JiPange

A financial literacy toolkit for Kenya: 26 calculators that answer money questions in Kenyan terms — PAYE and SHA on a payslip, Fuliza's real cost, a chama's payout order, whether a SACCO beats a bank, what a T-bill leaves you after withholding tax and inflation.

Everything runs in the browser. There is no account, no server-side profile, and nothing about a user's money leaves their device.

Getting started

npm install          # workspaces: the app lives in web/
npm run dev          # http://localhost:3000

From the repository root, npm run dev|build|lint|test all delegate into web/. To run a command directly against the app, use npm run <script> --workspace web.

Command	What it does
npm run dev	Next.js dev server
npm run build	Production build (next build --webpack)
npm test	Unit suite (Vitest)
npm run lint	ESLint
npm run sync:rates	Refresh the committed rates snapshot — see below

Inside web/ there are two more: npm run typecheck (tsc --noEmit) and npm run e2e (Playwright).

Layout

web/app/          Next.js App Router. Each calculator is app/tools/<name>/
web/lib/          The maths. Pure modules, no React — this is where correctness lives
web/lib/__tests__ Unit tests, deliberately beside the code they test
web/components/   UI. Presentation only; a component should not compute a number
web/e2e/          Playwright specs
scripts/          Build-time and CI tooling (rates sync, Netlify build gate)

The split that matters is lib/ computes, components/ renders. A calculator's arithmetic belongs in lib/ where it can be unit-tested without a browser; the component reads the result and displays it. When those blur, a number gets computed twice in two places and the two drift.

Where the rates come from

Interest rates are not fetched at runtime. scripts/sync-rates.mjs pulls Mwangaza Yield's published feed in CI and commits a snapshot, and the app ships that file (web/lib/rates-feed.ts explains the reasoning).

The script's governing rule is that it refuses rather than risks: an unreachable feed, an unparseable response, an unknown schema version, missing tenors, a yield outside a sane band, or a figure that moved implausibly far in one day all end the same way — the existing snapshot is left untouched and CI reports it. Yesterday's verified rate beats today's corrupt one in a tool people use to decide where to put money.

Exit codes: 0 updated or already current, 1 refused.

When the sync stops

npm run doctor answers one question — is anything still refreshing the rates? — and answers it from committed files, so it works offline, on a laptop, and outside CI. It reports the snapshot's age against the budget the test suite enforces, whether readers are currently seeing the stale notice, which rung actually pays most, and whether the live feed has moved ahead of us.

It exists because on 20 August 2026 every workflow in this repository began failing in about four seconds, before a single step ran and with no runner assigned — the twice-daily sync and CI alike. That is an account-level Actions problem, not a repository one, and no change here fixes it. The consequences were exactly what you would expect once all three alarms turn out to live inside the thing that died: the snapshot froze at 19 August, four commits landed on main with no CI, and a ladder preset went on telling readers the 364-day rung "pays most" while the feed in the same build said the 182-day rung paid more.

The reader-facing layer held — isStale() put the notice up at 14 days as designed. It was the operational alarms that were silent, and npm run doctor is the one that now lives outside CI.

If the workflow cannot run and you can reach the feed, run npm run sync:rates locally and commit what it writes. Do not hand-edit figures; if you cannot reach the feed, leave the snapshot alone.

And where the forward-looking figures come from

The registry in web/lib/sources.ts also carries three PROJECTIONS from the IMF's World Economic Outlook (April 2026 edition) — Kenya's 2026 growth, annual-average inflation and debt-to-GDP — each with its WEO subject code, an asOf date and a reviewBy tied to the WEO's April/October cadence, so the staleness gate in __tests__/sources.test.ts forces a re-read after every edition. They are exported from kenya-stats.ts as IMF_WEO_*.

They answer a different question from the measured figures, and the two are never interchangeable: the KNBS/CBK readings say what inflation IS, the WEO entries say what IMF staff EXPECT. A page that looks forward (a plan, a projection) may quote them; a page that states current conditions uses the measured outturns. The note on each registry entry says why.

Deployment

Netlify, from main. Configuration is in netlify.toml.

Two cost controls are deliberate and documented at the point they take effect:

• scripts/netlify-should-build.sh cancels a build when a commit touches nothing next build reads. Its exit polarity is inverted from every other script here — 0 skips, non-zero builds — so it says so at both exit points. Every uncertain case builds: a build we did not need costs one credit, a skipped build we did need ships a planner that disagrees with the one people were shown.
• Deploy previews and branch deploys are off. They were roughly half of all build spend, against a credit pool shared with Mwangaza that reached zero and left merged commits unpublished on both sites. Review happens through CI, the diff, and local screenshots.

Both are reversible: the comments in netlify.toml say exactly what to delete to bring previews back.

Stack

Next.js 16 · React 19 · Tailwind v4 · TypeScript · Vitest · Playwright · Netlify.

One trap is worth knowing before you touch the export feature. Tailwind v4 emits colours via color-mix() in oklab space, and html2canvas 1.4.1 throws unsupported color function oklab on it — which reaches the user as an export button that appears to do nothing. The app uses html2canvas-pro instead. Import that, never html2canvas.

Contributing

This is not an open-source project and pull requests from outside the team are not being accepted — see the licence below. For anyone working on it:

• Put arithmetic in lib/, with a test beside it. A calculator without a unit test is a calculator nobody can safely change.
• Prefer a test that fails when the behaviour is reverted. A test that passes against an empty result is worse than no test, because it reports coverage that does not exist.
• Money is Ksh, consistently, everywhere a user can see it.
• Run npm test and npm run lint before pushing; npm run build if you have touched anything under web/app.

Licence

Copyright © 2026 JiPange. All rights reserved.

This source is published for reference and is not licensed for reuse, redistribution, or derivative works. No open-source licence is granted, and nothing here should be read as one. For licensing enquiries, see /licensing.