Skip to content

neromtoobad/crossfire

BranchesTags

Repository files navigation

CROSSFIRE — the Live World Cup Arena

Five AI agents call the World Cup with real, chain-capped money. You back the ones with a proven record, and fade the rest. The agents can't bluff — conviction is USDC the chain refuses to let them spend past their cap.

Twitter tipsters are free and unaccountable. CROSSFIRE makes an AI forecaster's confidence costly and on-chain: every call is backed by USDC drawn from an ERC-7710 delegation the chain enforces. A bluffing agent literally cannot afford to look confident, and its record is a public, Brier-scored track record you can study before you stake. Accountability via cryptographic primitives, not promises.

Built for the MetaMask Smart Accounts Kit × 1Shot × Venice AI Dev Cook Off. Submission deadline 2026-06-15.

The idea in one screen

   WORLD CUP 2026 MARKETS   (outright winner, golden boot, every group fixture…)
        ▲
        │  agents read the matchup + buy evidence (x402)
        │
   ┌────┴──────────────────────────────────────────────────────┐
   │                  FIVE AI AGENTS                            │
   │                                                            │
   │   PHOENIX     ORION       NEXUS        ECHO       VEGA      │
   │   tactics     team-news   momentum     xG/data    contrarian│
   │      │           │            │           │          │      │
   │      ▼           ▼            ▼           ▼          ▼      │
   │   each reasons via VENICE, then STAKES capped USDC on a side │
   └──────────────────────────┬─────────────────────────────────┘
                              │  the call publishes
                              ▼
   THE ARENA   (live feed of calls — who called what, how much they staked)
                              │
                              │  you study the agent's record, then…
                              ▼
   FADE  or  FOLLOW   (grant a chain-capped mandate via ERC-7715, place your bet)
                              │
                              ▼
   THE VAULT   (your positions + mandates — revoke any of them, instantly)

The moat: adversarial conviction, metered on-chain. Conviction isn't a number an agent claims — it's USDC it actually staked, capped by a delegation the chain enforces. The five agents are opposed (a contrarian, VEGA, fades the favourites), so the net of their committed capital is the signal. Every drawdown is a real on-chain spend under a caveat enforcer. The costly signal is genuinely costly.

The five agents

All share one Venice account but get different role prompts, voices, and evidence. Internal role keys are stable; these are the public handles.

Agent Archetype Reads Internal role
PHOENIX Tactics engine shape, matchups, game management MacroScout
ORION Team-news scanner lineups, injuries, fitness NewsHawk
NEXUS Momentum model form, belief, the run of results CrowdPulse
ECHO xG & data model expected goals, shot quality, set pieces BookWatcher
VEGA Contrarian adversary fades favourites, calls the bottlers Skeptic

A call publishes only if ≥3 of 4 agree, VEGA doesn't veto, there's ≥5 points of edge over the market line, and the bond fits the cap. Each agent's Brier score sets a budget multiplier (sharp 1.5× → miscalibrated 0.7×) that sizes its next stake — being right literally compounds.

The pages

Route What it is
/ The Arena — dark broadcast landing; rotating World Cup trophy wallpaper, top agents by ROI, the live "Who wins the World Cup?" outright market, live feed
/markets Every market + all 72 group fixtures, tabbed (Live · Results · stage)
/calls/[id] A single call — the agents' votes & Venice reasoning, Fade/Follow with a chain-capped bet, and a Venice-generated verdict card
/agents/[handle] An agent's profile — ROI, calibration by stage, and its actual Venice reasoning per call with HIT/MISS outcomes
/leaderboard The standings — Brier-scored, with the accountability loop explained
/portfolio The Vault — your positions and chain-capped mandates, with one-tap revoke
/run The guided walkthrough — one signature → agents fight → bet lands → revert → relay → score
/lab The proof — on-chain receipts for every primitive, plus the tools to run each one live

On-chain proof (Base Sepolia + Base mainnet)

Every claim links to its tx on Basescan. Full receipts: PROOF.md.

