Basefox v0 · Section A follow-up · 25 Aug 2026
Card system, rebuilt Home, and product-aware search
Acting on the two notes from the Section A test pass: the Home redesign (mockup 1), the card system (mockup 2), and the search-intent problem.
What shipped
1. The card system is real code, not a picture
All six types and three densities from mockup 2 exist as one component. The five-part anatomy (eyebrow, headline, descriptor, visual, status) is the same shape at every density, and boards now compose cards instead of hand-rolling rows.
| Type | Leads with | What triggers it in the data | Live in v0? |
|---|---|---|---|
| A · Brand offer | the value | a usable best route | yes |
| B · Exchange event | the event | a live ExchangeEvent | yes |
| C · Category event | the category | a live event seen from a category board | yes |
| D · New benefit | what’s new | a newRoute change inside 14 days | yes |
| E · Ending soon | urgency | endsOn or event ends inside 3 days | events only |
| F · Fallback | what we do know | no usable route, or a verified negative | yes |
Type E fires from exchange events today. It cannot fire from a brand offer yet, because nothing extracts an offer expiry — see the gap below.
2. The visual asset hierarchy, with the fallback actually built
Four tiers, best first. The app asks for the best it could have and takes the best that exists, so a missing file costs richness and never correctness.
| Tier | Source | Where it comes from | Ships today |
|---|---|---|---|
| 1 · Campaign | from source | registry/assets/events/ | no files |
| 2 · Brand | curated by Basefox | registry/assets/brands/ | no files |
| 3 · Category | Basefox library | SF Symbol per category | yes |
| 4 · Generic | Basefox abstract | drawn in-app | yes |
larchpipe assets sync copies whatever is in the registry into the app’s
asset catalog, names what is still missing, and flags files nothing references. It reports
0 of 13 today and is safe to run against an empty directory.
3. Merchant signature accents
Sampled from your card-system mockup and assigned per brand in brands.yaml.
Every accent comes from this fixed set, so no card can introduce a colour the rest of the
app does not already speak.
4. Home, rebuilt on the cards
5. Search now answers a product query
Your instinct was right, and the old behaviour proved it: typing “hiking boots”
matched nothing and dropped to the empty state. The index still holds brands and categories
only. A new productTerms lexicon bridges the two.
- A term may sit under several categories on purpose. “hiking boots” is footwear and outdoor, so both boards are offered.
- Each result row names the term that matched. That is how the brand/category model gets taught without a tutorial.
- Ranking: category name beats synonym beats product term; within each, exact beats prefix beats word-start beats substring.
6. Sharp filters on the category board
The old board had four mutually-exclusive chips that mixed sorting with filtering, so choosing “Online only” silently discarded the discount ranking. They are now two separate controls, and the filters combine.
- Sort: Best discount · Recently verified · A–Z.
- Filters: Online · In-store · Stacks · Verified <30d · Has a discount · a verification-method menu (SheerID, ID.me, GovX, ID at register).
- Filtering happens at the route level, then the best surviving route is re-picked. A brand whose best route is in-store still appears under “Online”, showing the online route — the one being asked about.
- A running
6 of 9 brandsline states the effect, so a filter that hides most of the board says so instead of looking like a thin category.
The gap you should know about
Nothing produces an image, and the card system is largely image-driven
Before this work, grep for image/logo/asset/photo across
/pipeline, /registry, and the spec returned zero hits. The
“Visual Asset Hierarchy” box in your mockup described a system with no supply
of any of its four tiers.
I built tiers 3 and 4 so the app is complete and intentional-looking today, and wired
tiers 1 and 2 end to end. What remains is the licensed material itself: 10 brand
lockups and 3 campaign heroes. Specs and filenames are in
registry/assets/README.md; drop the files in, run
larchpipe assets sync, rebuild, and every card upgrades on its own.
Card type E cannot fire from a brand offer yet
Route.endsOn now exists in all four halves of the contract, but no source
extracts it — the Tier-2 extractor never asked for an expiry date. I left it
null across the bundled snapshot rather than inventing dates in shipped
sample data. Teaching the extractor to pull expiry is the next step if
“Ends tomorrow” matters for launch.
