Skip to main content

Policy Engine

The policy engine is the core of OnlyFence. It runs a pipeline of independent check functions in sequence. Each check evaluates a trade intent and returns pass or reject.

How It Works​

Every action your agent requests goes through the policy pipeline before execution:

Check Interface​

Every policy check implements the PolicyCheck interface:

interface PolicyCheck {
  name: string;
  description: string;
  evaluate(intent: TradeIntent, ctx: PolicyContext): CheckResult;
}

interface CheckResult {
  status: 'pass' | 'reject';
  reason?: string;
  detail?: string;
  metadata?: Record<string, unknown>;
}

Current Checks​

Token Allowlist​

Verifies both the source and destination tokens are in the configured allowlist.

[chain.sui.allowlist]
tokens = ["SUI", "USDC", "USDT", "DEEP", "BLUE", "WAL"]

Spending Limit​

Two-part check:

  1. Per-trade limit — trade USD value must be under max_single_trade
  2. Daily volume — rolling 24h total must stay under max_24h_volume
[chain.sui.limits]
max_single_trade = 200.0
max_24h_volume = 500.0

Config-Driven Loading​

Checks are registered based on which config sections exist. No config section = check not loaded.

Adding a New Check​

Adding a new guardrail requires three steps:

  1. Implement PolicyCheck — one file
  2. Define config schema — one TOML section
  3. Register in loader — one line

Zero changes to existing checks or pipeline logic.

Planned Checks​

Oracle Failure Handling​

If the price oracle is unreachable after 3 retries:

  • OnlyFence uses a fail-closed approach
  • Falls back to cached price for up to 5 minutes
  • If cache is stale or absent, the trade is rejected
  • Token allowlist checks always apply regardless of oracle status