Two players. One deck. Swipe YES/NO through live binary prediction markets — on-chain escrow pays whoever read the market better.
Note
The API runs on Render's free tier and sleeps when idle. The first request after a nap takes ~30s to wake — if the lobby looks empty, give it a moment and reload. Practice and Bot Arena need no wallet.
Prediction markets are powerful but look like a Bloomberg terminal. Swipe-betting apps proved retail loves feel-based prediction UX — but they're shallow, custodial, and have no real opponent.
DreamSwipe is both. Two players swipe through the same deck of live DreamDEX
event contracts. Scoring is the actual economics — netResult = realizedValue − entryCost — so the winner is whoever genuinely read the market better, not
whoever tapped faster.
Left: the swipe deck — one binary market per card, YES/UP or NO/DOWN, with a live strike and spot.
Centre: Bot Arena, dealing real BTC/ETH windows discovered on chain.
Right: the Practice intro. All three screens above are signed out — no wallet needed.
| Mode | Stake | Wallet | Chain |
|---|---|---|---|
| Practice | — | not needed | fully simulated |
| Bot Arena | — | not needed | live market data, 4 strategy agents |
| Free PvP | 0 | yes | same engine, money flow gated off |
| Staked PvP | tUSDC | yes | escrowed, winner takes the pot |
Free and Staked share one code path. The tier gates only whether collateral moves, and a Free duel carrying a stake reverts on chain.
Every address below is live. eth_getCode returns real bytecode for all three,
and decimals() on the collateral token returns 6 — read from the token, not
assumed.
| Chain | Somnia Shannon testnet · 50312 (0xc488) |
| RPC | https://dream-rpc.somnia.network |
| DreamSwipeDuel | 0x6b554BaFC2031b72AeB100cDAd111c2f70f2cC5c |
| ↳ deploy tx | 0xd35fee41…09b7809b · block 483885846 · 38,853,709 gas |
| SeasonPrizePool | 0xB380814066dcb5d0b4d4968742bFdC005D1FB4a7 |
| ↳ deploy tx | 0xd0b86f9d…575c140f · block 483959460 · 15,573,469 gas |
| Collateral | tUSDC 0x70a86D88…1BB25d8E — 6 decimals, permissionless faucet(uint256) |
| Gas token | STT — players sign swipes, a relayer pays |
| Venue | DreamDEX Event Contracts (binary Up/Down markets) |
Full record, including gas findings:
deployments/somnia-testnet.json
Every number here came from a command that was actually run.
This scans MarketCreated logs straight off chain — no indexer, no API key,
no wallet. Run it yourself with bun --filter server run check:dreamdex:
Eight live windows discovered by log scan, a real five-card deck dealt from
them, and can start a duel: YES. The full run also prints each order book —
they are genuinely empty on testnet, and the tool reports
no quotes resting rather than inventing a mid price.
Executed against the deployed contract with real transactions
(bun --filter server run e2e:somnia):
createDuel → joinDuel (2nd wallet) → revealDeck
→ 2× relayed swipe → settleCard ×3 → finalize
Read back from chain: status Complete · 3/3 settled · p0 +600000 · p1 −400000.
Those numbers are the economics exactly — backing UP at 0.40 and winning
redeems 1.00 (+0.60); the loser forfeits their 0.40 premium.
Neither player sent a transaction. Both signed EIP-712 messages and the relayer submitted them. No wallet popup mid-duel, no gas from the player.
Staked duels are verified with real collateral (e2e:staked): 5 tUSDC escrowed
per side, 10 held, the winner paid the whole pot, and escrow back to exactly 0
with nothing stranded. Seasons the same way (e2e:season): create → fund →
allocate → finalize → claim, with the claim sent by the winner, claimable
correctly 0 before finalization, and a second claim rejected with
AlreadyClaimed.
apps/web React 19 · Vite · Tailwind v4 · wagmi + viem
apps/server Bun · WS relay · matchmaking · relayer · keeper · MMR
apps/contracts-evm Solidity 0.8.24 (Foundry) · DreamSwipeDuel · SeasonPrizePool
packages/dreamdex venue boundary: types, CLOB math, Bot Arena agents
packages/ui shared shadcn components
There is no Sui left. Zero @mysten imports and zero @mysten
dependencies: the Move package, DeepBook adapter, Pyth oracle stream, zkLogin
auth and sponsored-gas service are all gone, replaced by the EVM equivalents
above. What was reused versus rewritten is recorded in
docs/MIGRATION_AUDIT.md.
