Update AI conference deadlines

/ai-deadlines is generated from the conferences and conference_deadlines tables. The upstream huggingface/ai-deadlines YAML repo is not imported anymore (PR #456). Don't edit YAML or open PRs there. Change data only through the paths below. Each path writes through the guarded, audited writer: it never deletes, past editions and past deadlines are immutable, and every edit leaves a catalog audit event with before/after snapshots.

How the pipeline works

A new edition needs a deadline or official conference dates (start_date). Editions with dates but no deadlines appear on the page as TBA after the upcoming deadlines, and the sweeps keep researching them until deadlines are published (PR #909). Run every command from the repo root.

Workflow

1. Check the pipeline before doing research yourself

uv run --env-file backend/keys.env python -m backend.scripts.conference_deadlines status
systemctl list-timers 'pwc-cron-*deadline*' --all --no-pager

• A sweep is running: it blocks new launches. Check its Job (step 4) and wait for it rather than duplicating its research.
• No sweep in the last ~8 days, or a run of failed: the pipeline is broken. Fix that first; editing rows by hand only hides it. The usual causes are below.

2. Find the real gaps

uv run --env-file backend/keys.env python .agents/skills/update-ai-deadlines/scripts/coverage.py --gaps --top-tier

• upcoming: fine.
• tba: dates are announced and deadlines are pending. The sweeps keep checking these.
• gap: nothing upcoming is listed.

For each gap row, read why the agents left it alone before searching yourself:

PGPASSWORD=$(grep '^PWC_DB_PASSWORD' backend/keys.env | cut -d= -f2-) \
  psql -h localhost -U paperswithcode_user -d paperswithcode_hf -At -c \
  "select s->>'series', s->>'status', left(s->>'reasoning', 400), s->'editions'
     from conference_deadline_sweeps, jsonb_array_elements(results->'series') s
    where id = (select max(id) from conference_deadline_sweeps where status = 'applied' and mode = 'weekly')
      and s->>'series' in ('icml','neurips')"

Most summer-deadline venues (ICML, NeurIPS, UAI, COLM, EMNLP, …) have no call for papers until a few months before. A gap with "official page 404 / no CFP" reasoning is correct, not stale. Biennial series (ICCV, ECCV) expect the next edition two years on. LREC is biennial too.

3. Research specific series now

When the user wants something checked immediately (a CFP just came out, or a deadline was extended), launch a manual sweep for just those series instead of editing rows by hand:

uv run --env-file backend/keys.env --env-file /etc/paperswithcode/production-cron.env \
  python -m backend.scripts.conference_deadlines launch --series icml,neurips

It refuses while another sweep is running. Series are researched in parallel. A manual sweep usually finishes in under 30 minutes and costs about $0.05–0.20 per series (weekly sweep 29 cost $2.84 for 30 series).

4. Watch a Job and apply it early

The local hf token gets 403 on namespace nielsr. Inspect Jobs with the service's credentials:

uv run --env-file backend/keys.env --env-file /etc/paperswithcode/production-cron.env python -c "
from backend.config import get_settings
from backend.services.conference_deadline_sweep_service import job_credentials
from huggingface_hub import HfApi
c = job_credentials(get_settings()); api = HfApi(token=c.token)
job = '<hf_job_id>'
print(api.inspect_job(job_id=job, namespace=c.namespace).status)
print('\n'.join(l[:200] for l in list(api.fetch_job_logs(job_id=job, namespace=c.namespace))[-10:]))"

Once the stage is COMPLETED, apply it now instead of waiting for the next refresh timer:

sudo pwc-cron refresh_conference_deadlines

Then read the sweep's changes_applied and per-series rejections in the ledger. A rejection is a guard decision, not an error. Report the useful ones to the user, for example "dates announced but past deadlines are immutable".

5. Replay proposals after a rule change

If a code change makes earlier rejections acceptable, re-plan that sweep's logged proposals against today's rows. Do the same if a collection failed but the Job logs are intact. Always run a dry run first:

uv run --env-file backend/keys.env --env-file /etc/paperswithcode/production-cron.env \
  python .agents/skills/update-ai-deadlines/scripts/replay_sweep.py <sweep_id> --series icml,iccv
# inspect, then add --apply

Replays are idempotent: already-applied changes plan to nothing. Run them from a worktree at the deployed commit, so the rules match production.

6. Correct one edition by hand (last resort)

Use this only for a verified official source the agents keep missing, or for an urgent correction. The admin endpoint replaces the edition's whole deadline list, so send every deadline, not just the changed one. Omitted deadlines are deleted. Dates are local times in each deadline's timezone (AoE, UTC, UTC-7, or an IANA name). short_name creates a missing edition.

curl -sS -X PUT https://paperswithcode.co/api/v1/conferences/<series>-<year>/deadlines \
  -H "X-Admin-Token: $(grep '^PWC_ADMIN_API_TOKEN' backend/keys.env | cut -d= -f2-)" \
  -H "Content-Type: application/json" \
  -d '{"short_name": "CVPR", "url": "https://cvpr.thecvf.com/Conferences/2027",
       "deadlines": [
         {"type": "abstract", "label": "Abstract Submission", "date": "2026-11-06 23:59:59", "timezone": "AoE"},
         {"type": "paper", "label": "Paper Submission", "date": "2026-11-13 23:59:59", "timezone": "AoE"}
       ]}'

Fetch the edition's current list from GET /api/v1/conferences/deadlines first and edit that. Never invent or estimate dates: aggregator sites (OpenCurious, WikiCFP, and the like) often publish guesses. Only use the conference's official page.

7. Verify live

curl -s "https://paperswithcode.co/api/v1/conferences/deadlines?v=$RANDOM" | python3 -c "
import sys, json; r = json.load(sys.stdin)
print(r['count'], 'editions,', sum(not c['deadlines'] for c in r['results']), 'TBA')
for c in r['results']:
    if c['slug'] in {'icml-2027', 'cvpr-2027'}: print(c['slug'], c['start_date'], [(d['type'], d['deadline_at']) for d in c['deadlines']])"

The page defaults to Top tier only (tier = 'a'). A new edition inherits its series' tier and tags. Data changes need no deploy. Code changes to the rules or the page go through the normal PR → Bugbot → merge → pwc-deploy flow.

Known failure modes

• Stuck running row: it blocks every launch. collect marks it failed once the Job is past its timeout plus 1 h. If the Job is truly gone, investigate before editing the ledger.
• Every aggregation fails with HTTP 400: the Job's pinned claude-agent-sdk bundles a Claude Code CLI that is too old for the aggregation model. Check _cli_version.py whenever you bump the model.
• SweepConfigurationError: missing … HF_TOKEN: you didn't pass --env-file /etc/paperswithcode/production-cron.env.
• Only a couple of changes after a big sweep: this is usually correct. Read the rejections first. Most are "past deadline is immutable", or a new edition without deadlines or dates.
• Tests: test_missing_credentials_fail_loudly fails under --env-file backend/keys.env because the key is present there. Run pytest backend/tests/test_conference_deadline_service.py without the env file.

Report

Tell the user:

• which sweeps ran and what they cost
• what changed: editions created, deadlines added or moved
• which gaps remain and why ("no official CFP yet" vs. a rejected proposal)
• the live verification