Changelog

What changed on this site. The text comes straight from the code repository.

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

• Instagram lookup: after the passes of each batch, discovery runs a separate call (prompts/instagram_lookup.md) that lists the cafés, bars, pubs, clubs and venues of its municipalities and searches " instagram" for each. Every account discovery finds, and every mined account never read yet, is now checked against its profile before it is kept; a username that is only the bare place name (e.g. schongau) no longer counts as proof.
• Instagram candidates: discovery's venues pass searches " instagram" for cafés, bars and venues without a known account and reports the names directory sites give as instagram_candidates. Each candidate is checked against its profile (Graph API: name, biography, website) and kept only if it names the municipality or links the organiser's website; rejected names are stored in data/instagram.json (rejected) and retried after 180 days.
• Discovery collects Instagram accounts: every call reports the accounts of local organisers it came across, and the venues pass looks for those of cafés, bars, clubs and venues on purpose. New accounts go into data/instagram.json and are read by the next events run. With more mined accounts than instagram.max_accounts_per_run, the least recently read go first, so all of them take turns.
• The Actions list names each run by what it does: the job and "(dry run)" for Run, "daily" for scheduled runs, the places for Crawl check, the minutes for Geo check.
• Instagram: the events crawl reads recent posts of Instagram business and creator accounts (Graph API Business Discovery; secrets INSTAGRAM_ACCESS_TOKEN and INSTAGRAM_USER_ID). Accounts come from sources.instagram_file and are also mined from links on pages the crawl reads. New posts (caption and images) are read by the AI once and cached in data/instagram.json; settings in the new instagram config section.
• window.store_days (default 180): events the AI sees anyway up to this far ahead are kept instead of dropped; searching still stops at crawl_days, the website still shows display_days, the feeds contain all events.
• Flyer cache data/flyers.json: flyer files are fingerprinted (SHA-256) and read by the AI only once; unchanged files reuse their stored events.
• Merge: events with the same day, start time and place always go to the AI duplicate check (also when the titles share few words), including pairs already stored, which cleans up existing duplicates.
• A manual events run also rebuilds the site; the deploy job is skipped instead of failing when no site was built.
• Flyer pages: instance/sources/flyers.yaml (sources.flyers_file) lists pages whose programme is only in posters, brochure images or PDFs. The crawl downloads their images/PDFs (robots.txt respected, logos and icons skipped) and one AI call per page reads them without web access. Starts with Reaktor, Brauhaus, the parish newsletter and the Schlossbergler club calendar.
• New website look: green header band with brand icon, date cards and search, one colour per category (chips, card stripe, tags; mapped from the configured tags), events as cards with place pin, section icons, nicer pager, footer and content pages, refined dark mode. Days without events are dimmed in the date bar.
• depth config section for more thorough runs. Discovery makes three focused passes per batch (official calendars, clubs, venues), each with a German query checklist per municipality and a coverage report per category, which ends up in the Coverage section of data/suggested.md. The event crawl splits each batch into source sweeps that read every known source completely (no cap of 12) and one open search with a query checklist per month, topic and district.
• The Crawl check workflow also runs discovery (into a throwaway data/).
• site.indexing switch: false puts noindex on every page and stops inviting crawlers in robots.txt, to keep a site out of search results until it is ready.
• Geo retries the whole Overpass instance list up to three times (waiting 60 s, then 120 s) when every public instance is at capacity.
• Phase 7 discovery: Opus finds and rates sources per batch of municipalities (prompts/discovery.md), data/sources.json with origin, rating, update frequency and hits, whitelist synced in, blacklist enforced, data/suggested.md with whitelist and blacklist candidates (ADR-004); the events stage uses known good sources per batch and counts hits; forced after a geo change.
• Phase 6 scheduler and workflow: --job auto runs only due jobs (catch-up of missed runs, frequency changes apply at once, discovery forced after a geo change, build after every events run), a failing stage does not stop the build; last_run in state.json; job summary with event changes and AI usage. Workflow run.yml: daily cron, "Run workflow" button with job choice and dry run, concurrency group, data branch worktree and bot commit, Pages deploy.
• Phase 5 AI readability: schema.org Event JSON-LD on every day page (program items as subEvent), llms.txt, public data/events.json, iCalendar feeds per language, sitemap.xml with hreflang, robots.txt allowing AI crawlers, Open Graph and description per day.
• Phase 8 content and wrap-up: content pages from instance/content/ (about, "Why is my event not here?", imprint, privacy) in DE and EN with values filled from the config and [TODO] markers listed by letsgau check; feedback page (form service or mailto: fallback, honeypot, link to the issue forms); changelog page; CNAME from base_url; letsgau init creates a fresh instance from examples/instance/ (interactive or with options).
• Phase 4 website (letsgau run --job build): day pages DE/EN for display_days with coverage sentence, quick links, date bar with counts, sections (all day collapsed, morning, afternoon, evening), expandable events with sources, opening hours, program items and status; tag chips with include/exclude, search, grouping by time/category/place, swipe, sliding header, settings in localStorage; everything readable without JavaScript. New option site.timezone. Workflow site-check.yml (Lighthouse, curl).
• Phase 3 merge and status: fuzzy matching of crawled and stored events (dates, municipality, start time, normalized titles), one Haiku call per run for unclear pairs and the topic check of new events (prompts/merge.md), misses/unconfirmed for events not found again in crawled municipalities, cancellations and postponements from the crawl, cleanup; merge summary in the log (also with --dry-run). AI steps can run without web tools.