What landed Tx
The revert — over-cap redemption refused with ERC20TransferAmountEnforcer:allowance-exceeded. No code stops it; the chain does. 0xa8d4…ee45
A2A redelegation — an agent redeems its capped sub-budget through the user → arena → agent chain 0x5cdc…ba41
x402 evidence buy — buyer-with-delegation pays for evidence, real USDC moves 0x0bd9…cf23
A staked call lands — an agent's capped USDC stake settles on its side, on-chain 0x44a7…7a4c
1Shot mainnet relay — confirmed (200), EIP-7702 in-flight upgrade, gas paid in USDC 0x5a09…2651

Reproduce it yourself

No coverage badge over mocked chain calls — every core claim is a script that hits the real chain and prints a Basescan link.

Command Proves Chain
npm run proof The hero shot: signs a 50 USDC mandate, redeems 1 USDC (ok), attempts 60 USDC → reverts at the caveat enforcer. Base Sepolia
npm run duel:skeleton A2A redelegation, leaf-to-root chain, over-sub-cap reverts, child.delegator == parent.delegate. Base Sepolia
npm run conviction x402 evidence buy (metered USDC) → Venice reasons over the evidence. Base Sepolia
npm run test:unlock:direct The unlock path end-to-end: EOA pays USDC, server verifies, content unlocks. Base Sepolia
npm run relay:bet One real Base-mainnet 1Shot relayer_send7710Transaction — 7702 upgrade, gas in USDC. Base mainnet
npm run council:test The 5-agent Venice panel: 4 votes + VEGA veto + quality gate. Venice only. off-chain
npm run test:venice-x402 Venice paid inference over x402 (TEE model path). off-chain

Three guarantees are enforced by the build, not by trust:

  • Venice is the only model provider. grep -rniE "groq|api\.openai\.com|api\.anthropic" lib app scripts returns nothing — the sole LLM base URL in the repo is https://api.venice.ai/api/v1. If Venice is down, the agents do not decide. There is no fallback.
  • The cap is the contract's, not the code's. There is no if (amount > cap) reject in the agent path. The over-cap revert comes from MetaMask's ERC20TransferAmountEnforcer, on-chain.
  • The result is the oracle's, not ours. We never type an outcome. The agents' record is graded on real, already-played matches; live markets settle against UMA's Optimistic Oracle (the decentralized, disputable, on-chain resolver Polymarket uses), read in lib/polymarket.ts and exposed at GET /api/settlement?slug=… (returns status, umaStatus, and the on-chain resolvedBy). Anything without a real result yet stays PENDING, forever.

Tracks targeted

Track How CROSSFIRE earns it
Best x402 + ERC-7710 Agents pay evidence APIs via buyer-with-delegation; users place bets / unlock via the same capped-delegation primitive. Both metered on-chain.
Best A2A Coordination Redelegation chain: user → arena orchestrator → each agent's own keypair + budget. Real keypairs redeem through the full chain; child.delegator == parent.delegate verified on-chain.
Best Use of Venice AI Venice is the only model provider (grep-enforced). Every agent reasons via Venice; the on-screen verdict card is rendered by Venice's image endpoint (venice-sd35) — visible Venice output in the main flow.
Best Use of 1Shot Relayer One real Base-mainnet relayer_send7710Transaction, EIP-7702 in-flight upgrade, gas paid in USDC, webhook handler wired.
Smart Accounts Kit (main flow) The user grants a capped spend via MetaMask's native ERC-7715 Advanced Permissions dialog when they Fade/Follow — and revokes it in The Vault.

Code usage (MetaMask submission checklist)

Direct links to the exact code, per the Dev Cook Off submission checklist.

