BurningDrill
Source of truth for the poker-trainer POC: a personal-first NLHE cash drill trainer (no live play, no solver in the web app). pnpm monorepo — Next.js App Router UI plus static strategy packs. Solvers never ship in the deployable web app or Vercel/Docker web image.
Default pack: packs/toy-6max-expand v0.5.0 (ships_solver_code: false). The browser loads the copy under apps/web/public/packs.
Quick start
pnpm i
pnpm --filter @poker-trainer/pack-format build
pnpm test
pnpm dev # syncs packs, then http://localhost:3000
No .env is required. Do not commit secrets.
Other commands:
| Command | What it does |
|---|
pnpm build | Build pack-format, sync packs into public/, build Next app |
pnpm test | Vitest: PackManifest + drill pack schema + scoring |
pnpm typecheck | tsc --noEmit in packages |
pnpm validate-packs | CLI validate toy packs, including toy-6max-expand, and weeklock |
Requires Node 20+ and pnpm.
Layout
apps/web/ Next.js UI (Vercel Hobby target)
packages/pack-format/ Zod PackManifest (CDN) + compact drill spots + scoring
packs/toy-6max-expand/ Default drill pack v0.5.0 (synced → public/packs)
packs/toy-6max/ Slice-1 compact drill JSON
packs/toy-6max-multistreet/ Flop/turn/river practice nodes
packs/weeklock/ Week-lock fixture
packs/examples/… Full PackManifest fixture (CAS blob refs)
tools/pack-builder/ Stub only — offline/CI solver runner lives here later
tools/generate-toy-expand/ Regenerates toy expand / multistreet JSON
apps/web/scripts/sync-packs.mjs copies packs/ → apps/web/public/packs/ (predev and prebuild).
How packs are loaded (MVP vs production)
MVP ($0): compact drill packs under packs/ (default toy-6max-expand v0.5.0) are copied to apps/web/public/packs and fetched by the browser. No object store required.
Production target: Cloudflare R2 (not wired in this slice):
packs/{pack_id}/{version}/manifest.json
blobs/{sha256} # content-addressed, Cache-Control: immutable
Schema: packages/pack-format/src/manifest.ts (schema_version, pack_id, ships_solver_code: false, blob sha256 refs, provenance, …). Validate with validatePackManifest. Toy drill UI keeps a separate compact spot JSON path (DrillPackSchema / validatePack) so demos stay simple.
Never put AGPL solvers (TexasSolver, postflop-solver, …) in the web Docker/Vercel image. Separate pack-builder (offline/CI) vs web. Pack artifacts = manifest + blobs only. See NOTICE and pack provenance.ships_solver_code: false.
Scoring rule
Live drills grade with gradeDrillNode only.
- Best — the highest-EV action. A line within 0.01bb of that EV is still Best.
- Correct — Best, or EV loss ≤ 0.03bb, or the action’s mix is at least 0.15.
- Wrong — anything else.
EV loss (bb) = max(0, bestEv − chosenEv).
User stub
localStorage key poker-trainer:user holds { id, display_name, created_at, pack_prefs } — shape documented for a later Postgres users table. Saved drill settings: poker-trainer:saved-drill.
CI
.github/workflows/ci.yml on push to main and on pull requests: pnpm i --frozen-lockfile → pack-format build → pnpm test → pnpm typecheck → sync packs → web build. No solver install.
Deploy notes
- Aimed at Vercel Hobby; do not deploy solvers.
- AGPL: solvers offline/CI only — not legal advice; counsel before commercial launch.
Next steps
tools/pack-builder job that emits PackManifest + blobs- R2 publish (
aws s3 sync compatible) with immutable blob cache headers
- Multi-user Postgres behind the existing
user_id stub
- Expand toy → real cash lines / table sizes
