Sheshiyer Blockchain / Web3 Core
Shared model for the blockchain-web3 cluster. On-chain code is immutable, public, and
adversarial — once deployed, anyone can call it, read its state, and reorder/observe its
transactions. That single fact drives every rule below.
1. The adversarial model (decide everything from here)
- Anyone can call any external function in any order, any number of times. Never assume a caller, a sequence, or a single invocation.
- Everything on-chain is public — "private" variables are readable; no secrets in contract state.
- Transactions are front-runnable / reorderable (MEV) — the mempool is visible; assume an adversary can sandwich, front-run, or back-run any state-changing call.
- Deployed code is immutable (absent an upgrade proxy) — a bug is permanent and exploitable for the contract's full TVL. There is no "ship and patch."
2. Non-negotiable contract-safety rules
- Checks-Effects-Interactions (CEI): validate → update state → then make external calls. The
canonical defense against reentrancy (+ a
nonReentrantguard for value transfers). →defi-amm-security - Access control: explicit, least-privilege roles; prefer two-step ownership transfer
(
Ownable2Step); never leave an unguardedselfdestruct/delegatecall. - Oracle manipulation: never price off a spot AMM reserve an attacker can move in one block;
use TWAP / multiple sources / circuit breakers. →
prediction-market-oracle-research - Integer & decimals precision: track token
decimalsexplicitly; multiply before dividing; beware rounding that leaks value; validate the donation/inflation attack on share-based vaults. →evm-token-decimals - Validate all inputs (the webview-is-untrusted rule, on-chain edition): the caller is hostile.
3. Routing within the cluster
| Concern | Spoke |
|---|---|
| AMM / vault / DeFi security review | defi-amm-security |
| Token decimals / precision bugs | evm-token-decimals |
| LLM/agent that touches a wallet or trades | llm-trading-agent-security |
| Hashing / keccak256 utility | nodejs-keccak256 |
| Prediction-market oracle / risk research | prediction-market-oracle-research, prediction-market-risk-review |
4. Non-advisory stance (hard guardrail)
The prediction-market and trading spokes produce research and risk analysis, not
financial advice. Never present an output as a recommendation to buy/sell, never auto-execute a
trade or move funds, and always surface assumptions + downside/risk. An agent with wallet access
(llm-trading-agent-security) must gate every value-moving action behind explicit human confirmation.
5. Tooling / version matrix
| Need | Tool |
|---|---|
| Build / test (Solidity) | Foundry (forge) or Hardhat |
| Static analysis | Slither, Mythril; forge invariant/fuzz tests |
| Standards | OpenZeppelin (audited primitives) over hand-rolled |
| Hashing off-chain | nodejs-keccak256 (match on-chain keccak256 exactly) |
6. Shared guardrails
- Assume adversarial callers, public state, and MEV — always.
- Checks-Effects-Interactions + reentrancy guards on every external interaction.
- No spot-reserve price oracles; TWAP / multi-source.
- Track decimals explicitly; multiply-before-divide.
- Audited primitives (OpenZeppelin) over bespoke crypto.
- Trading/prediction outputs are non-advisory; never auto-move funds; gate on human confirmation.
- Treat deployed code as permanent — audit + test (fuzz/invariant) before mainnet.