[moremeds]

Problem

MarketAggregates.pcr_oi was set from UW screener put_call_ratio. UW's spec (docs/uw-samples/unusual_whales_api_spec.yaml, /api/screener/stocks) defines that field as put volume / call volume. So pcr_oi has been a copy of pcr_vol:

• prod pcr_history: 38,376 rows (2026-01-02 → 2026-11-06); 38,342 have pcr_oi == pcr_vol, 0 differ, 34 have NULL pcr_vol.
• Affected readers: watchlist "PCR (OI)" tile, pcr_delta_30d, market_overview.positioning, the Trade Insights AI prompt input.
• Example: TSLA showed pcr_oi 0.9397 (= pcr_vol); put_oi/call_oi is ~0.80.

Fix

• pipeline.aggregates_from_screener() computes pcr_oi = put_oi / call_oi and pcr_vol = put_vol / call_vol from the screener's own counts. Zero puts → 0; missing/zero calls → NULL. run_single_stock uses it.
• Repository.repair_pcr_history_oi() (in _MarketDataMixin) + scripts/backfill/pcr_history_oi_repair.py rewrite pre-fix rows from options_volume_daily. The match is the latest session <= snapshot_date within --max-lag-days (default 4), because the UTC-stamped snapshot_date can be a day ahead of the ET session (weekend rows). No match → NULL, never the old volume ratio. Dry run by default; idempotent; bounded by --before.
• A script, not a migration: migration DML must be insert-only.

Post-deploy (manual, needs operator confirmation)

Pass the day after the deploy as --before, and run only after the deploy day's 18:15 ET flow refresh has written options_volume_daily for that day.

python scripts/backfill/pcr_history_oi_repair.py --before <day after deploy>          # dry run
python scripts/backfill/pcr_history_oi_repair.py --before <day after deploy> --apply

Prod check (read-only, 2026-11-08): of 17,203 rows that map to their last-writing scan run, the repair reproduces that scan's own OI ratio on 16,661; 275 match an earlier session (morning-only last scans saw pre-publish OI); 267 have no exact match (248 from notes-tagged, non-full-scan runs).

Until it runs, pcr_delta_30d mixes the old volume ratio with the new OI ratio.

Tests

• tests/unit/test_pipeline_aggregates_from_screener.py — UW's AAPL spec example (OI ratio ≠ volume ratio), zero/missing counts.
• tests/integration/storage/test_pcr_history_oi_repair.py — dry run rolls back, apply values, weekend row takes Friday, no-source → NULL, post-cutover row untouched, rerun changes 0.
• Local: full lint+unit gate green (3195 unit passed); integration 5 passed; dry run on option_wizard_local: 34,657 in scope, 34,652 with OI source, 5 → NULL.

Review

Tribunal: Opus 5.5 lead + Fable 5.1 + GPT-6-Astra (herdr panes). Fixed: cutoff guidance (both), negative --max-lag-days now raises (Astra), stamping docstring (Fable, Astra), tests for zero call OI + per-ticker isolation (Fable; mutation-checked). Astra's "repair picks a later session" CRITICAL was withdrawn after the prod evidence above. All findings verified resolved by the raising seat.

Not in scope

snapshot_date uses date.today() in a UTC container, so an ad hoc rescan after ~20:00 ET stamps the next day. scan_runs.aggregates JSON history keeps the old pcr_oi value; nothing reads old runs for PCR history. Separate issues.

🤖 Generated with Claude Code