Skip to content

STE-23: the landing becomes its own app, for sterish.xyz - #66

Merged
Lin1er merged 4 commits into
mainfrom
sharkzneverending/ste-23-landing-own-app
Sep 28, 2026
Merged

Lin1er merged 4 commits into
mainfrom
sharkzneverending/ste-23-landing-own-app

Conversation

@Nabil-Aufa

Copy link
Copy Markdown
Collaborator

Moves the landing out of frontend/ into its own deployable app, so it can serve
sterish.xyz while the dashboard keeps app.sterish.xyz.

cc @axel @fable @Ancung — @Ancung, nothing in frontend/ changes except that
app/landing/ and src/modules/landing/ are gone. globals.css is untouched.

Why separate

The two URLs answer different first questions. Someone arriving at sterish.xyz has usually
never heard of Sterish and needs the argument; someone opening app.sterish.xyz already
knows and wants the registry.

Serving both from one app meant the landing inherited the product nav and a "Connect
wallet" button above the fold
, which asks a stranger for something before it has told them
anything. That is now gone.

What the landing does not carry

It is a one-page app, so it does not take the dashboard's dependencies:

Dashboard Landing
API client 391 lines: x402 payments, settlement receipts, wallet 2 GET requests
Types 327 lines, whole API the shapes this page reads
Skeleton shadcn + cn 3 lines, local
Deps 14 runtime 3

Both the client and the types are written against docs/api-spec.md, which is the contract
the two apps actually share.

The API timeout drops from 30s to 10s here. The dashboard's has to resolve because somebody
is waiting on a verdict before installing; the landing falls back to a dated snapshot and
says so, and a page that hangs for half a minute has already lost the reader.

Two things the split exposed

The lockup and the footer came from the dashboard's layout. Without them sterish.xyz
carried no logo at all and ended on a bare paragraph. Both are now the landing's own.

The tokens are a copy, and the copy is guarded

landing/app/globals.css is frontend/app/globals.css minus two imports whose packages this
app does not install. Everything else is byte-for-byte identical.

A copy is the arrangement that drifts quietly: retune --danger on one side, both apps keep
building, and the two faces of the product diverge over a month with nothing to catch it.
That is the exact failure /brand was built to avoid, so it is not left to memory:

tokens.drift.test.ts reads both files and fails the build when they disagree.

Verified by breaking it on purpose — changing --danger in the dashboard's file turns the run
red, restoring it turns it green. A guard nobody has seen fail is not a guard.

The Landing workflow therefore also triggers on frontend/app/globals.css, so a token edit
made only on the dashboard side still runs the guard.

CI

New .github/workflows/landing.yml. Without it nothing under landing/ would be type
checked, linted, tested or built on any PR, since the Frontend workflow is filtered to
frontend/**.

Separately, for @axel: a docs-only PR cannot merge today. install + lint + build is a
required check but its workflow is path-filtered, so it never reports and the PR sits BLOCKED
forever. That is what happened to #63, which is still open and needs an admin merge. It
is a branch-protection setting, not something a file in this repo can fix.

Deploying

Set each Vercel project's Root Directory: landing/ for sterish.xyz, frontend/ for
app.sterish.xyz. Neither needs environment variables to build; NEXT_PUBLIC_API_URL
defaults to https://api.sterish.xyz. Recorded in docs/deployments.md.

Tests

Landing: 13 passing (9 on the figures the page states as fact, 4 on token drift).
Dashboard: 165 passing, unchanged by this PR. Typecheck, lint and production build clean
on both.

Measured on the standalone build: TTFB 80ms, and the evidence figures now read 30 of 38
on chain, 21 VERIFIED
— up from 24 of 30 on 20 September, with nothing edited. That is the
point of reading the numbers instead of typing them.

Still open

  • 375px not visually checked. The browser harness will not apply a window resize. Mobile
    rests on responsive classes; worth an eyeball in devtools.
  • sterish.xyz still 404s until the Vercel project is pointed at landing/.

Linear

STE-23 — https://linear.app/sterish/issue/STE-23

🤖 Generated with Claude Code

Nabil-Aufa and others added 4 commits September 28, 2026 19:54
sterish.xyz and app.sterish.xyz answer different first questions. Someone
arriving at the landing has usually never heard of Sterish and needs the
argument; someone opening the dashboard already knows and wants the registry.
Serving both from one Next app meant the landing inherited the product nav and
a "Connect wallet" button above the fold, which asks a stranger for something
before it has told them anything.

So landing/ is its own pnpm project with its own layout, deployed from its own
Vercel root directory. The page moves to `/` because it is now the root of its
own site rather than a route inside another one.

