track-conferences

name	track-conferences
description	Watch tracked AI conferences and ingest them as their papers go live. For the current and next year, web-search each conference's date; when a conference is within ~3 months, check (web search + `ai-trend probe`) whether its accepted papers are published on OpenReview/thecvf; if so, ingest and analyze it and update CONFERENCES.md. Use for the monthly/scheduled conference-watch run.

Track Conferences (web-search driven watch)

Encodes the running logic: discover dates → wait until ~3 months out → check if papers are published → ingest & analyze. You (Claude) drive this using your web search tools plus the deterministic CLI. Run from the repo root with PYTHONNOUSERSITE=1 ./env/bin/ai-trend ....

Inputs

• today — the current date (from the environment).
• config/conferences.json — tracked venues (label, source, openreview_group, month).
• CONFERENCES.md — the maintained table of dates + source URLs (state of record).

Procedure

1. Scope: current & next year, what's unprocessed

• Compute the current year and next year from today.
• Run ai-trend check-sources to see which conference-years already have a source URL, and ls data/<year>/ to see what's already ingested.
• Build the worklist: tracked conferences for {current year, next year} that are not yet ingested.

2. Discover dates (WebSearch)

For each worklist item missing a date in CONFERENCES.md:

• WebSearch e.g. "<Conference> <year> dates" / "<Conference> <year> accepted papers".
• Read the official site (WebFetch) to confirm the conference date / camera-ready / proceedings date.
• Record the date into CONFERENCES.md (add the row if absent).

3. Window check: is it ~3 months away?

• If the conference date is more than 3 months in the future, do nothing now (it'll be re-checked next run). Stop here for that item.
• If it is within 3 months (or already passed) and not yet ingested, proceed.

4. Are the papers published? (web search + probe)

• Deterministic check first:

  PYTHONNOUSERSITE=1 ./env/bin/ai-trend probe <CONF> <year>
  

  It constructs the expected source URL (OpenReview content.venueid or thecvf <CONF><YEAR>?day=all) and reports AVAILABLE (N papers) or not yet.
• If probe says not yet, WebSearch/WebFetch the venue (OpenReview group page, thecvf, or the official proceedings) in case the URL pattern changed this cycle. If you find a working URL, update CONFERENCES.md with it.
• If still nothing is published, stop (re-check next run).

5. Ingest & analyze

When papers are available, run the full pipeline (see the add-conference skill):

PYTHONNOUSERSITE=1 ./env/bin/ai-trend ingest-url "<the live URL>"
# then: candidates -> curate (curate-topics) -> assign -> trends -> export-site

Then fetch citations for the new conference-year's emerging-topic papers (ai-trend citations ... --scope emerging) if desired, and snapshot (ai-trend snapshot-citations).

6. Record & publish

• Update CONFERENCES.md: ensure the row has the date + the working URL.
• Open a PR with the regenerated data/site (and the CONFERENCES.md update).

Notes

• Only the date discovery and fallback URL discovery use web search; the availability check and ingestion are deterministic CLI tools.
• New venues (not in config/conferences.json) need an entry first (key/label/tokens/month/source/openreview_group).
• Be conservative: never ingest a partially-posted venue — re-check next run if the paper count looks far below the expected size.