Repository navigation
feat(TASK-24893): creator invite pages on /go with a three-page test for Nita - #4107
Conversation
G2 is the conversion reviewers' first pick for the landing-page test: a frame of the Reel the visitor just left, with ten words and one button. The sheet keeps a 92px band free at the bottom, because in-app browsers lay their toolbar over the viewport there. The frame is dimmed so nobody taps it as a paused video. TASK-24893
…perties A creator's /go link shows one of several pages, so the test needs a weighted, sticky pick and one set of labels on the tap event. These parts are pure so the server registry and the browser shell share them. TASK-24893
The shell makes the variant sticky from the browser, because a page render cannot set cookies. The button is the single place that holds the engine's link and sends creator_page_cta_clicked, so the one-tap iOS swap is one edit. No view event: the engine's tracked_link_visit is the only view count. TASK-24893
selectCreatorPage is the seam the /go engine calls (go-links contract v1). experiments.ts is the file Peanut edits to change a page or the split. TASK-24893
…cookie Pages render only behind /go once the engine PR lands, so reviewers need a route that shows each variant from a fixture. Web-only: the native export skips it. TASK-24893
The repo sizes icons through the Icon size prop, not through CSS, so the page no longer overrides it. TASK-24893
…button inside Nita's followers arrive from Instagram already knowing her X post. A1 shows that post large and puts the store button in the card, so the tap needs no scroll at 375x667 with an in-app toolbar over the bottom. A pinned button takes over whenever neither in-page button is on screen. Copy differs from the mockup where product facts or the copy rules required it: no card line, no 'the app already knows' claim, no exit link to X, and the payer limits per currency are pointed to the app.
E1 tests a loud benefit headline with her real post as proof against the near-empty G2. The poster scales with the column width instead of a fixed 390 px canvas, and the button sits on a paper strip pinned to the bottom of the visible viewport, because the review found it under the in-app toolbar. The copy drops the saved-invite claims, which no device check backs.
selectCreatorPage called pickVariant inside the lookup, so each variant was compared with a new random number: a1 33%, g2 22%, e1 15%, and 30% of new visitors got no page. The tests pinned Math.random to one value and could not see it; they now feed a changing sequence and count 6,000 real draws. TASK-24893
Both are link placements (th, li), and their visitors were filed with Safari and Chrome under none. The TikTok test also matched trill inside other words. An iPad with the desktop user agent stays other, now stated. TASK-24893
…p as a beacon The shell was a client component that started every visitor on the neutral label and swapped it after hydration, which is slow in the in-app browsers these pages are opened in. It is now a server component that reads the request user agent, so the first HTML has the right label and store icon. The cookie write and the context moved to CreatorPageProvider. The shell also hands the experiment id to the browser. The id used to come from a second file that had to be edited together with the weights; now experiments.ts holds id and weights, and page-components.ts holds the components, with a test that keeps the two in step. The tap left through PostHog's batch queue and was lost when the page unloaded without a flush; it now leaves at once as a beacon. It carries only the agreed keys: signup_journey_id, not the other two journey properties. The store icon carries custom-size at 24px, so .btn svg no longer shrinks it to 18px and adds a margin. The preview route is a server component that renders a page the way /go will. New tests render each variant on the server and pin the href, the language, the legal line and the per-device label; the navigation assertion can now fail. TASK-24893
… large text The sheet kept a 92 px empty reserve under the legal line, which pushed the Reel still up until only her chin showed at 667 px and below. Only the button has to clear an in-app toolbar, so the reserve is now the foot itself (invite line, disclosure, legal line) and the still is anchored on her face. - The column grows instead of clipping, and the headline may wrap, so Android text zoom 130% and a 20 px root font scroll rather than cut the headline. - The tick beside the invite line is gone; it also broke the iconOffScale ratchet. The store icon keeps the 24 px the shared button gives it. - Ruled labels and the shared disclosure line by the legal line. - The Reel still was a 2x upscale; it is now the 425x758 source at native size, with a stronger dim so the softness reads as a backdrop. - Copy comments point to engineering/ui/creator-pages.md.
The real page will render at /go without the wallet providers. Its dev preview ran on the app tree, so every screenshot carried the app's connectivity toast and globals the real page will never have. Only /dev/creator-pages and its children move; other dev routes and native builds stay on the app tree.
The root layout locks zoom for the app. A landing page opened from Instagram must let a reader zoom into small legal text, so this route overrides it.
…at is true Review round 1 on Nita's poster page. Layout: with Android text zoom the headline ran off the column and her post rode over it, because every piece was absolutely placed and sized from the column width. The headline is now drawn from glyph outlines, so it fits its box at any text size, and the poster is a column in the flow: label, headline, post, then a hidden copy of the pinned strip that keeps exactly the strip's height free. The collage pieces sit under the strip and can no longer cover the button. Strip: on a short screen it is the button and one line. On a taller one it keeps 94 px under the button, so a 90 px in-app toolbar laid over the page covers the lines and not the button. Every line a visitor must read is at least 11 px. Copy: the chip is no longer in her voice, the card is gone from the steps and the fine print, the verified badge and the tick after the invite line are removed, step 02 no longer promises bank details from verifying alone, step 03 says that who can pay depends on the currency, and the disclosure is the line agreed for all three pages. Font: the Ubuntu Sans Mono subset kept the original family name, which its licence does not allow for a modified file. The labels use the device monospace and the file is deleted.
…ne pinned-bar rule The headline touched the mascot on every 375 px phone and sat under it with a 20 px root font, because the mascot was absolutely placed. It now shares a grid row with the subhead, so text wraps beside it at any width or text size. The pinned bar covered the post card and its own button on short Instagram viewports, and two store buttons showed at the handover. One rule replaces the thresholds: the bar shows only while no part of an in-page button is on screen, and it starts hidden. Copy follows the review rulings: "Abrir en el App Store", "tu usuario", "tu identidad", "la app te lo muestra", the shared made-by line next to "Peanut no es un banco", and a footnote that also owns up to the "…" cuts. The button grows instead of clipping when large text wraps its label.
Every page asked for the icon, so the opt-in prop was one more thing to forget. Tests follow the ruled iOS label.
The branch now stacks on the /go engine, which owns CreatorPageProps and GoVisit. Our interim copy in types.ts could drift from it, so it goes and every file imports the engine's types.
…t labels creator_page_cta_clicked has to join tracked_link_visit on link_id, signup_journey_id, page_variant and page_experiment. Hand-building those keys beside the engine let the two events drift, and the experiment id came from our own table while the visit takes it from the handoff. goVisitPosthogProperties(page) is now the only source: the click reads the experiment from the same tl_e the visit is counted under, and gains opened_in. Our own keys stay: creator_username, device_os, in_app_browser, cta_position. The fixture builds its handoff with goHandoffPath so it cannot drift from a real one.
…gine's isGoPageLabel The engine refuses a selection whose labels fail isGoPageLabel (it throws in development and redirects in production). Our own copy of the pattern could accept an id the engine rejects, which would turn Nita's page into a silent redirect. The cookie parser and the experiment table test now use the engine's check.
The page only renders if the engine accepts our selection, puts it into the handoff and the click reads the same labels back. This test runs those steps with the engine's contract module, so a change on either side fails here before it fails on the route. It also records that no page sets link-preview metadata yet. The marketing-routes comment said the page will render at /go; it does now.
…-app browsers, set the tab title
The shell awaited headers() to pick the button label. On /go the engine
renders the page under <Suspense fallback={null}>, so an async shell put
every page in a hidden block that only a script reveals. selectCreatorPage
now reads the user agent next to the cookie and returns the variant with
deviceOs bound; the shell and the pages are plain functions.
in_app_browser gains `other` (Snapchat, WhatsApp, Telegram, the Google app,
Android webviews, iPhone webviews without Safari/). Before, all of these
were counted as `none`, the comparison group of the read-out.
selectCreatorPage returns one Spanish tab title per creator, the same for
every variant, because an in-app browser shows the title in its top bar.
Also: deviceOs() uses the wallet button's rule instead of a copy, the
fixture file says it is for the preview and tests, and the preview route
folder is named after the username like every other place.
…review gate and the real route The review found four untested behaviours: when A1's pinned button shows, whether a page renders and hydrates inside the engine's CreatorPageBoundary, that the preview does not answer on peanut.me, and that the tap event on the real /go route names the variant of the button link. The Playwright spec uses the engine's switches and its own lookup stub. It marks two checks as expected failures for A1 and E1: React streams a Suspense boundary over 12,800 bytes as a hidden block, so these two pages are blank without JavaScript until the engine's boundary changes. G2 passes.
