🚰 Sluice

A strategy composer for 1inch Aqua. ETHGlobal Lisbon 2026.

1inch Aqua's SwapVM lets you ship programmable market-making strategies as bytecode ([opcode_index][args_length][args_data]). That is powerful and almost nobody can author it safely — and once a strategy is shipped it is immutable; the only exit is dock() or a _deadline unwind. There is no patching a bad strategy after the fact.

Sluice turns a plain-language intent plus a budget you already hold into concrete, risk-rated Aqua strategy recommendations you review and ship in a single signature.

You describe what you want and select the tokens/amounts you're willing to commit. Sluice composes one or more strategies, checks them with deterministic code, and presents them; you sign one Multicall that ships them. Tokens never leave your wallet — you are the maker. Sluice acts once, at creation time: it is never in the per-swap path and does not manage your positions afterwards.

How it works

Three integrations, each load-bearing:

• 1inch Aqua + SwapVM — the venue and the language. We run against a Base mainnet fork at a pinned block (real, already-deployed Aqua), self-deploying only our own contracts. Strategies are authored through a fixed six-slot grammar an LLM can safely fill, then compiled deterministically to SwapVM bytecode. A SluiceTaker produces the fills a fork otherwise lacks.
• 0G — private, verifiable recommendations. Composition runs inside an Intel TDX enclave and is signed; the chain verifies the signature to prove provenance (a real 0G TEE produced it, our committer authorised it). Review splits by what has a right answer: a deterministic validator rejects non-compliant recommendations and re-infers (never rewriting a signed one), while a reviewer agent judges what does not — risk rating, intent match, band sensibility. Traces live encrypted in 0G Storage for audit.
• The Graph — market & book context. A net-new subgraph indexes the user's own Aqua book (Position/Fill); composed external subgraphs supply market data (depth, realised vol, fees). Together they form the MarketContext the recommendation is built from.

Status

Hackathon build, pre-alpha. Nothing here is audited; don't point it at real funds.

Running so far: a generic 1inch Aqua subgraph — the first that exists — live on Ethereum mainnet and Base against real protocol activity (subgraph/); and the 0G inference CLI in packages/arbitration-sdk (npm run infer -- "…"), which runs a prompt on a live 0G provider and prints the answer plus a verifiable EIP-191 proof.

The venue is a Base fork sharing Base's chainId, so a wrong-network mistake looks perfectly correct — guarded by a fork probe plus an explicit SLUICE_ALLOW_MAINNET opt-in, with addresses pinned in config/addresses.8453.json.

Documentation

Notion is the source of truth for the concept, the plan and every schema; this repo is the implementation. The page index and the precedence rule live in CLAUDE.md.

Local docs are organised by the same four features as Notion:

F1 — Aqua & the Strategy VM	Aqua + SwapVM: the fork venue, the slot grammar, the compiler, the taker
F2 — Verified Private Recommendations	0G: sealed inference, signed recommendations, validator, encrypted trace
F3 — Market & Book Context	The Graph: the user's book subgraph + composed market context
Wiring & Delivery	Shared vocabulary, the per-user request flow, transaction shape, demo

PRDs and issue files live in docs/prds/.

A note on language

• Recommendation → compiles to one or more Positions; one recommendation is one user signature over a Multicall.
• The over-committed book / "fillability" framing is parked future work, not the current product. So is "headroom" and its cousins (balanceFloor, largestAllOrNothingDraw, exposureHeadroom) — using them unqualified is a real bug.