SWMMike's Tekken Repo
Mike's Tekken 8 training system: battle data collected automatically, reviewed
after every session and every month, and turned into anti-character guides.
The loop: play → collect → review → prepare → play.
What runs on its own
| What | When | Where |
|---|
| Collect battles from ewgf.gg (Pro API) | Scheduled hourly; GitHub actually runs it every 3–5 hours | .github/workflows/collect-battles.yml → scripts/collect_battles.py |
| Add ratings from Wavu Wank | Scheduled midnight, 8 AM, 4 PM Eastern (runs often start hours late), plus once per End session | .github/workflows/collect-wank.yml → scripts/collect_wank.py |
| Build one anti-character guide | Fridays, 9:47 AM | Claude scheduled task, reads queue.md |
| Monthly ranked review with coach's notes | 1st of the month, 9:55 AM | Claude scheduled task → reviews/monthly/ |
| Coach's take on each session review | Nightly, 10:13 and 11:13 PM Eastern; skips sessions already covered in chat | Claude scheduled task → reviews/coach/ |
GitHub delays scheduled jobs when it's busy, so run times drift. No games are
missed: each ewgf pull covers your last 100 battles and each Wavu Wank pull
your last ~200 games. End session always fetches fresh battles and ratings first (one Wavu Wank
request, ranked reviews only), so reviews are complete no matter when the
scheduled runs last happened.
A single failed ratings run (Wavu Wank didn't answer) is caught up by the next one.
What you run
- End-of-session review: GitHub → Actions → End session → Run workflow.
Pick last session or today, and ranked (default) or all. The report
lands in
reviews/LATEST.md and on the run's summary page.
- Tekken app (replaces the old overlay poller): see
app/README.md.
Runs on your PC: stage tracker and guides at localhost:8787, live poller,
stream overlay (Twitch strip or full scoreboard) and a GitHub sync every
15 minutes, so games played on stream reach this repo without waiting for
the hourly collector. Your ewgf key and GitHub token never leave the PC.
- Old stream overlay:
overlay/ (PowerShell poller). Superseded by the app;
don't run both, or they split the hourly ewgf request limit.
- Guide sources for any character:
python scripts/fetch_sources.py "Armor King"
downloads and parses the Wavu movelist + counterplay and TekkenDocs frame data +
anti-strat into build/sources/<Character>/sources.json (not committed), with a
Wavu-vs-TekkenDocs block-frame cross-check. The weekly guide task uses it.
- Backlog import from saved Wavu Wank pages (one-time, e.g. Jin or Armor King):
python scripts/import_wank.py --since 2026-01-01 page1.html page2.html.
Never commit the saved pages; they contain names and IDs.
The data
data/battles.jsonl is the database: one battle per line, ranked from
Jan 1, 2026 plus all modes since Oct 3, 2026. data/battles.js (for the
website) and data/battles.csv (for the Google Sheet) are copies rewritten on
every save.
| Field | Meaning | Source |
|---|
battle_at | When the game ended (UTC) | both |
mode | ranked / quick / player / group | ewgf (Wank games are all ranked) |
my_char, opp_char | Characters | both |
opp_rank | Opponent's rank tier | ewgf only (null on backlog games) |
result, rounds | W/L and round score, yours first | both |
stage_id, stage | Stage (names in data/stages.json) | ewgf only (null on backlog games) |
my_rating, opp_rating | Each player's rating going into the game, per character | Wavu Wank (ranked only) |
Privacy rule: no player names, tags or IDs, ever. It's enforced in code:
every write goes through save_rows() in scripts/collect_battles.py, which
keeps only ALLOWED_FIELDS and stops the run if a value looks like a player
ID. Adding a field means changing that list on purpose.
Coaching rules
- Coaching reads ranked only. Quick match is played loosely to feed the
collector and isn't graded.
- Numbers under 10 games are hints, not verdicts.
- Opponent strength is judged by rating (every ranked game has it) rather than
rank tier (only ewgf games have it).
Where things are
| Path | What |
|---|
index.html | Website: stage tracker, live records, Matchup Guides tab |
guides/ | Anti-character guides + manifest.js (powers the Guides tab) |
queue.md | Order the weekly guide task follows; edit freely |
sidestep.md | Community sidestep chart, used as secondary evidence in guides |
reviews/ | Session reviews (LATEST.md = newest) and monthly/ reviews |
scripts/ | Collectors, reviews, Wank parser and importer, guide source fetcher |
overlay/ | OBS stream overlay and its local poller |
data/ewgf-schema.json | Field list of the ewgf API (no personal values) |
data/lag-test-2026-10-10.json | Result of the Pro lag test: median 0.7 min from game to API (worst 9.7) over 26 games |
Outside the repo
- Google Sheet "Tekken 8 Battle Log": reads
data/battles.csv. The Mode filter is set to Ranked.
- Secrets:
EWGF_API_KEY is a repository secret. Nobody, Claude included, sees its value.
- If a collector fails, GitHub emails you about the failed run (Settings → Notifications → Actions).