…n-image icon The only Reel frame is 425 px wide. A DPR 3 phone stretched it 2.75x with a coarse filter, which showed stripes and blocks on skin. The file is now that frame at 2x (850x1516, 81 kB); a comment names the size the creator's original must have. The frame's name moves to its wrapper, so a failed load leaves a plain ground. The bank line keeps "banco." with the word before it, as on A1.
The preview kept the English site title, so a review of the page could not see what an in-app browser shows in its top bar. The route now reads the title from the experiment, the same field selectCreatorPage hands the /go engine.
…proved poster The outlines were Roboto Flex at weight 400. The old generator read the WOFF2 with fontkit, which does not apply the weight axis to it, so the poster headline came out thin. The data is now made from the app's own font file with the weight pinned to 1000 and kerning on. The generator moved to mono (engineering/ui/creator-pages/e1-lettering), where docs and tools about app code live, and no longer reaches into node_modules. A test measures the ink of each line, so a thin drawing fails, and checks that every headline line has a drawing. The shapes are in the HTML once and drawn three times with <use>. With the smaller path data the page's <main> is about 10 kB, under the 12.8 kB at which the engine hides server HTML until a script runs. The drop is the letters without their outline, and one em is 86 units, both as on the mockup. Short screens (up to 620 px high): the note sits under the button and the strip takes 24% of the screen, so the button stays in view when an in-app toolbar of 90 px lies over the page. Copy, from the round-2 review: the footer says the quote is word for word, cut at [...] and highlighted by Peanut; "banco.", "saldo." and "Nita." keep the word before them; step 2 says the accounts are requested after verification; the ticket says what the link does. Step titles are headings. At 320 px the chip fits on one line. The five torn-edge point lists move to e1/torn-edges.module.css: they were 626 of the stylesheet's 1,151 lines and nobody reads them. A failed image on a flat ground no longer shows the browser's broken-image mark.
The wordmark file has a dark grid ground baked in, a little lighter than the page. Under the lighten blend it showed as a rectangle. The same file with its black point raised: the ground falls below the page colour and the mark is unchanged. No new export from the creator was needed.
Round 2 review found gaps where no store button of 44px was tappable, and positions with two. The pinned bar now hides only while an in-page button is at least 95% visible (95% of 48px is above 44px), so a scroll sweep in 10px steps finds one whole button at every position from 320x568 up. The hero card is tighter at narrow widths so the in-card button is on the first screen at 320x568 and 360x640 without JavaScript, on one line. The bar puts the button on top with the invitation line under it and is 146px tall, so 44px of the button stays above a 90px in-app toolbar. Copy: the sample rows show USD and EUR instead of a bare dollar sign, the copy labels that copied nothing are gone, the disclosure and 'Peanut no es un banco.' sit on their own lines, and the post card says who added the colours and numbers.
…ript E1's <main> is now 10,309 bytes, under the 12,800 bytes at which the engine's Suspense boundary hides server HTML until a script runs. The spec still marked its two no-JS checks as expected failures, so a run would have reported them as unexpected passes. Only A1 (17,048 bytes) stays an expected failure. The comment gives the measured sizes and E1's remaining room.
…he inviter, since the invite can be lost on install
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
/chip review |
There was a problem hiding this comment.
Chip review — no blocking findings — this is not an approval
Two major findings: the A1 pinned CTA can disappear behind an in-app toolbar, and an experiment cookie persists after analytics opt-out.
Findings
-
MAJOR · src/features/creator-pages/pages/juanacervio/a1/A1StickyCta.tsx:34 · Pinned CTA disappears while the footer CTA is covered
On a 667px viewport with the 90px bottom toolbar this page explicitly accommodates, scroll until the 48px footer CTA sits around y=619–667. IntersectionObserver reports nearly full viewport intersection, so this branch hides the pinned bar, even though the toolbar covers the footer button. The page then has no usable CTA until another ~90px of scrolling. Measure the usable viewport (for example, an appropriate negative bottom rootMargin) before retiring the pinned bar, and cover the footer handoff with a toolbar-sized test. -
MAJOR · src/features/creator-pages/CreatorPageProvider.tsx:30 · Variant cookie survives analytics opt-out
A first visit with analytics enabled writes a 30-day cpv cookie. If the visitor later enables GPC/DNT or opts out of PostHog, this early return leaves the cookie intact; the next /go request still sends it and registry.ts selects the same variant from it. That contradicts the documented no-consent/no-cookie, fresh-pick behavior. Expire the cookie with the same Path=/go when eligibility is false, and do not honor an existing cookie when request privacy signals disallow it; add a consent-revocation case.
Checked clean
- Exact head and base, trusted PR metadata, /go selection and handoff, all three variant components, consent and attribution paths, and relevant tests inspected.
- Design focus styling remains inherited from the shared button class; product copy was cross-checked against the live Lexicon and product source.
- At this SHA GitHub reported successful guard, bot-approval and Vercel checks, but no unit-test/typecheck check runs; local Jest was unavailable because this detached worktree has no node_modules.
- A1 observer tests mock only intersection ratios and do not model a bottom browser toolbar; production-phone behavior remains unverified.
Security review by moonshotai/kimi-k3: 0 finding(s), marked with the model name. It reads the diff only and answers only security, privacy and money, so treat its findings as advice.
Third opinion by claude-opus: 0 finding(s), marked with the model name. It answers only product truth, missing tests and the cross-repo contract, so treat its findings as advice.
Copy check by claude-haiku-5-5: 0 finding(s), titled "Copy now wrong". Every quoted passage was checked against the file.
Exact head: 23cf3cd38c66 · Context: repo, design, product, lexicon · Took 6m
…tracking A visit with analytics allowed writes a 30-day cookie. If the visitor later opted out or switched on DNT or GPC, the cookie stayed and kept choosing the same page, against the documented rule (no consent, no cookie, fresh pick). The browser now removes the cookie when analytics are not allowed, and the server ignores it on a request that carries DNT or GPC. Found by Chip review.
|
/chip review |
There was a problem hiding this comment.
Chip review — no blocking findings — this is not an approval
The cookie opt-out finding is fixed. The previously raised A1 CTA visibility gap remains and was explicitly deferred to a real-phone/design follow-up.
Answered by a human, not raised again
These are still present as far as this review can tell. You resolved the thread or deferred the fix, so Chip does not repeat them. If one needs a follow-up, file the task yourself.
- MAJOR · src/features/creator-pages/pages/juanacervio/a1/A1StickyCta.tsx:34 · Pinned CTA disappears while the footer CTA is covered
Checked clean
- P1 remains in the unchanged A1 observer logic; the supplied author reply explicitly defers it to other work.
- P2 is fixed: an ineligible active page expires the experiment cookie with the original /go scope, and server selection ignores an old cookie for DNT: 1 or Sec-GPC: 1. Browser-only PostHog opt-out can still affect one server selection before the page expires the cookie, as disclosed by the author.
- Reviewed the six-file follow-up diff and its consent, prerender and request-header tests against the full /go selection and handoff path; no additional reachable regression found.
- At this SHA the guard and bot-approval check runs succeeded; Vercel status was pending. Local Jest was unavailable in the detached worktree because node_modules is absent.
Security review by moonshotai/kimi-k3: 0 finding(s), marked with the model name. It reads the diff only and answers only security, privacy and money, so treat its findings as advice.
Third opinion: did not run — it reads only the first review of a pull request; the first reviewer checks later rounds. This review is one reviewer short.
Copy check by claude-haiku-5-5: 0 finding(s), titled "Copy now wrong". Every quoted passage was checked against the file.
Exact head: 40c1ec738851 · Context: repo, design · Took 2m
Code-analysis diffPainscore total: 11090.26 → 11089.28 (-0.98) 🆕 New findings (43)
…and 23 more. ✅ Resolved (2)
📈 Painscore deltas (top movers)
|
🧪 UI test report — ✅ all greenSuites
📊 Coverage (unit)
⏱ 10 slowest test cases
|
…'s pages The owner ruled on 15 Nov 2026 that the pages need no disclosure line: the context makes it plain who made them. A1, G2 and E1 no longer say 'Nita usa Peanut. Página hecha por Peanut.'; 'Peanut no es un banco.' stays on every page and the page test now asserts the line is absent. A1 reaches its notes by markup order, so the selectors move up one place. E1 and G2 keep the min-height under the button, so the button stays clear of an in-app toolbar (92-94 px) as before.
Instagram's and Facebook's in-app browsers on an iPhone can draw their own controls over about the bottom 60px of the page without shrinking the viewport. 100vh and 100svh still report the full height and an IntersectionObserver cannot see the overlay, so A1's pinned bar could hide while the only in-page button sat under those controls. - A1: an in-page button counts as on screen only when it is at least 95% visible above the bottom 100px (observer rootMargin). On short screens the pinned bar now shows from the first screen. - The server marks <main> with data-lift-cta for Instagram and Facebook on iPhone/iPad, from the same user-agent detection the events use. A1's bar and closing section, G2's sheet and E1's strip then move their content up by 60px. E1 lifts by less on a screen too short to give 60px, so the strip does not cover the end of the post; its button already ends 92px above the bottom edge there. - G2 sizes its page with svh and a vh fallback instead of dvh. The lift is in the server HTML, so it works without JavaScript.
There was a problem hiding this comment.
Chip review — no blocking findings — this is not an approval
No findings at this head. The A1 CTA hand-off gap is fixed, and the analytics opt-out cookie fix remains intact.
Checked clean
- Exact detached head and dev merge base verified; reviewed the 69-file creator-page PR diff, not unrelated dev merges.
- P1 fixed: A1 observer excludes the bottom 100px, with a lifted pinned bar and closing CTA for affected iOS in-app browsers.
- P2 remains fixed: disallowed analytics expires the Path=/go cookie, while DNT and GPC bypass it during server selection.
- Reviewed the copy removal and selector adjustments in A1, E1 and G2; legal not-a-bank text remains on all three.
- Checked request-to-HTML UA lift, variant props, UI geometry, design rules, trust boundaries and tests; no additional defect established.
- Exact-head unit, typecheck, lint, design-lint, screen, security-scan and preview checks passed; capture_after (en) was still in progress.
Security review by moonshotai/kimi-k3: 0 finding(s), marked with the model name. It reads the diff only and answers only security, privacy and money, so treat its findings as advice.
Third opinion: did not run — it reads only the first review of a pull request; the first reviewer checks later rounds. This review is one reviewer short.
Copy check by claude-haiku-5-5: 0 finding(s), titled "Copy now wrong". Every quoted passage was checked against the file.
Exact head: f4b0b3190e78 · Context: repo, design · Took 8m
Summary
Nita's invite link sent 1,089 Meta visitors to a generic English page and got zero claim taps. This PR gives a creator their own page on
/go/<username>and a small service to test page variants on the same link, so we share one link per creator and change the page ourselves.src/features/creator-pages/): a registry the/goengine calls through its page seam, weighted variant pick, a sticky per-creator cookie (written only when analytics are allowed, removed when they are not, and ignored on a request that carries DNT or GPC), and one click event,creator_page_cta_clicked. The event carries the engine's visit properties (placement, code, utm, page experiment and variant) pluscreator_username,device_os,in_app_browserandcta_position, so results split per traffic type.juanacervio, equal weights, experimentnita-2026-11-a: A1 (long page with her post), G2 (one screen over her Reel still), E1 (poster)./dev/creator-pages(404 on peanut.me).Stacked on #4104 (page seam), which stacks on #4102. It targets
devonce those land. The engine files (src/app/go/**,src/features/tracked-links/**) are not touched here.One PR, about 4,700 lines: most of it is the three pages and their tests. E1 is the piece to split out if a smaller review is wanted.
Task
Personalised invite page per creator on /go/ (TASK-24893)
Risks
e2e/flows/creator-pages.spec.tsfails if one grows past it.next dev; the tap test in the spec needs a production build.NitaA1.tsxsays so.src/utils/marketing-routes.ts(+ test) puts the dev preview on the marketing tree,src/constants/analytics.consts.tsadds the event name,scripts/native-build.jsexcludes the preview from native builds.public/creator-pages/juanacervio/need Konrad's approval. G2's Reel still is a 2x enlargement of a 425x758 frame until the original arrives.QA
/dev/creator-pageson a dev build lists the three pages; open each at 375x667 and 320x568.e2e/flows/creator-pages.spec.ts(run alone, commands in its header): first HTML and JavaScript-off checks pass for all three onnext dev; the tap test needs a build.tracked_link_visitandcreator_page_cta_clickedin PostHog withpage_experiment = 'nita-2026-11-a'.engineering/ui/creator-pages.md.Screenshots
Dev preview (
/dev/creator-pages/juanacervio/<variant>, fixture data), headless Chromium at 375x667 with an Instagram iPhone user agent. Not a real phone.The images live on the
pr-assets-4107branch, which is deleted after merge.