extract-deck

name	extract-deck
description	Reconstruct a player's exact 60-card PTCG decklist from their Kaggle ladder replays (any submission_id → deck JSON). Use to copy/clone/mimic a top-ranker's or opponent's deck, or seed a counter. Triggers — "copy charmq's deck", "リプレイから60枚を複製", "clone sub <id>", "extract the top player's decklist".

Extract a deck from replays

Given a Kaggle submission_id, download that submission's ladder replays and reconstruct the owner's MAIN deck (= the most frequent exact 60-card list across their games) as a JSON ready for a build. Works for any submission_id — ours or a rival's reached by traversing episodes.

Steps

1. Get the target submission_id. If the user gave one, use it. Otherwise find it: it appears as an opponent's submission_id in any episode you can list.

   • From our submissions / a player we faced:

     uv run python -c "from kaggle.api.kaggle_api_extended import KaggleApi; a=KaggleApi(); a.authenticate(); \
       [print(x.submission_id, x.team_name) for e in a.competition_list_episodes(OUR_SUB_ID) for x in e.agents]"
     

   • To reach a higher-rated player, list one of their episodes (their submission_id from the step above) and read off their opponents. See workspace/exp011_meta_watch/top_meta.py for the traversal pattern.
   • Cross-reference team names with kaggle competitions leaderboard pokemon-tcg-ai-battle --show.

2. Run the extractor (replays are cached under references/raw/replays/, gitignored):

   cd workspace/exp011_meta_watch
   uv run python extract_deck.py <submission_id> [out.json]
   

   • Omit out.json to just print the decklist + archetype.
   • Pass an output path (e.g. ../exp012_nonex/<name>_deck.json) to save the 60-card-id JSON list for a build script.

3. Report the archetype label, ex-card count, and the Pokémon/Trainer/Energy breakdown (the script asserts it sums to 60).

4. (Optional) turn it into a submission with the /build-submit skill (deck JSON → tar.gz → smoke-test → submit after approval).

Notes

• Needs Kaggle API auth + the cabt engine; run via uv run.
• Pilot choice (updated 07-14): if a PUBLIC DEDICATED pilot exists for the archetype, prefer it over generic-piloting — the two current-era wins (v020 Archaludon, v021/v023 LO) both shipped a public notebook's own pilot (+ our crash wrapper + targeted patch), while generic-piloting a foreign deck caps far below the dedicated pilot (deck⊗pilot, 5+ confirmations). The generic lucario_v2 route (Crustle v004, non-ex v006) is the fallback when no dedicated pilot exists.
• Also reusable: exp053_bandpool/extract_band_decks.py (per-archetype REAL decklists from any cached replay set — how the band pools are built), and scanning all caches for a deck's signature card ids to find same-deck players (exp054-C).
• A copied deck is only as good as its pilot — validate with /build-submit's smoke test before trusting it.