The shop is a pure-Python rewrite of Princeton's WebShop. It is on by default and needs no setup step:

uv run launch.py                                  # shop at /onlineshop
uv run launch.py apps.onlineshop.enable=False     # turn it off entirely
uv run launch.py apps/onlineshop/content=default  # chrome only: no products, no shop

config/apps/onlineshop/content/webshop.yaml holds 999 products converted from the WebShop item dump, and is the shop's default content pack. The catalog is a property of the pack rather than of the app, so a pack with an empty products list yields no /onlineshop route and no start-page tile — the app is absent rather than empty.

The webshop pack sets product_images: hotlink, because it is the one catalog whose image URLs point at real photos. The default configuration therefore references Amazon's CDN. On a node without egress the images fail silently while the page still returns 200, so for eval runs there:

uv run launch.py apps.onlineshop.product_images=glyphs

tests/test_no_egress.py records this as the one content-driven entry in its external-host allowlist; every other entry is a styling dependency.

Varying the shop

It composes the same way as every other app:

uv run launch.py apps/onlineshop/layout=grid            # default | grid | compact_table
uv run launch.py apps/onlineshop/content=german         # or long_descriptions, adversarial_descriptions, misleading_descriptions
uv run launch.py apps.onlineshop.theme=solarized        # or apps/theme=solarized globally
uv run launch.py apps.onlineshop.products_per_page=5

The text variations (german, long_descriptions, adversarial_descriptions, misleading_descriptions) layer on top of webshop, so they change the chrome and keep the catalog. default.yaml is chrome only (title, promo text, category labels, currency) and deliberately carries products: []. Override apps.onlineshop.products from your own content file to swap the catalog wholesale.

content/fixture.yaml is a small mechanical catalog (18 generic products, no prose) that exists so the test suite has something fixed to run against.

The product seed

A content pack is plain Hydra YAML. set_environment clears and re-seeds all four tables from it on every launch and every reset, so the catalog is a variation axis like any other text in OpenApps — not a fixture you migrate.

# @package apps.onlineshop
defaults:
  - default          # inherit title, promo text, category labels, currency

products:
  - sku: 'furn-desk-205'          # stable id; used in URLs and in cart/order state
    title: 'Standing Desk'        # display name, and the top-weighted search field
    price: 649.00                 # float, in `currency_symbol` units
    category: furniture           # must be a key of `categories` in the pack
    breadcrumb: ['Furniture', 'Desks']   # crumbs on the item page
    rating: 4.2                   # 0-5, rendered as stars
    options:                      # option name -> selectable values; may be {}
      finish: ['walnut', 'natural']
      width: ['120 cm', '140 cm']
    bullets:                      # short feature list; may be []
      - 'Electric height adjustment'
    description: 'Sit-stand desk with a programmable controller.'
    images: []                    # optional; only drawn under product_images=hotlink

# Optional starting state. Both are seeded after `products`, and a line
# referencing an unknown sku is dropped with a warning rather than crashing
# the launch.
cart:
  - sku: 'furn-desk-205'
    options: {finish: 'walnut'}   # values not offered by the product are dropped
    quantity: 1

orders:
  - order_id: '0efcf51f'          # omit to get a random one
    name: 'Alex Morgan'
    address: '14 Bridge Street'
    date: '2026-08-24 19:54:34'
    status: 'Delivered'
    items:
      - sku: 'furn-desk-205'
        options: {finish: 'walnut'}
        quantity: 1
    # total: 649.00               # omit to compute from current prices

Option values are indexed for search alongside the title, bullets and description, so a query for walnut finds the desk above even though the word appears nowhere in its text. Cart lines are keyed on sku plus a canonical form of the chosen options, so the same product in two finishes is two lines, and the same options in a different order is one.

Rebuilding the catalog

scripts/fetch_webshop.py generated the committed pack and can regenerate it — a different dump, more products, synthesized ratings. Nothing requires it to run, and it overwrites webshop.yaml in place, discarding hand-edits. It pulls from a HuggingFace mirror of the item dump.

uv run scripts/fetch_webshop.py --inspect          # print the real schema; writes nothing
uv run scripts/fetch_webshop.py                    # rebuild the committed pack as it stands
uv run scripts/fetch_webshop.py --file items_shuffle.json --limit 5000   # bigger (~1.5 GB download)
uv run scripts/fetch_webshop.py --synth-ratings    # stable, invented ratings when the dump has none
uv run scripts/fetch_webshop.py --out /tmp/webshop.yaml                  # e.g. to diff two conversions

It downloads through huggingface_hub when that is importable, so re-runs hit the shared HF cache.

On prices. About one record in eleven has an empty pricing field, and a few ranges open at a $0.01 placeholder. A product nobody can be charged for is useless to a shopping task, so a missing price is filled with the median price of the products in the same category, and sub-dollar values are ignored when reading a range.

On field names. The dump has been re-exported several times and the casing is not stable — the current one uses name, full_description, small_description and pricing. Each output field is read from a list of candidate keys (_FIELDS in the script), so if a conversion comes out empty, run --inspect and add the real key to that list.

On options. WebShop's customization_options is {"Color": [{"value": "Bath Ball", "url": ..., "image": ..., "price": ...}]} — the sibling keys hold amazon.com links. Only value is carried across, and the script aborts if any URL survives into a text field.

Product images

apps.onlineshop.product_images chooses what gets drawn:

Under hotlink, each <img> carries an onerror that swaps the whole carousel for the glyph, so a failed fetch degrades to line art rather than a broken-image icon. The carousel is one hidden radio per slide and a clickable dot per image — real elements an agent can click, no JavaScript. layout=compact_table stays imageless either way. An unrecognised value falls back to glyphs, so a typo in a sweep override cannot silently start hotlinking.

Glyphs. _GLYPHS in src/open_apps/apps/onlineshop_app/main.py holds ~35 hand-written shapes on a 100×100 box; _GLYPH_KEYWORDS maps title keywords onto them, first match wins, on word boundaries with an optional plural (pen hits "Pens" but not "Open"); _CATEGORY_GLYPHS catches the rest. The hue is derived from the sku, so a product looks identical on every page. To review a catalog's art at once:

uv run scripts/render_glyph_sheet.py            # the webshop pack; or pass `fixture`
open /tmp/glyphs.html

Inspecting the shop's data

Each launch writes a fresh SQLite database under the run's log_outputs directory, so the newest one is the run you were just clicking through:

DB=$(ls -t log_outputs/*/databases/onlineshop.db | head -1)
sqlite3 -header -column "$DB" "SELECT * FROM cart_items;"

Four plain tables — products, cart_items, orders, order_items — plus the products_fts* search index. The same state is served as JSON at /onlineshop_all, which is what rewards are computed from.

How it works

FTS5 is compiled into Python's sqlite3 module, so BM25 ranking costs no dependency. User input is accent-folded and tokenised, and each token is quoted before being OR-ed into a MATCH expression, so FTS5 operators typed into the search box (*, NEAR, a stray quote) are matched literally, and creme and crème find the same products. Tokens are OR-ed rather than AND-ed so a query returns its best partial matches.