| name | pipeline-doctor |
| description | Use when checking whether the data + model pipeline is healthy, when returning to the project after time away, or when a workflow looks stuck. Runs the read-only health snapshot and triages any breach. |
| argument-hint | [--refresh] |
/pipeline-doctor — One-screen pipeline health + triage
Strictly read-only on the money path — never places bets, never writes the
ledger. It composes pipeline_health() (fit/odds freshness, persisted Stan
diagnostic drift, orphaned bets, bankroll/drawdown) into a single status table
and helps triage anything that is WARN or FAIL.
1. Refresh the snapshot
cd /Users/brynjolfurjonsson/sports && Rscript scripts/07_healthcheck.R
This prints the per-check table, lists the breaches, and writes
data/health/status.json (overall status OK < WARN < FAIL, plus PAUSED
for an off-season cell). It always exits 0 — it is observability, not a gate.
2. Add operational context
git -C /Users/brynjolfurjonsson/sports log --oneline -8 -- data/
gh run list --repo metill-is/sports --limit 12
3. Triage each breach
For every WARN/FAIL row, map the check to its runbook in
docs/runbooks/ and follow the symptom → diagnose → fix → verify steps:
| check | likely cause | runbook |
|---|
odds_freshness FAIL | Lengjan scrape failing or off-season | docs/runbooks/stale-odds.md |
fit_freshness FAIL/WARN | fit.yml failing, or results not moving | docs/runbooks/failed-fit.md |
divergence_drift / rhat_drift WARN | model degrading under the abort gate | docs/runbooks/failed-fit.md |
orphaned_bets WARN | a bet whose match never settled | docs/runbooks/orphaned-bet.md |
bankroll FAIL | realised PnL ran the pool to/under zero | docs/runbooks/orphaned-bet.md |
check_error | a health sub-check itself errored | inspect the message; usually a missing/corrupt partition |
Confirm-intent before "fixing": a PAUSED cell (basketball/handball off-season)
and the schedule fail-open are intentional — surface them, do not treat
them as faults.
Why a skill (and a SessionStart banner), not a hot-path hook
Health is checked on demand (this skill), as a twice-daily cron
(healthcheck.yml), and via a cheap SessionStart banner that reads the
committed status.json. A PostToolUse-after-fit hook was deliberately
rejected: the Stan abort gate already blocks a bad fit synchronously, so such a
hook would add always-on cost for no marginal signal.