Smart Accounts Kit usage

Advanced Permissions (ERC-7715)

Delegations

Redelegation

  • Create: duel.ts#L46-L75 and duel-engine.ts#L124-L140 — child delegation with parentDelegation = signedRoot (the user → orchestrator → Bull/Bear chain; child.delegator == parent.delegate).

x402

1Shot API usage

Venice AI usage

Feedback

Honest developer-experience notes from building on the kit (Best Feedback track):

  • ERC-7715 requestExecutionPermissions is gated by wallet build. It only resolves on the MetaMask extension with smart accounts / Flask; on a stock wallet it throws an unknown-method error. We added explicit detection + a clear message, but a documented capability check (wallet_getCapabilities-style) would save everyone the trial-and-error.
  • Smart-account payments break naive x402 verification. When the payer is a MetaMask smart account, a USDC.transfer is redeemed through the DelegationManager, so the outer tx.to is the manager, not the asset. Server-side verification must read the USDC Transfer event in the receipt, not tx.to/calldata. A note about this in the x402 buyer guide would prevent a common false-reject.
  • Embedded-wallet chain default. With MetaMask Embedded Wallets, the wallet seeded a different default chain (Ethereum Sepolia, 11155111) than our target (Base Sepolia, 84532), so transactions asserted a chain mismatch. We auto-switch (wallet_switchEthereumChain + 4902 add-chain) before redeeming; clearer guidance on setting the embedded default chain would help.
  • 1Shot was the smoothest piece. getCapabilities → getFeeData → send7710Transaction with a single EIP-7702 authorizationList entry worked first try on Base mainnet, and pointing destinationUrl at a public webhook endpoint (a Vercel route, no tunnel) gave clean status pushes. Splitting fee + work into two transfers in one redemption was the one non-obvious bit.
  • Testnet/mainnet split. The relayer is mainnet-only while the enforcement story is cheapest on Base Sepolia; a testnet relayer (even rate-limited) would let teams prove the full 7710-via-1Shot loop without spending real USDC.

Social Media

Posts from @neromtoobad sharing the CROSSFIRE build, tagging @MetaMaskDev (Best Social Media Presence track):

Each post shows how MetaMask Advanced Permissions (ERC-7715) shaped the UX — the one-signature capped mandate and the over-cap revert.

What works now vs what's next

The arena is built and running end-to-end — the chain primitives, the 5-agent Venice panel, the World Cup feed (82 markets + 72 fixtures, 49 resolved), agent profiles, Fade/Follow, the Venice verdict card, The Vault, and the proof console. The production build is green. What remains is packaging.

Done

  • ERC-7710 mandate signing + the user-facing ERC-7715 grant (components/GrantCouncilMandate.tsx) — the kit in the main flow
  • ERC-7710 revert proof at the caveat enforcer (npm run proof)
  • A2A redelegation chain (user → arena → agent), real keypairs redeem on-chain
  • x402 seller route + buyer-with-delegation — metered evidence buys; Venice paid inference over x402
  • Venice is the sole engine (lib/venice.ts) — reasoning + the verdict-card image (lib/verdict-card.ts)
  • 1Shot mainnet relay client + webhook — one real Base-mainnet relay confirmed
  • The arena landing with the rotating World Cup trophy wallpaper; agent profile pages; The Vault (positions + mandates + revoke)
  • World Cup feed — 82 markets, 72 group fixtures, 49 resolved; the resolution loop moves agent records
  • Brier-scored standings with the accountability loop (record → budget multiplier → next stake)
  • wagmi + MetaMask connect; dark broadcast-gold design system pinned across the app

Run it locally

git clone https://github.com/neromtoobad/crossfire.git
cd crossfire
npm install
cd contracts && forge build && cd ..

cp .env.example .env.local
# Fill in: VENICE_API_KEY, ONESHOT_API_KEY, 4 EOA private keys

npm run dev        # http://localhost:3000
npm run build      # production build (green)

# on-chain proofs (need a funded .env.local)
npm run proof              # the ERC-7710 revert (hero shot)
npm run duel:skeleton      # A2A redelegation
npm run conviction         # x402 evidence buy → Venice
npm run relay:bet          # real Base-mainnet 1Shot relay

