TLV Minyanim

Where can I daven in the next 40 minutes?

Nobody in Tel Aviv can currently answer that. There are synagogue directories — they list buildings, addresses, phone numbers. What they cannot do is sort today's prayer times into a timeline, because most of those times are not clock times at all. They are rules: twenty minutes before sunset, ten minutes before candle lighting, at dawn.

This is a bilingual (Hebrew/English) site for the person who is not where they usually pray — a visitor, a business traveller, someone saying kaddish in an unfamiliar city.

Read that Mincha row: 20 דק' לפני שקיעה = 18:47. The rule is stored; the clock time is computed for today. Tomorrow it will say something else, and it will still be right — because sunset moved, not because anyone edited a row.

---

The one idea

A minyan time is a structured value, never a display string.

type MinyanTime =
  | { kind: 'fixed';    time: string }                       // "13:30"
  | { kind: 'relative'; anchor: Zman; offsetMinutes: number } // shkia − 20
  | { kind: 'unknown';  rawText: string }                     // "בזמן"

LA Jewish Times — the closest comparable — stores relative times as the literal text "~25 Min before Netz". Those minyanim are therefore invisible to its own next-minyan feature: a string cannot be sorted into a timeline. That single decision breaks the only feature that matters.

The invariant is enforced in Postgres, not just TypeScript:

CONSTRAINT minyan_time_shape CHECK (
  CASE kind
    WHEN 'fixed'    THEN fixed_time IS NOT NULL AND anchor IS NULL AND …
    WHEN 'relative' THEN anchor IS NOT NULL AND offset_minutes IS NOT NULL AND …
    WHEN 'unknown'  THEN raw_text IS NOT NULL AND fixed_time IS NULL AND …
  END
)

An import script cannot write in SQL what the types forbid in the app. All three malformed shapes were tried against a live database and rejected.

The third kind matters most

About 60% of the source data says only בזמן — "at the proper time" — with no offset. That is kind: 'unknown', and it is never guessed.

A hallucinated Mincha sends someone across a neighbourhood to an empty room, and it does so with a confident-looking number on the page. A blank is honest. The parser has a test asserting that two shuls with identical field shapes — one stating shkia − 20, one saying בזמן — do not let the second inherit the first.

The site says so out loud rather than hiding it:

טעון בדיקה — לא מאושר המקור לא ברור, והשעות האלה לא אושרו. אל תסתמכו עליהן.

Where it stands

Parser	Hebrew free text → structured values. All 26 raw fields parse, zero leftovers
Data	16 synagogues, 65 minyanim — 42 fixed, 19 unknown, 4 relative. 1 verified against its own notice board
Zmanim	@hebcal/core, GRA, resolved at read time and never stored
Tests	284, including edge cases asserted against published luachot
Contrast	AA in both modes — 130 elements, 0 failures, worst 5.05 light / 5.73 dark

Not built: geo/radius search, the nightly refresh job, and any deployment — the site currently runs locally only.

The real gap is not code. Shacharit is known for every shul. Mincha is 69% unknown, and exactly one synagogue in the neighbourhood publishes a real offset. No feature closes that; twelve conversations with gabbaim would.

---

A few decisions worth the detour

Candle lighting in Tel Aviv is shkia − 22, not 20. @hebcal/core ships 20 for the Tel Aviv geoname and Hebcal's own pages print 20 — but the Tel Aviv-Yafo Religious Council, the authority for exactly these synagogues, publishes 22, and MyZmanim labels it 22 דקות קודם השקיעה. Checked against all 34 published Fridays of 5786 falling in 2026: no date fits 20, 21 or 23. Where published authorities disagree, take the earlier — two minutes early costs nothing, two minutes late is chillul Shabbat.

tzeit names two different times, so no tzeit minyan is published. On a luach, יציאת שבת is the stringent 8.5° value. A synagogue writing צאת הכוכבים on its Arvit line means the nightfall it actually davens at — up to 26 minutes earlier. Resolving one against the other listed a real minyan 26 minutes late, so the anchor is kept and the record held back until a gabbai settles it.