• Whitelist URL templates with {municipality_slug} (one URL per municipality) and letsgau sources [--check] (instance whitelist updated accordingly).

• Phase 2 event crawl (letsgau run --job events): places in batches of places_per_request, one AI call per batch with prompts/events.md, answers forced into a JSON schema and validated with Pydantic (28-day window, places in the area, tags from the list, exactly one primary tag and source, no blacklisted source); invalid entries are dropped and logged.

• AI runner for both modes: Claude Code headless (claude -p --json-schema, only WebSearch/WebFetch, blacklist as WebFetch deny rules, API keys removed from its environment) and the Claude API (web_search_20260209 / web_fetch_20260209 with blocked_domains, strict submit_result tool).
• Search budget (max_web_searches_per_run) spread over the batches, runtime limit, whitelist check per batch, usage in data/state.json.
• Models Event and EventsFile; id-based merge into data/events.json (sources appended, past events deleted); fuzzy matching follows in Phase 3.
• whitelist.yaml entries take optional places.
• Places get their municipality (OSM admin_level=8) in the geo stage; the crawl batches by municipality (districts like Herzogsägmühle are searched with Peiting), --limit-places counts municipalities.

• Workflow crawl-check.yml: live dry run of the crawl on GitHub Actions with the CLAUDE_CODE_OAUTH_TOKEN secret; LETSGAU_RECORD_DIR saves raw answers.

• Phase 1 geo stage (letsgau run --job geo): OpenRouteService isochrone (driving-car) via the new HeiGIT host api.heigit.org/openrouteservice, places from the Overpass API, drive times from the ORS matrix; writes data/places.json, data/state.json and geo/isochrone.geojson.

• Geo fingerprint of center and drive time: recompute only on change, then state.json gets force_discovery; changing only the place filters reuses the stored isochrone. --force recomputes anyway.
• include_places (looked up by name in OpenStreetMap) and exclude_places.
• Pydantic models Place, PlacesFile and a first State.
• Workflow geo-check.yml: live geo check on GitHub Actions (manual or on changes to the workflow) with the ORS_API_KEY secret; prints the place list and uploads re-recorded fixtures.

• .env is loaded by letsgau run; tests replay ORS/Overpass fixtures (tests/fixtures/geo/) with a script to re-record them.

• Phase 0 scaffold: repository structure, pyproject.toml, Ruff, pytest and CI.

• instance/config.yaml loaded and validated with Pydantic; errors report file, line and field.
• CLI python -m letsgau with check, run and init (stages are stubs).
• First versions of README.md, CLAUDE.md, docs/ARCHITECTURE.md, docs/CONFIGURATION.md, docs/DATA_MODEL.md, docs/OPERATIONS.md.
• Specification docs/CONCEPT.md and task list docs/ISSUES.md.
• Issue forms for bug, feature request and source suggestion.