Three things in mockup 1 I did not build
- The bottom tab bar (Home / Categories / Saved / Alerts / More). You chose the card-library scope, not the everything scope. Saved and Alerts have no data model, no store, and no spec section, so they are a feature rather than a restyle.
- The hamburger at top-left. There is no drawer to open, so the profile button remains the only header control.
- “See all” next to Live at your exchanges. It needs an events board that does not exist. All three live events are reachable by scrolling the row.
What changed, by file
| Area | Files | Change |
|---|---|---|
| Contract | brands.yaml, categories.yaml, events.yaml, models.py, registry.py, dataset.schema.json, Models.swift, dataset.json | Six new fields carried through all four mirrors: accent, logo, glyph, productTerms, hero, endsOn |
| Card system | Views/Cards/ (new: CardModel, CardView, CardBuilders, AssetResolver) | The six types, three densities, the anatomy, and the four-tier resolver |
| Home | HomeView.swift | Hero event cards, circular glyph tiles, compact change feed, one card tap handler |
| Search | Domain.swift | CategoryMatch with the matched term; scored name/synonym/product-term ranking |
| Filters | BoardFilter.swift, FilterBar.swift, CategoryBoardView.swift (new + rewired) | Sort split from filters; route-level filtering with best-route re-pick |
| Assets | larchpipe/assets.py, cli.py, registry/assets/ | larchpipe assets sync plus the file spec |
| Tests | CardSystemTests.swift, test_assets.py | 27 new Swift tests, 11 new pipeline tests |
| Spec | basefox-spec-v03.md | New §Card system; §Screens, §Design system, §Data model, and the search order updated |
One tidy-up worth naming: FlowChips was left with no callers after the
category chips became tiles, so it is gone. FlowLayout survived and moved to
Components.swift with the other shared primitives.
Review
What this buys
- Boards B, C, and D inherit the card system for free from here. The route board and occasion board are the obvious next candidates.
- Adding a brand image is now a one-step change with no code involved.
- The contract stayed in lockstep across YAML, pydantic, JSON Schema, and Swift, and the round-trip test still passes, so the two halves cannot drift quietly.
Where it is thin
- The hero layout is stacked, not side-by-side. Mockup 2 draws hero cards landscape (copy left, product right); mockup 1 draws Home’s event cards tall. I built tall, because that is what Home needs at 174pt wide. A landscape hero for a full-width board will need a second variant.
- Tier-1 photography may fight the copy. The copy block sits on the card ground above the visual, so it cannot collide — but a hero image cropped into a short band may read worse than the mockup’s full-bleed treatment. Worth re-checking the moment a real image lands.
- The product lexicon is hand-written and short. Roughly 60 terms across 7 categories. It will miss real queries, and there is no telemetry to learn which. Growing it is editorial work, the same as the brand registry.
- Filters have no persistence. Leaving a board and returning resets them. That is probably right for v0 and worth revisiting once boards are visited repeatedly.
- The change feed’s event rows fall to tier 4. A
ChangeItemcarries arouteIDbut no event id, so aneventOpenedrow cannot reach its event’s category glyph and shows the house mark instead.
Confidence
High on the parts with a clear right answer: the four-tier resolver, splitting sort from filters, filtering at the route level, and keeping the contract mirrored. Lower on two judgement calls that only a real asset will settle — the single stacked hero variant, and how a licensed hero image behaves in a short band. The lexicon’s coverage is a guess until someone uses it.
Decisions waiting on you
- The 13 asset files. Nothing else blocks tiers 1 and 2.
- Offer expiry. Teach the Tier-2 extractor to pull an end date, or accept that “Ending soon” is an exchange-event state only in v0.
- The tab bar. Still unbuilt by choice. Saved and Alerts need a spec before they need code.
- Discount vernacular. Your mockup writes “10% OFF”; the app writes “−10%” everywhere. I kept the app’s form for consistency. Changing it is a one-line change in the card builder plus a sweep of the boards.