One rule holds the design together: nothing outside packages/dreamdex may
import the venue SDK. This repo already lost ~12 days of playability when a
previous venue was torn down, so "what are the cards" is swappable without
touching "run a duel".
specs/06_FRONTEND.md says "No fake live values", and that is a type
signature here, not a convention:
- anything not derivable returns
null, never0— a zero spread and an unknown spread must never look alike insufficientLiquiditycarries a partial fillable size, so the UI shows what the book can really fill- no tx hash, order id, fill or price is ever synthesized; the adapter throws
VenueUnavailableErrorinstead getFillsthrows rather than returning[], because an empty array is indistinguishable from "not implemented"
The player signs an EIP-712 message binding
(duelId, cardIdx, direction, nonce, deadline). The relayer submits it and pays
the gas. It can neither forge nor replay a swipe, and it cannot touch escrow.
premium and filled are derived server-side from the live venue, never
taken from the request body — a client-supplied premium would let a player
understate their entry cost and inflate their own PnL.
Bots see exactly what a human sees, and that is enforced by the type rather
than by policy: PredictionContext has no winner, settlementPrice or
resolved field, so an agent cannot see an outcome even in principle. A
test pins the exact key set, so adding one later trips a tripwire.
Difficulty raises the signal threshold a bot needs before acting — it never changes what the bot can see.
Prizes are escrowed, not promised. Allocations freeze at finalization, so an
operator cannot rewrite who gets what after seeing the leaderboard; each winner
claims once; and the surplus sweep is bounded by funded - allocated, so an
unclaimed prize can never be swept out from under a late claimer.
bun install
cp .env.example .env # add a funded Shannon key for writes
bun dev # web :5173 + server :3001tUSDC has a permissionless on-chain faucet — call faucet(uint256) on the
token; no external step is needed for collateral. STT for gas comes from the
SomniaHacks group: https://t.me/+XHq0F0JXMyhmMzM0
bun typecheck # 4 packages
bun run test # 264 tests
bun build
bun --filter server run check:dreamdex # live venue probe, no key needed
cd apps/contracts-evm && forge test # 71 tests, incl. 2 fuzzed invariantsTip
forge on some machines is shadowed by an unrelated CLI of the same name.
If forge test behaves oddly, use ~/.config/.foundry/bin/forge.
Both cost real debugging time and are documented in
docs/DREAMDEX_INTEGRATION.md.
Somnia deploy gas. Foundry estimated 3,394,836 gas; the deploy actually
needed 38,853,709 — 11×. Worse, an under-limit deploy mines with
status: 0 and burns the entire limit while eth_call simulation still
succeeds, so simulation cannot catch it. Three deploys failed that way before
60M worked. The estimate is unreliable in both directions: the tUSDC faucet
call was estimated at 1,379,707 gas, actually used 253,138 — and a 500,000
limit still mined status: 0 and burned the lot.
The spec pack's DreamDEX constants were wrong in three ways — package name,
collateral token, and decimals. Building on them unchecked would have mispriced
every order by 10¹² on mainnet with nothing reverting to warn you. Corrections
are recorded in docs/MIGRATION_AUDIT.md §2.2.
This is not a hypothetical. A stale 9-decimal formatter survived the migration
into the UI and rendered every practice strike as $0 — the value was correct
on the wire and 1000× too small on screen, with nothing to revert. Fixed in
c683146, now pinned by tests.
MIGRATION_AUDIT.md |
what was reused, replaced, and verified |
DREAMDEX_INTEGRATION.md |
venue facts, traps, and a real bug the health check caught |
DEPLOYMENT.md |
deploying the web app, server, and contracts |
deployments/somnia-testnet.json |
addresses, tx hashes, gas findings |
Apache-2.0