Changed

• Instagram lookup: the discovery passes list the cafés, bars and venues they saw (venues_seen, the venues pass from the municipality's directory), and the lookup call gets one fixed search per place instead of a free task; the log shows known accounts that were checked and kept, with the evidence.
• Instance: ten more Instagram accounts in the list (cinema, brewery, cafés, beer garden, markets, clubs), found with site:instagram.com searches.
• Discovery's Instagram searches use site:instagram.com <town> …; plain searches like " Café Instagram" returned no Instagram links.
• Discovery's venues pass runs three Instagram searches per municipality (cafés, bars, clubs) and takes account names from the links in the search results; cafés, bars and pubs are reported even without known events.
• Cheaper discovery: the default model is Sonnet instead of Opus (models.discovery), depth.discovery_searches_per_request is 20 instead of 30, and known sources that are whitelisted or were rated or hit in the last 90 days are only listed in the prompt, not opened and reported again; older ones are re-rated.
• Instagram: the step runs last in each batch, so accounts found by the batch's other calls are read in the same run.
• Content pages (about, missing event) say that Instagram business and creator accounts are read, but not Facebook or other social networks.
• Instance: four more Instagram accounts in the list (a pub, a cinema, a bar and a café).
• Cheaper discovery: the default model is Sonnet instead of Opus (models.discovery), depth.discovery_searches_per_request is 20 instead of 30, and known sources that are whitelisted or were rated or hit in the last 90 days are only listed in the prompt, not opened and reported again; older ones are re-rated.
• Instagram: the step runs last in each batch, so accounts found by the batch's other calls are read in the same run.
• Fewer tokens per events run, same coverage:
• blacklist entries can be a url instead of a domain: such a page is not read as a known source (for a second front end of an event database that is already read elsewhere), while the domain stays usable;
• re-check calls leave out the known events that a source call of the same run reported again, and are skipped when none remain;
• the open search defaults to 13 searches (one call) instead of 40, and every run starts the query checklist at a different point, so over a few runs every query is still run.
• Open search: a call that stops before using its searches is no longer asked a second time. In a live run the second attempt repeated the same queries and found nothing new, at about $0.60 per run.
• Contact form: it is sent in the background and a thank-you message replaces the form on the page (the service's raw answer is no longer shown). The email address is mandatory, and there is a new type "General inquiry" (the default). Without JavaScript the form is posted normally.
• "Why is my event not here?" page: explains why Facebook and Instagram are not read, asks for the title and a short description as text (not only date, time and place), and that a public calendar only helps if its dates can be read, with a simple web-search test.
• site.contact_email is optional when site.feedback_form_endpoint is set (at least one is required). With a form service, the about, missing-event and privacy pages and llms.txt link the form instead of showing the address; the imprint shows the address only if one is set, otherwise the form.
• The search field is on every page, not only on day pages. On content pages the results replace the page text and link to the events; category filters apply only where the category chips are shown.
• The clear button of the search field is a white cross on a light circle instead of the browser's blue default.
• Search also matches category names in every language: "Konzert" or "concert" finds all events of that category, not only those with the word in their title.
• Shorter coverage sentence: "Events within N min of · D days ahead".
• The browser title is the same on every page: site name and region label; the day-specific text stays in the description and the Open Graph title for previews when sharing.
• Event crawl prompt: regular church services are not events; special services and religious occasions are.
• New category religion (Religion) in the default tag list; the category palette has 12 colours, so up to 12 categories get distinct colours.
• Event crawl: the re-check of the known events of the next days is its own call (up to 20 events each) instead of part of the open search. The open search budget is split into calls of about 13 searches, each with a part of the checklist (the log lists each search query once). Checklist topics include youth centres and live music, and press previews count as a source (with the search result when the article is blocked).
• The Run workflow uploads every AI answer and, in subscription mode, the transcript of each call as the artifact ai-recordings; the log lists the search queries of each call.
• Category chips are folded by default on phones (until the visitor opens them). Grouped by category or place, each group can be folded; the choice is remembered. Size budget of app.js raised to 12 KB.
• About page template: explains that Instagram and Facebook do not allow their pages to be read automatically.
• Search covers all displayed days instead of the current day: results are listed by date (earliest first), 50 at a time with a "Show more" button, after at least 3 characters. Cancelled, postponed and unconfirmed events are marked as on the day pages. Category filters apply to the results. The index data/search.<lang>.json is written by the build and fetched only when a search starts.
• On a phone the search field takes the whole header row while it has focus.
• The category chips can be folded in and out with a "Categories" button; the choice is remembered, and the button shows the number of active filters.
• Every page (content pages, feedback, changelog) has the quick links and the date bar in the header, so a click on a day leads back to the events.
• "About" and "Is my event missing?" are linked in the header (next to the coverage sentence on desktop, below it on a phone) instead of the footer.
• Merge: a series split by weekday ("Market (Fridays)" / "Market (Saturdays)") is one event. Weekday words are ignored when comparing titles, events only match when they share an actual day, an entry covering only part of a stored series keeps the stored schedule and texts, and clear duplicates already in events.json are merged without the AI. The crawl prompt asks for one entry per series with the times per weekday in opening_hours.
• Site design in black and yellow: dark header band with a thin yellow line, yellow for the selected day, "Today" and buttons; the page itself stays light.
• Date bar: arrow buttons and mouse-wheel scrolling on desktop, so all 21 days are reachable; a thin scrollbar without JavaScript.
• Category chips wrap onto several lines instead of scrolling sideways.
• The "Weekend" quick link is removed.
• Day pages show one list instead of time-of-day sections: all-day, multi-day and events without a start time come first, then events by start time. The "Group by: Time" option is now "None"; category and place grouping stay.
• Event crawl prompt: no summary entries such as a cinema's whole programme; each film or event is reported on its own.
• In subscription mode the usage summary shows the cost at API prices (the CLI's total_cost_usd) instead of 0, as a measure of plan usage; discovery usage is saved in state.json too.
• max_runtime_minutes and the search budget apply per AI call (and stage), not per batch.
• The Crawl check workflow runs only on demand, no longer on pull requests that change a prompt: it uses about as much of the plan as a real run.

Fixed

• Markets, exhibitions and other multi-day events with opening hours show that day's opening time in the day list, with the closing time below it, and sort by it (before: only "until ").
• Social media (Instagram, Facebook, TikTok, …) is never added to data/sources.json: events from Instagram posts no longer add instagram.com as a source when hits are counted, discovery answers with such URLs are skipped, and existing entries are removed.
• Search results expand like the cards on the day pages: description, venue, source and a link to the day (the index has the fields t, v, l).
• Start times that differ between the dates of a film or series: opening hours can be given per date ("days": "2026-11-21"), which wins over the weekday, and the events prompt asks for them instead of a weekday pattern or one time for all dates. A fresh crawl result updates the stored opening hours also when the stored event keeps its longer schedule.
• Films and other events shown on several days at a different time per day show that day's start time in the day list (with "to " below) and sort among the timed events; their details say "Today 20:00" instead of "Open today 20:00–20:00".
• Social media (Instagram, Facebook, TikTok, …) is never added to data/sources.json: events from Instagram posts no longer add instagram.com as a source when hits are counted, discovery answers with such URLs are skipped, and existing entries are removed.
• Search results expand like the cards on the day pages: description, venue, source and a link to the day (the index has the fields t, v, l).
• Start times that differ between the dates of a film or series: opening hours can be given per date ("days": "2026-11-21"), which wins over the weekday, and the events prompt asks for them instead of a weekday pattern or one time for all dates. A fresh crawl result updates the stored opening hours also when the stored event keeps its longer schedule.
• Films and other events shown on several days at a different time per day show that day's start time in the day list (with "to " below) and sort among the timed events; their details say "Today 20:00" instead of "Open today 20:00–20:00".
• Social media (Instagram, Facebook, TikTok, …) is never added to data/sources.json: events from Instagram posts no longer add instagram.com as a source when hits are counted, discovery answers with such URLs are skipped, and existing entries are removed.
• Search results expand like the cards on the day pages: description, venue, source and a link to the day (the index has the fields t, v, l).
• Start times that differ between the dates of a film or series: opening hours can be given per date ("days": "2026-11-21"), which wins over the weekday, and the events prompt asks for them instead of a weekday pattern or one time for all dates. A fresh crawl result updates the stored opening hours also when the stored event keeps its longer schedule.
• Films and other events shown on several days at a different time per day show that day's start time in the day list (with "to " below) and sort among the timed events; their details say "Today 20:00" instead of "Open today 20:00–20:00".