DIME
Get the picksJoin

PROJECT · findings note

project_wnba_sport_agents_playbook

Concrete work breakdown to build the WNBA sport-agents adapter. Three touchpoints, each with NBA regression risk; documented 2026-05-21 after attempting the build and hitting the complexity wall.

← All findingssource · memory/project_wnba_sport_agents_playbook.md

WNBA sport-agents adapter is THE last gap between data-hydrator's WNBA snapshot (verified live with FIX 1-5 firing on tonight's GSV@NYL) and the automated blog_generator LLM pipeline. Without it, python -m core.blog_generator --league wnba --date X fails at adapter registration.

Why this isn't a quick mirror of NBA

NBAAdapter is ~539 lines + 429-line prompts + 37-line archetypes. Surface-level "copy and replace NBA → WNBA" misses three real dependencies on shared infrastructure:

Touchpoint 1: core/hydrator_client.py is NBA-hardcoded

Today (verified 2026-05-21):

  • HydratorClient.get_snapshot(game_id) builds URL as /api/v1/snapshot/{game_id} — no league prefix.
  • HydratorClient.get_snapshots_for_date(date) same shape.

For WNBA, the URL must be /api/v1/wnba/snapshot/{game_id} etc. (data-hydrator's multi-sport routes per project_sds_is_the_data_layer).

Two options:

  • (a) Parameterize HydratorClient on optional league arg, default None=NBA-historical-path. Touches NBA's currently-working pipeline.
  • (b) WNBAAdapter owns its own snapshot fetch path. Duplicates ~30 lines of HTTP/cache logic.

(a) is cleaner; (b) is safer. Either is bounded but needs care.

Touchpoint 2: extract_team_bundle / extract_player_bundles assume NBA snapshot shape

In core/hydrator_client.py:88-, both extractors read:

  • snapshot.get("stats") — typed for NBA, in raw_sections for WNBA
  • snapshot.get("game") — typed for NBA, in raw_sections.game.game (envelope!) for WNBA
  • snapshot.get("injuries") — typed for both (post my data-hydrator work)
  • snapshot.get("lineups") — typed for both (post f114e9b/87befcc)
  • snapshot.get("players", {}).get("players", {}) — typed for both (post f114e9b/87befcc)

The first two will silently return None/empty for WNBA, producing empty bundles. Fix options:

  • (a) Update the extractors to fall back to raw_sections.game.game for WNBA. Touches NBA path (but the fallback is harmless if first lookup succeeds).
  • (b) Modify assembleMultiSport in data-hydrator to populate typed snap.Game for WNBA (parsed from raw_sections.game.game). Cleaner architecturally; requires another data-hydrator deploy.

(b) parallels what hydrateMultiSportPlayerBundles did for Players + Lineups — the same pattern applied to Game would close this gap once and for all.

Touchpoint 3: WNBAAdapter implementation (~150 lines)

Mirror NBAAdapter, swap:

  • __init__: pass league="wnba" to HydratorClient (after touchpoint 1)
  • get_config(): WNBA SportConfig(league_key="wnba", display_name="WNBA", odds_api_sport_key="[REDACTED]", season_months=[5,6,7,8,9,10], max_players_per_game=10, prop_types same as NBA — basketball)
  • get_team_ids(): WNBA team_ids — sportlib has 15 WNBA teams at wnba.teams with team_id field. Build the abbrev→id map from sportlib's registry instead of hardcoding.
  • _get_espn_team_id(): WNBA ESPN team IDs — sportlib likely doesn't have these. Either hardcode (look up from ESPN) or skip the ESPN fallback for WNBA (the snapshot covers it).
  • get_odds_markets(): sportlib wnba.odds_api_markets has 5 market keys (verified 2026-05-21).
  • get_prop_stat_map(): identical to NBA (basketball stats).
  • calc_combo_prop(): identical to NBA (pra/pr/pa/ra).
  • build_game_slug(): identical to NBA.
  • fetch_games_for_date(), fetch_team_bundle(), fetch_player_bundle(): identical pattern to NBA, just use WNBA team_ids map.

Reuse NBA's prompts (get_agent_prompts() returns NBA PROMPTS) since basketball discipline is identical. The LOCKDOWN PRE-CHECK already added (sport-agents fcd4e2e) works for WNBA the moment WNBAAdapter forwards prop_diagnostic in summarize_player_data — and since we'd inherit that method from NBAAdapter, it carries through automatically.

Mirror NBA's archetypes.json (same positions).

Sequenced execution

  1. Pick HydratorClient strategy: (a) parameterize or (b) duplicate
  2. Pick extract_*_bundle strategy: (a) fallback or (b) typed-snap.Game in data-hydrator
  3. Build WNBAAdapter subclass / direct
  4. Build wnba/prompts.py (re-export NBA PROMPTS)
  5. Build wnba/archetypes.json (copy NBA)
  6. Verify python -m core.blog_generator --league wnba --dry-run parses + runs without erroring (full pipeline smoke)
  7. Backtest one WNBA final-game date to verify the LLM prop_picks output is sensible (FIX 1-5 gates fire correctly)

Estimated: 2-4 focused hours. NOT a 1am ship.

What's already done that makes this easier

  • WNBA snapshot reachable end-to-end (commits b2dba69, f114e9b, 87befcc)
  • WNBA snap.Players + snap.Lineups + snap.Injuries all typed
  • LOCKDOWN PRE-CHECK added to player_agent prompt (fcd4e2e), works for both NBA and WNBA the moment a WNBA adapter forwards prop_diagnostic
  • Health endpoint covers WNBA (4ba5d34) so a regression in the WNBA pipeline will fire before picks are attempted

See also: project_wnba_lockdown_live (the data layer this builds on), project_sds_is_the_data_layer (architectural anchor), feedback_verify_before_extend (why scope-checking matters).