shopper workspace

Files: existing concierge/experience owner and storefront selection components.

Interfaces: registered visual blocks and authoritative native product/cart totals.

[x] Test a budgeted multi-product request, rejected selection and editable cart handoff.
[x] Connect intent, visual comparison/configuration and native checkout; keep UI and customer-agent actions aligned without model-supplied executable frontend code.
[x] Verify native browser purchase and scope/language regressions; record the native German simulated receipt, return, empty-cart reload and retained privacy choice.

Task 6: Migration-to-preview, documentation and delivery

Files: migration app/service existing owners, docs/status.md, related feature docs, public site/README entry links and existing CI owners when affected.

Interfaces: existing source jobs/reconciliation, private target shop and release.

[x] Test imported source -> comparison -> private preview -> chosen release.
[x] Connect preview, missing-field repair and scenario testing in migration UI.
[x] Update truthful feature docs, tutorials and public entry points.
[x] Run relevant full checks and fresh whole-branch review; fix important findings.
[x] Land and deploy the verified artifact using the existing authorized path, then repeat public simulated acceptance and record what was actually delivered.

Progress

22 October: fetched origin/main (defac3dc), clean existing isolated checkout, branch feat/agentic-scenarios. Existing tasks, approval, model checkpoints, native carts, experiments and stage admission inspected. Implementation started.
Local simulator acceptance is complete: real qwen3.6:35b customer decisions, two native private purchases, exact-budget rejection/abandonment, reload, comparisons, private order grid and mobile layout. See the recorded evidence.
Existing connected app/configuration, agentic cart/UCP, catalogue, staging and five-source migration suites passed again. Checked items above may reuse those owners; they do not mean a second implementation was introduced in this branch.
The connected per-product language/image readiness projection and app reference diagnostics are implemented in a separate follow-up. Native catalogue reload, failed-image persistence, foreign scope and selected publication passed; Chrome opened the failed-image repair path. See the same evidence document.
Independent critical-path review completed. Late UI selection, incomplete snapshot comparison, cancellation and post-lock lease-expiry findings were fixed and regressed. PR 195 passed complete CI, merged and was deployed; hosted stock-failure/repair, exact-budget rejection, two purchases, saved comparison and original embedded checkout passed. The connected catalogue/app and visual shopper follow-ups landed in PR 199, with readable profile destinations in PR 200. The final f74d8ff9 artifact is public. Catalogue repair links, app diagnostics, saved simulator comparison, visual selection and native purchase RAC-eb78b7d7 passed there; original Storyfront embedded order RAC-3b44ea85 returned to an empty bag after full reload. Payments remained simulated.