daf-ai — דפי חזרה ולימוד לדף היומי

Live site: https://daf.tsemach.dev (the old daf-ai.pages.dev redirects there)

Interactive Hebrew study pages for Daf Yomi, generated by an AI agent and verified against the Gemara text. Content lives as JSON; a zero-dependency Python generator renders a static site.

Layout

Path	What
data/<slug>/<daf>.json	One daf: header, nav, storyline (הקדמה), sugyot with steps + argument tree, quiz, links, source texts
data/glossary.json	Shared glossary (sages, concepts). Longest key wins; first occurrence per page is marked
data/sources.json	Shared verse / halacha texts shown in the in-page source dialog
content/home_about.html	Static blocks on the home page
site/app.css, site/app.js	Shared runtime: glossary popovers, source dialogs, tree view, chavruta mode, settings, quiz
build.py	data/ → dist/ (/, /<slug>/, /<slug>/<daf>/)
validate.py	Schema + content checks; non-zero exit on errors
tools/extract_from_hub.py	One-time migration from the original single-file hub

Run locally

python validate.py        # must pass
python build.py           # writes dist/
python -m http.server -d dist 8000

Daf JSON (summary)

{
  "key": "bechorot9", "tractate": "Bekhorot", "slug": "bechorot", "tractate_he": "בכורות", "daf": 9,
  "title": "בכורות ט׳", "eyebrow": "…", "thesis": "…", "meta": ["<hebrew date>", "עמודים א–ב", "~20 דקות לימוד"],
  "nav": [{"id": "s2", "amud": "עמוד א", "title": "…"}],
  "storyline": "<one paragraph, html allowed>",
  "sections": [
    {"id": "s2", "kind": "sugya", "ref": "Bekhorot.9a.7-13", "amud": "עמוד א · גמרא", "title": "…",
     "quote": {"text": "…", "cite": "…"} | null,
     "flow": "<בקצרה: 1–3 sentences>",
     "pre": "<html before the steps>", "post": "<html after: verdicts, think boxes>",
     "steps": [{"tag": "קושיה", "type": "q|a|c|", "hideable": true, "p": 0, "body": "<html>"}]},
    {"id": "sum", "kind": "raw", "amud": "סיכום", "title": "…", "html": "<table…>"}
  ],
  "quiz": [{"q": "…", "o": ["…","…","…"], "a": 1, "e": "…"}],   // exactly 8
  "links": [{"href": "…", "text": "…"}],
  "srctext": {"Bekhorot.9a.7-13": {"t": "בכורות ט׳ ע״א", "p": ["<verbatim segment>", "…"]}},
  "card": {"date": "…", "topics": "…", "stats_html": ["<span>…</span>"]}
}

p = index of the earlier step this step responds to (argument tree). Step types: q question, a answer, c conclusion, empty = statement.

Deploy — Cloudflare Pages

1. Cloudflare dashboard → Workers & Pages → Create → Pages → Connect to Git → tsemachh/daf-ai.
2. Build command: python build.py · Build output directory: dist · Root: /.
3. Environment variable PYTHON_VERSION=3.11.
4. Custom domains → add your domain (DNS on Cloudflare → one click).

Every push to main rebuilds and deploys.

Reader feedback

Every daf page has "💬 הערה על הדף" and a per-sugya "הערה על סוגיה זו". Notes are stored in Cloudflare D1 (db/schema.sql) by Pages Functions in functions/:

Endpoint	Who	What
POST /api/feedback	public	save a note → {id, token}; honeypot, optional Turnstile, 6/hour per IP hash
GET /api/feedback/status?n=<id>.<token>,…	the author	status + public reply of their own notes
GET /api/feedback/page?p=bechorot/9	public	notes marked fixed on a page (no text/name/email)
GET /api/feedback/config	public	Turnstile sitekey
/admin/ , GET /admin/api/list?status=new , POST /admin/api/update	owner / nightly job	dashboard and API, behind Cloudflare Access

Readers stay anonymous. The browser keeps id.token pairs in localStorage; the tracking page is /feedback/ (or /feedback/?n=<id>.<token> from the link shown after sending). An email is kept only when "עדכון במייל" is ticked, and should be cleared (clear_email) after the update is sent. Pages show "✓ תוקן בעקבות הערת קורא" on sections with fixed notes.

Statuses: new, in-progress, fixed, feature, rejected, needs-info. Update body: {"id":12,"status":"fixed","reply":"…","commit_sha":"abc1234","notified":true,"clear_email":true}.

Setup (once, Cloudflare dashboard)

1. D1 → Create database daf-ai-feedback → Console → paste db/schema.sql → Execute.
2. Pages → daf-ai → Settings → Bindings → D1 database, variable name DB → daf-ai-feedback.
3. Turnstile → Add widget (hostnames daf.tsemach.dev and daf-ai.pages.dev, Managed). In Pages → Variables and Secrets add TURNSTILE_SITEKEY (text) and TURNSTILE_SECRET (secret). Optional: IP_SALT (secret).
4. Zero Trust → Access → Applications → Self-hosted, domains daf.tsemach.dev and daf-ai.pages.dev, path admin, policy "Allow" for the owner's email. Copy the app's AUD tag; add Pages variables ACCESS_TEAM = <team>.cloudflareaccess.com and ACCESS_AUD = the AUD tag. For the nightly job: Access → Service Auth → create a service token and add a second policy (action "Service Auth") for it. The job sends CF-Access-Client-Id / CF-Access-Client-Secret.
5. Redeploy (Deployments → Retry) so bindings and variables take effect.

Without DB the form answers 503 ("not saved"); without ACCESS_* /admin answers 503.

Local test

npx wrangler d1 execute DB --local --file db/schema.sql   # with a wrangler.toml that declares DB
npx wrangler pages dev                                     # serves dist/ + functions/

Credits

Every page shows the sources used and "Powered by Shefing" with the logo (site/static/shefing-logo.png, shefing-logo-white.png for dark mode) linking to shefing.com.

Portability

The nightly instructions live in tasks/daily-daf.md. Content, code and hosting are in GitHub and Cloudflare, independent of any Claude account: to move the job, create a scheduled task on the new account with that file as the prompt and give it write access to this repo.

Optional section fields

• table: opinions × cases matrix — {"title": str, "cols": ["", "<שיטה>", …], "rows": [["<מקרה>", "<תא>", …]]}.
• decide: if → then cards — {"title", "branches": [{"if": str, "then": [str, …]}]}.
• chain: source chain — {"title", "links": [{"k": "פסוק|דרשה|דין|תוצאה", "t": str}], "note"?: str}.
• calc: worked numbers — {"title", "rows": [[label, value, cite?]], "lines": [[label, value, cite?]], "note"?}.
• seq: order of events — {"title", "items": [str, …]}. table / decide / chain / calc / seq render together in a collapsed "עזרים" panel.
• cont: "next" / "prev" for a sugya split across dapim (otherwise inferred from the amud label).

What's new

data/changelog.json — newest first (new days at the top, and new items at the TOP of the day's list), [{"date": "YYYY-MM-DD", "items": ["…"]}]. Home shows the latest 5 items; /changes/ lists everything. Add an entry for every site improvement.

Model comparison lab

lab/index.json lists dapim that have an alternative version (lab/<slug>-<daf>-<model>.json, same schema as data/). build.py renders /lab/<slug>/<daf>/ with both versions side by side (blind, randomized order) and a vote (/api/lab/vote, table lab_votes); the regular daf page links to it.