מנחה 1:30 means 13:30, and that is reading rather than guessing. Mincha at 01:30 does not exist, so the time is stated and only the clock convention is open. Every such shift is recorded in clockNormalisation and stays queryable. Where neither reading is possible, the record is flagged and withheld.

Hebrew sorts with COLLATE "C". The cluster default is en_US.UTF-8, which orders Hebrew into nonsense — it put המרכזי before אוהל. Invisible to anyone not reading Hebrew.

Contrast is measured by painting a pixel, never by reading a colour string. color-mix() computes in oklab, and a regex reads oklab(0.72 0.13 0.11) as near-black — 81 elements on the homepage are affected. See scripts/contrast-audit.js.

Tests never assert that a library returns what the library returns. Expected zmanim come from published luachot, sourced independently of this code (docs/zmanim-ground-truth.md). Reverting candle lighting to 20 fails six tests; switching tzeit to the 7.083° opinion fails ten.

---

Stack

Next.js (App Router) + TypeScript, server-rendered — every synagogue page must be crawlable. Postgres + PostGIS for real geo queries. @hebcal/core for zmanim, never hand-rolled. Zero runtime dependencies in the parser, which is a standalone module with its own tests.

Bilingual routing (/he, /en) from the first commit, because retrofitting it means losing every ranking. CSS logical properties only; Asia/Jerusalem throughout, where the Hebrew date rolls at sunset.

Running it

brew install postgresql@17 postgis && brew services start postgresql@17
createdb tlv_minyanim
psql -v ON_ERROR_STOP=1 -d tlv_minyanim -f db/migrations/0001_init.sql
psql -v ON_ERROR_STOP=1 -d tlv_minyanim -f db/migrations/0002_shiurim_and_parse_issues.sql
psql -v ON_ERROR_STOP=1 -d tlv_minyanim -f db/migrations/0003_minyan_nusach.sql
psql -v ON_ERROR_STOP=1 -d tlv_minyanim -f db/migrations/0004_erev_shabbat_day_type.sql
psql -v ON_ERROR_STOP=1 -d tlv_minyanim -f db/migrations/0005_synagogue_nusachim.sql
psql -v ON_ERROR_STOP=1 -d tlv_minyanim -f db/migrations/0006_minyan_validity.sql

Put DATABASE_URL=postgres:///tlv_minyanim in .env.local, then:

npm install && npm run import:seed && npm run dev

npm test	the suite (284)
npm run typecheck	parser and app
npm run import:seed	load seed data through the parser — idempotent
npm run coverage	how much of the source actually parsed

Layout

src/minyan-times/   the parser — standalone, no framework, no runtime deps
src/zmanim/         rules to instants. Computes nothing itself; @hebcal/core does
src/db/             plain SQL, no ORM — the schema is the source of truth
src/app/[locale]/   server-rendered pages, Hebrew and English
db/migrations/      schema. The time invariant is a CHECK constraint here
docs/               zmanim ground truth, sourced independently of the code

CLAUDE.md is the project's binding contract — the invariants, the taxonomy, and the decisions that must not be undone by accident.

Two things that will trip you up

data/seed-ramat-aviv.json is gitignored and must stay that way. It carries gabbai and rabbi names and personal phone numbers — personal data under Israeli privacy law. There are no columns for them in the schema, the importer deletes those fields at the read boundary, and nothing logs or fixtures them.

Do not serve the repo root over HTTP. python3 -m http.server does not read .gitignore. The artboard preview in .claude/launch.json is bound to 127.0.0.1 for exactly this reason.

Licence

MIT for the code. The photographs are not licensed for reuse, and the synagogue data derives from the Tel Aviv-Yafo municipality GIS layer and Religious Council under their own terms. See LICENSE.