It does not carry the dashboard's dependencies. The client here is two GET
requests instead of 391 lines that also build x402 payments, decode settlement
receipts and talk to a wallet -- none of which belongs on a deployment that can
never take a payment. Same for types: the shapes this page reads, not the whole
API. Both are written against docs/api-spec.md, which is the contract the two
apps actually share. The shadcn Skeleton becomes a local three-line one rather
than pulling shadcn in for a single placeholder.

Two things that were invisible until the split: the lockup and the footer came
from the dashboard's layout, so without them sterish.xyz carried no logo at all
and ended on a bare paragraph. Both are now the landing's own.

The API timeout drops from 30s to 10s. The dashboard's has to resolve because
somebody is waiting on a verdict before installing; here the caller falls back
to a dated snapshot and says so, and a landing page that hangs for half a
minute has already lost the reader.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two deployments now means two copies of globals.css, and a copy is the
arrangement that goes wrong quietly: somebody retunes --danger on one side,
both apps keep building, and the two faces of the product drift apart over a
month with nothing to catch it. The whole reason /brand reads its tokens
instead of listing them was to stop exactly this.

So the copy is not defended by anyone remembering. This reads both files off
disk and fails when they differ anywhere outside the two imports the landing
drops (shadcn/tailwind.css and tw-animate-css, whose packages it does not
install).

Verified by breaking it on purpose: changing --danger in frontend/app/globals.css
turns the run red, and restoring it turns it green again. A guard nobody has
seen fail is not a guard.

Two cheaper assertions sit beside it because a byte comparison of two empty
files would also pass: the five brand values and the four verdict tokens have
to actually be present.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Frontend workflow is filtered to `frontend/**`, so nothing in landing/
would have been type checked, linted, tested or built on any PR.

It also triggers on `frontend/app/globals.css`, which looks wrong until you
follow it through: the drift guard lives in this job, so a token edit made only
on the dashboard side has to run it. Without that path the guard would sit
green while the two apps diverged, which is the precise failure it exists to
prevent.

Note for whoever hits it next: a docs-only PR cannot merge today, because
"install + lint + build" is required but its workflow is path-filtered and
never reports. That bit PR #63. Not fixed here -- it is a branch-protection
setting, not a file in the repo.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
sterish.xyz now has an owner in the table instead of a "reserved, currently
404" note, with the Vercel root directory each project needs.

The part worth reading is the last paragraph: the apps share tokens by copy,
not by import, and the copy is guarded by a test rather than by discipline.
Change frontend/app/globals.css first, then bring it across.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@vercel

vercel Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
sterish Ready Ready Preview Sep 28, 2026 12:55pm UTC

Lin1er added a commit that referenced this pull request Sep 28, 2026
A ruleset named `master` appeared on the default branch today at 19:20 (+07) and
froze the repo: every open PR sits BLOCKED with green CI. Two diagnoses were
wrong before the real one, so all three are written down.

Not a path-filtered required check that never reports, as #66 suggests:
`install + lint + build` is SUCCESS on both of Nabil's PRs and the ruleset has no
`required_status_checks` rule at all. Not `required_signatures` either — that
rule was active at 19:20 and removed at 20:05, and all three PRs stayed BLOCKED
afterwards.

It is the `update` rule, "Restrict updates". Merging a PR updates the ref, so it
blocks merges and not only direct pushes, and with `bypass_actors: []` and
`current_user_can_bypass: "never"` an admin merge cannot get around it either.
The symptoms actively mislead: `mergeable: MERGEABLE` next to
`mergeStateStatus: BLOCKED`, and an error that names no rule.

Also records that the fix has to come from the web UI: the gh OAuth token is
answered 404 for `PATCH /repos/.../rulesets/{id}` even though the account is a
repository admin, so the API cannot undo this.
@Lin1er
Lin1er merged commit bd6143a into main Sep 28, 2026
4 checks passed
Lin1er added a commit that referenced this pull request Sep 28, 2026
For about ninety minutes today no PR in this repo could be merged, with green CI
on all of them. Worth recording, because the symptoms point away from the cause
and the first two diagnoses were both wrong.

Not a path-filtered required check that never reports, as #66 suggested:
`install + lint + build` was SUCCESS on both of Nabil's PRs and the ruleset
carried no `required_status_checks` rule at all. Not `required_signatures`
either — that rule was removed at 20:05 and all three PRs stayed BLOCKED.

It was the `update` rule, "Restrict updates". Merging a PR updates the ref, so it
blocks merges and not only direct pushes, and with an empty bypass list an admin
merge cannot get around it. GitHub reports `mergeable: MERGEABLE` next to
`mergeStateStatus: BLOCKED` and names no rule in the error.

The ruleset has since been deleted, so this is written as the recipe for next
time: how to read the effective rules, which combination protects main without
freezing it, and why the fix has to come from the web UI — the gh OAuth token is
answered 404 for the rulesets API even though the account is a repository admin.

This branch was successfully deployed

1 active deployment
Preview — faedf1b3 Deployed Sep 28, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants