afiche

Cartelera curada de Buenos Aires

afiche pulls the weekly programming of Buenos Aires' independent and repertory cinemas into one clean, editorial-feel cartelera — one screen for the whole indie circuit, built for people who already know the difference between a Saturday at Lugones and a Saturday at Hoyts. Live at afiche.ar — Next.js on Vercel, libSQL on Turso.

What it does

• One row per film, scoped to a time window. The homepage groups every screening by film (not showtime) for a relative window — Hoy (default), Este finde, Esta semana, Próximamente — switchable via ?ventana=hoy|finde|semana|prox (shareable; an unknown value falls back to hoy). A single-showtime film renders an inline time · venue; a multi-showtime film collapses to a {n} funciones · {venue} summary that taps open. Prod data earns the layout: 64% of films have one showtime all week and 95% play a single venue, so the common row is one clean line.
• A curated "Destacados" band. Above the film grid sits a full-bleed row of 0–4 hand-picked poster cards. Curation is visible, not buried behind a toggle.
• An exhaustive day-by-day view at /cartelera. One tap from the homepage ("Ver todo →"): a 14-day rolling window of full cards grouped by day, a sticky date strip, and a Próximamente text index for the longer tail.
• Per-film pages with cross-venue discovery. /pelicula/<slug> gives each film a title block, an optional 16:9 editorial still, poster, synopsis, top-billed cast — and the killer feature: every upcoming screening of that film across BA, in one list. It answers "when else can I catch this?", which no single venue's own site can.
• Rich, deduplicated film cards. Each film is enriched from TMDB (poster, backdrop, director, runtime, country, original title, Spanish synopsis, genres, cast). The same film surfacing with title drift across cinemas collapses to one entry, and a title-ambiguity guard keeps the three Nosferatus apart.

What it doesn't do (yet)

• No on-site search or filters — the cartelera is the index. Title + director search exists in the data layer and is exposed over MCP, but it has no UI on the site yet; that's queued behind /pelicula/ usage data.
• No per-cinema programming pages, and no /programa/<slug> pages for curated cycles (Lugones Olivera-Aries, MALBA Cineclub Nocturna). Cycles are denormalized text on screenings for now.
• No chains. afiche covers the indie / repertory circuit only; Cinépolis and the multiplexes are intentionally out. This is the indie-circuit cartelera, not all of BA cinema.
• No dark mode — the editorial cream + carmine palette is the brand.

Cinemas covered

Cinema	Neighborhood	Type	Source
Sala Leopoldo Lugones	San Nicolás	indie / repertory	complejoteatral.gob.ar
MALBA	Palermo	indie / museum	malba.org.ar
Cine Lorca	San Nicolás	indie / repertory	cinelorca.wixsite.com (VLM-extracted poster)
Cine Cosmos	Balvanera	indie / repertory	cinecosmos.uba.ar
Cine York	Olivos	indie / repertory	lumiton.ar
Centro Cultural Munro	Munro	indie / community	lumiton.ar
Lumiton	Munro	indie / museum	lumiton.ar
Cine Gaumont	San Nicolás	indie / repertory (INCAA)	cinegaumont.ar (Cinexo ticketing JSON API)
CineArte Cacodelphia	San Nicolás	indie / repertory	cineartecacodelphia.com.ar (adro.studio JSON API)
Centro Cultural Borges	San Nicolás	indie / cultural center	centroculturalborges.gob.ar
Cineclub Lucero	Palermo	indie / cineclub	eventbrite.com (organizer feed)
Cinta	Palermo	indie / open-air terrace	passline.com (producer page)
El Cairo	Rosario	indie / public cinema (Santa Fe province)	elcairocinepublico.gob.ar (The Events Calendar REST API)

Architecture

graph LR
  V["12 venue sites<br/>lugones · malba · gaumont<br/>lorca · cosmos · york · …"]

  subgraph pipeline["Scrape pipeline · scrape-prod.sh (residential IP)"]
    R["run.ts<br/>orchestrator"]
    P["Provider.fetch() x12<br/>→ ScrapedScreening[]"]
    I["ingest()<br/>upsert films · replace future<br/>screenings · enrich"]
  end

  TMDB["TMDB API<br/>match (similarity) + enrich"]
  D[("Turso / SQLite · Drizzle<br/>cinemas · films · screenings<br/>providers · scrapeRuns<br/>tmdbOverrides · healRuns")]

  subgraph web["Next.js 16 · RSC on Vercel"]
    Q["queries.ts"]
    PAGES["/ · /cartelera<br/>/pelicula/[slug] · /sala/[id]"]
    ADMIN["/admin<br/>signed-cookie auth<br/>manual TMDB match"]
    REV["/api/revalidate"]
  end

  U(("visitor"))

  V -->|HTML / JSON| P
  R -->|drives| P
  P -->|screenings| I
  I <-->|lookup + enrich| TMDB
  I -->|upsert| D
  R -.->|POST when done| REV
  REV -.->|revalidate ISR| PAGES
  D --> Q --> PAGES
  ADMIN <-->|read / fix| D
  U --> PAGES

Getting data in. The scrape doesn't run in the cloud. scrape-prod.sh runs on an always-on machine behind a residential connection, because lumiton.ar and complejoteatral.gob.ar 403 datacenter IPs and a GitHub Actions runner can't reach them. It's scheduled twice daily and pulls main before each run, so a shipped scraper fix reaches the pipeline without anyone deploying it by hand. It walks all 13 providers (src/scrapers/run.ts) and ingests each through src/scrapers/ingest.ts — upsert films → replace that cinema's future screenings → TMDB-enrich. Every provider is pure () → ScrapedScreening[]. Three of the thirteen (Cine York, Munro, Lumiton) share lumiton-agenda.ts because one Lumiton agenda page drives all three venues; Cine Lorca has no HTML schedule at all — its week lives on a single poster image, so its provider hashes the image and, on a cache miss, runs a Claude vision call (temperature: 0, structured JSON) to read the showtimes. Cinta has no site at all — its listing lives on a Passline producer page behind a fingerprint-based Cloudflare challenge, so its provider tries the origin first and falls back to a reader proxy that returns the same post-challenge markup, recording the fallback as a run warning.

Staying correct. Ingest is idempotent (unique index on (scraped_title, scraped_year)) and self-healing. A film that shows up with title drift across cinemas — different localizations, year-null collisions — collapses to one row keyed on tmdb_id after enrichment, and a title-ambiguity guard plus director verification stop silent wrong-matches on common titles (the Nosferatu class: Eggers 2024 vs Herzog 1979 vs Murnau 1922 all share a Spanish title). A failed match parks the row and stamps it with the matcher version that failed it (MATCHER_VERSION, src/tmdb/match.ts), so improving the matcher re-opens exactly the rows it was written to rescue — without that, a better matcher never gets a second look at its own backlog, because nothing about those rows changed. A pre-write circuit breaker guards the other direction: ingest replaces a venue's future schedule wholesale, so a provider that silently breaks against a site redesign would publish an empty venue — when the fetch comes back empty and future rows exist, the replace is refused, the last good schedule stands, and the run reports circuitBroke. Each run records one scrape_runs row per (cinema, run) with status + counts; when it finishes, scrape-prod.sh POSTs /api/revalidate (even when some providers failed, so the ones that landed still reach the site) so the deployed pages pick up the fresh data. A daily Vercel cron reads the newest successful run and warns if the cartelera has gone stale — the one failure a scraper can't report about itself is the run that never happened.

Repairing itself. Deterministic matching still misses, and until now the gap between a miss and a fix was however long it took someone to notice. A three-actor loop runs after every scrape, pass or fail (scripts/self-heal.ts, scripts/actor2-fix.ts). Actor 1 audits the run, judges the films still stuck against the candidates the matcher already fetched, and writes only what it can corroborate — a candidate-judged proposal at confidence ≥ 0.90 with the year or director agreeing, or at ≥ 0.95 when the title is an exact, dominant match nothing contradicts; a web-researched id never auto-applies. Applied matches land in tmdb_overrides (durable across reset-programming, and always outranked by the hand-curated tmdb-overrides.json seed) and re-open the film for the next enrich; everything else queues for a human. Misses that no single override fixes are grouped by cause and filed as matcher-pattern issues, and Actor 2 picks up the mechanical ones, prepares the fix in a throwaway worktree, and opens a PR — it never merges. Each write is recorded in heal_runs for a 7-day trend in the digest, and SELF_HEAL_ENABLED, SELF_HEAL_APPLY_ENABLED, LAYER2_ISSUES_ENABLED and ACTOR2_ENABLED pause any stage without a code change. Full walkthrough in docs/agents/self-heal-loop.md.

Getting data out. Every serving route — and the MCP server — reads the same DB through src/db/queries.ts. The homepage (src/app/page.tsx) calls getWindowScreeningsByFilm (one row per film, bounded by ?ventana=), getFeaturedFilms (the Destacados band), and getJsonLdScreenings (structured-data feed). /cartelera (src/app/cartelera/page.tsx) calls getTwoWeeksScreenings (the 14-day window), getUpcomingScreenings (Próximamente), and getLastScreeningPerFilm (the ÚLTIMA FUNCIÓN anchor). /pelicula/[slug] calls getUpcomingScreeningsByFilm for the cross-venue list. The window registry (src/lib/windows.ts) is the single source for the nav, the ?ventana= validation, and the bounded query.

The cartelera as data. /api/mcp exposes the same live cartelera over the Model Context Protocol (Streamable HTTP, stateless), so any MCP client can answer "¿qué dan esta noche en Palermo?" against real showtimes. Four read-only tools — search_films, whats_on, get_film, list_cinemas — wrap the query layer in src/mcp/, return validated structuredContent, and emit Buenos Aires local times. Every showtime it advertises is filtered to still-catchable, so it never sends anyone to a screening that already started. No credentials; it's the same public data the site renders. The reading pages meet agents halfway too: /, /cartelera and /acerca answer an Accept: text/markdown request with markdown at the same public URL, and /llms.txt describes the site and how to call the MCP tools.

claude mcp add --transport http afiche https://afiche.ar/api/mcp

Tech stack

• Next.js 16 — App Router, React Server Components, force-dynamic on the cartelera + film pages
• React 19 + Tailwind CSS 4 — design tokens via @theme (see DESIGN.md)
• Type — Geist (body) + Geist Mono (eyebrows, caps) + Instrument Serif (masthead, day banners, film titles, italic time accents)
• Drizzle ORM + libSQL — SQLite locally, Turso in production
• cheerio for HTML scraping; Claude vision (claude-sonnet-4-6) for Cine Lorca's image-only poster, image-hash cached
• TMDB API for film metadata, posters, and editorial stills
• Vitest — ~1,130 tests across 61 files (scraper providers, ingest, TMDB enrichment + the self-heal safety gate, date/window helpers, layout invariants); fixtures are real HTML captures, so tests never hit the network or burn vision credits
• MCP — mcp-handler + @modelcontextprotocol/sdk serving the read-only cartelera tools at /api/mcp
• Vercel for hosting, ISR revalidation, and the daily freshness cron

Credits

afiche uses the TMDB API but is not endorsed or certified by TMDB. TMDB provides the film metadata, poster images, and editorial stills rendered alongside each screening. afiche hotlinks TMDB's CDN (image.tmdb.org/t/p/...) rather than re-hosting, per TMDB's best-practice guidance, and never resells or commercially redistributes TMDB data. Cine Lorca's image-only weekly programming is read via Anthropic's Claude vision API; the image is hashed and the result cached locally, so a given poster is parsed at most once.

License

Licensed under the PolyForm Noncommercial License 1.0.0 (source-available). Noncommercial use is permitted; commercial use is reserved to the author.

---

Built by Benjamin Delasoie. Spot a screening afiche missed or got wrong? Issues welcome.