Standing Order Agent / Technical project guide
Technical project guide

A grocery workflow the user can see, trust, and control.

Standing Order Agent prepares a Stop & Shop cart from household preferences, rules, and past behavior; selects an allowed delivery window; then stops for review before the user completes checkout.

12editable diagrams
14technical pages
3runtime entrypoints
2delivery phases
The product

What we are building

A scheduled helper, not an invisible purchasing bot. The user sees the plan, the live progress, every exception, and the exact review snapshot.

Prepare

Build the right cart

Combine recurring product preferences, bounded purchase history, saved lists, current availability, and deterministic household rules.

Schedule

Reserve an allowed window

Prepare early enough to find a preferred slot, rank acceptable fallbacks, and never silently choose an unapproved day.

Review

Explain what changed

Freeze exact products, substitutions, quantities, prices, fees, total, and delivery slot into a reviewable snapshot.

01
Human confirmation is the first release boundary

Live Phase 1 ends at manual final checkout. The architecture contains a future approval gate, but checkout remains disabled by deployment policy until separate authorization and security gates are met.

System map

How the pieces connect

A modular monolith keeps the prototype practical. Provider and retailer boundaries remain replaceable through explicit ports and adapters.

High-level architecture
High-level architecture
The API owns user traffic, the worker owns browser execution, and PostgreSQL holds the durable workflow. Retailer, browser, and LLM details remain behind adapters.
Module map

Explore each responsibility

Every card opens a focused page with an editable draw.io diagram, contracts, data, failure behavior, and implementation notes.

01

Intelligence

Provider-neutral orchestration for GPT-5.6 and Claude Opus 4.8 with structured results and evaluation gates.

Open module
02

Browser control

Isolated local, visible, remote-CDP, or managed-cloud browser sessions with evidence and takeover.

Open module
03

User preferences

Versioned household budget, delivery, automation, substitution, notification, and pause settings.

Open module
04

Rules engine

Deterministic safety, item, basket, slot, and approval policies with stable reason codes.

Open module
05

Preferred products

Household product intent from manual choices, Past Purchases, retailer lists, and confirmed suggestions.

Open module
06

Scheduling & jobs

Timezone-aware recurrence, durable dispatch, deduplication, retry, skip, and pause.

Open module
07

Retailer integration

A stable Stop & Shop commerce contract over reviewed browser actions and authorized observed traffic.

Open module
08

Order orchestration

The durable state machine that coordinates cart preparation, review, approval, and safe confirmation.

Open module
09

MVP implementation

The smallest live solution that proves Past Purchases, deterministic rules, cart preparation, and dashboard review.

Open MVP section
10

Docker & local Chrome

Deploy the prototype to a Docker VPS while a constrained reverse SSH tunnel reaches a dedicated Chrome profile on the user's computer.

Open deployment guide
11

Implementation plan

A module-by-module path from architecture foundation through assisted live cart and private pilot.

Open plan
Product phases

One workflow, two purchase boundaries

The first phase proves useful automation with low risk. The second phase adds explicit approval-bound confirmation only after separate release gates.

Phase A · Simulate

Build every domain module against fake browser, retailer, LLM, and notification adapters.

Architecture foundation

Phase B · Assist

Prepare the live cart, select the slot, freeze the review snapshot, and hand control to the user.

Prototype target

Phase C · Confirm

Future approval-required checkout with a snapshot hash, amount limit, idempotency, and confirmation verification.

Authorization required
Before implementation

Five decisions that change the build

1

Account scope

Single private household or multi-household onboarding from day one?

2

Browser runtime

Confirm the Docker VPS → reverse SSH → dedicated local Chrome topology for the first live run.

3

Checkout boundary

Confirm that Phase 1 always stops before the retailer’s final order action.

4

List integration

Read-only Stop & Shop list import or no list integration in the first live test?

5

Model policy

Primary provider, allowed fallback, provider retention, and household-data boundaries.