Architecture

USER (browser + MetaMask)
  · browse the arena (no wallet needed)
  · Fade/Follow → grant a chain-capped mandate (ERC-7715), place a bet
  · The Vault → revoke any mandate, instantly
        │ HTTP / SSE
        ▼
CROSSFIRE SERVER (Next.js 16, App Router)
  · reads markets via viem
  · runs the 5-agent panel through Venice (the only provider)
  · pays evidence APIs via x402 buyer-with-delegation
  · applies the quality gate, posts the bond (ERC-7710), publishes the call
  · renders the verdict card via Venice's image endpoint
  · relays the hero proof on Base mainnet through 1Shot (7702 + USDC gas)
        │ on-chain
        ▼
BASE SEPOLIA / MAINNET
  · DelegationManager (kit) · USDC · BinaryMarket (lib/markets.json) · 1Shot relayer

Tech stack

  • Next.js 16 (App Router, Turbopack) + TypeScript + viem 2.x
  • @metamask/smart-accounts-kit@1.6.0 — delegations, caveats, scopes, ERC-7715
  • wagmi 2 + MetaMask connector
  • openai@6 pointed at Venice's OpenAI-compatible endpoint — never at OpenAI itself
  • Foundry — BinaryMarket.sol, dependency-free, ~75 lines
  • Base Sepolia for delegations / redemptions / the revert; Base mainnet for the one 1Shot relay

Venice models: qwen3-235b-a22b-instruct-2507 (chat) + venice-sd35 (verdict-card image) — chosen to avoid Venice-routed third-party models that would muddy the "Venice as sole engine" claim.

Build note: the repo uses NodeNext module resolution with explicit .js import extensions (the tsx operational scripts need it). Turbopack bundles this correctly; the build's tsc step is non-blocking (next.config.mjs → typescript.ignoreBuildErrors) because of the App-Router/NodeNext friction. Type-check separately with npm run typecheck.

Architectural decisions worth knowing

  • The bond is an ERC-7710 mandate, not a separate escrow. The treasury signs a delegation capped at the bond; settlement redeems against it. No new contracts.
  • Two distinct x402 surfaces. Agents pay APIs (server-to-server); users place bets / unlock (browser-to-server). Same primitive, both on-chain.
  • Fresh salt on every delegation (generateSalt()) — otherwise the kit produces deterministic hashes and the enforcer accumulates spend across runs.
  • buyOnBehalf(buyer, isYes, amount) — the mandate's Erc20TransferAmount scope only allows USDC.transfer, so a bet is two steps: redeem transfers USDC to the market, then anyone credits shares.
  • Venice is the only LLM provider — enforced by grep, not just intent. Switching providers mid-build would invalidate the Venice track claim.

Bug catalog — landmines already cleared

  • Wrong execution encoding — use createExecution + encodeSingleExecution, not hand-rolled abi.encode.
  • 1Shot permissionContext is the full delegation object array, not encoded hex bytes.
  • 1Shot executions[0] must be a USDC.transfer to the feeCollector, or "no valid payments to the feeAddress" is thrown.
  • signAuthorization wants the raw privateKey, not a viem account.
  • Public Base Sepolia RPCs serve inconsistent block heights — backdate BlockNumberEnforcer.afterThreshold by ~1000.
  • Counterfactual smart accounts can't sign via ERC-1271 — deploy + fund before the first redemption.
  • wagmi hydration race — cookieStorage + cookieToInitialState.
  • Mixed import conventions break a single tsconfig — see the build note above.

What CROSSFIRE deliberately does NOT do (yet)

  • Mirror Polymarket directly. It lives on Polygon with UMA-resolved ConditionalTokens; cross-chain settlement is its own project. The agents bet on our BinaryMarket on Base; the agent code is identical to what would run against Polymarket.
  • Resolve calls fully on-chain. Resolution needs an oracle; for now match results move agent records at the data layer (the standings + payout direction), with manual settlement of bonds.
  • Verify webhook signatures. 1Shot uses Ed25519 against their JWKS; we accept-and-poll getStatus. Production fix documented.

None change the architecture or invalidate a track claim.

Files worth reading first

License

MIT.

About

Adversarial prediction-market agents on a chain-enforced mandate. MetaMask Smart Accounts Kit × 1Shot × Venice.

crossfire-nu.vercel.app

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages