sports-update

name	sports-update
description	Use when refreshing models or running the full pipeline. Runs ingest + odds + fit + decide + publish for specified leagues.
argument-hint	[ingest|odds|fit|decide|publish|all] [--league LEAGUE] [--sex male|female] [--force]
context	fork
effort	high

/sports-update — Full pipeline update

Wraps the five scripts/0N_*.R entry points for the active Icelandic leagues. Each script has its own freshness guard: 02_scrape_odds.R skips when there are no upcoming games; 03_fit.R skips when no new games have been played since the last fit. Pass --force to bypass.

Step 1: Parse arguments

Pick the default selector based on the user's phrasing:

User phrasing	Default action
"run the pipeline", "update the models", "refresh", "full update"	Run all five scripts in order
"just refresh fits" / "rescrape odds" / "republish"	Run a single script (mapping below)
explicit --league, --sex	Honour the user

Selectors:

• --league KEY — one league (football_iceland, basketball_iceland, handball_iceland)
• --sex SEX — restrict to male, female, or all
• --force — bypass the freshness guard in 03_fit.R and 02_scrape_odds.R

Single-step mapping (when the user asks for one phase only):

Phrasing	Script invocation
"ingest", "scrape results"	scripts/01_ingest_results.R
"odds", "scrape odds"	scripts/02_scrape_odds.R
"fit", "refit", "refresh models"	scripts/03_fit.R
"decide", "recompute recs"	scripts/04_decide.R
"publish", "regenerate JSONs"	scripts/05_publish.R
"settle", "resolve bets", "settle ledger"	scripts/06_settle.R

Step 2: Plan and warn

For full runs, warn that model fitting takes significant time:

• 1 league × 1 sex: ~5–15 min (1000 MCMC iters)
• All active × both sexes: ~1–2 h (basketball fits dominate)

The freshness guards mean that re-running with no new completed games short-circuits in <1 second per (league × sex) pair. So scheduled re-runs are cheap; fresh fits only happen when there's new data.

Approximate inference (Pathfinder, ADVI) is disabled on the current models — they crash. Always use the default sample method.

Step 3: Execute (detached launch for long runs)

Always use a detached launch for long fits. This skill runs with context: fork — a plain run_in_background: true Bash call inside the fork creates a process that dies when the fork terminates, which has silently killed pipeline runs partway through chromote-backed scrapes. nohup … & disown detaches the process so it survives.

For a full refresh (chained):

cd /Users/brynjolfurjonsson/sports && \
  LOG=/tmp/sports-update-$(date +%Y%m%d-%H%M%S).log && \
  nohup bash -c '
    Rscript scripts/00_active_competitions.R && \
    Rscript scripts/01_ingest_results.R && \
    Rscript scripts/02_scrape_odds.R && \
    Rscript scripts/03_fit.R && \
    Rscript scripts/06_settle.R && \
    Rscript scripts/04_decide.R && \
    Rscript scripts/05_publish.R
  ' > "$LOG" 2>&1 & \
  disown && \
  echo "Launched PID $! (log: $LOG)"

For a single-step run (replace 03_fit.R with the relevant script):

cd /Users/brynjolfurjonsson/sports && \
  LOG=/tmp/sports-update-$(date +%Y%m%d-%H%M%S).log && \
  nohup Rscript scripts/03_fit.R > "$LOG" 2>&1 & \
  disown && \
  echo "Launched PID $! (log: $LOG)"

Capture the PID and log path. Report both to the user so they can tail or kill if needed.

Monitoring. Poll the log, not ps. The fork's sandbox can hide processes from user-level ps even while running. A simple until loop on grep -qE "complete|Error" "$LOG" with a generous timeout (1–3 h depending on selector + step) is sufficient.

For short steps (02_scrape_odds.R, 04_decide.R, 05_publish.R) running inline (without nohup) is fine — they complete in seconds.

Step 4: Verify outputs

After completion, check the relevant Parquet stores:

Step	Verifying read
ingest	data/facts/results/sport=*/country=*/ partitions touched recently
odds	data/facts/odds/sport=*/country=*/scraped_date=*/ has a fresh scraped_date
fit	data/beliefs/latest/sport=*/country=*/sex=*/beliefs.parquet mtime is recent
decide	data/decisions/recommendations/sport=*/country=*/run_date=*/ has a fresh run_date
publish	data/publish/{football,basketball,handball}/iceland/{karla,kvenna}/*.json mtime is recent

Or, more uniformly, via DuckDB:

Rscript -e 'sports::rebuild_duckdb(); con <- DBI::dbConnect(duckdb::duckdb(), "sports.duckdb", read_only = TRUE); print(DBI::dbGetQuery(con, "SELECT sport, country, MAX(scraped_at) FROM odds GROUP BY 1,2"))'

Step 5: Offer to commit

If the pipeline modified data/, offer to commit. Tracked Parquet stores should be committed alongside model/code changes; ad-hoc rebuilds without data changes are fine to leave uncommitted.

cd /Users/brynjolfurjonsson/sports && git add data/ config/ && git status

Only commit if the user confirms. Use a descriptive message such as "Refresh fits + publish ({date})" or "Scrape odds ({date})".

Common workflows

Pre-betting refresh

# Update everything, then preview pending bets. Settle runs before decide
# so current_pool reflects realised PnL from any newly-resolved bets.
Rscript scripts/00_active_competitions.R
Rscript scripts/01_ingest_results.R
Rscript scripts/02_scrape_odds.R
Rscript scripts/03_fit.R
Rscript scripts/06_settle.R
Rscript scripts/04_decide.R
Rscript scripts/05_publish.R
Rscript scripts/preview_bets.R

Quick model test (smoke check)

To smoke-test a single fit manually with fewer iterations, call the model layer directly:

Rscript -e 'devtools::load_all(); fit_one(load_leagues()$basketball_iceland, "male", iter_sampling = 200, write_archive = FALSE)'

Just refresh decide + publish (after a manual fit)

Rscript scripts/04_decide.R
Rscript scripts/05_publish.R

Reference

• Pipeline entry points: scripts/0N_*.R (one per layer)
• Freshness predicates: R/pipeline-freshness.R
• League registry: config/leagues.yml
• Per-step R modules: R/{ingest,model,decide,publish}-*.R
• Stan models: Stan/{league_key}/{file}.stan
• Publishing scaffolds: R/publish-{football,basketball,handball}-iceland.R
• Bet placement (separate, local-only): /place-bets skill