Spot only, one venue, one adapter. Only exact-input swaps through the Orca Whirlpool swap instruction on explicitly allowed pools. No leverage, lending, perps, bridging, aggregators or multi-hop routes. SPL Token mints only; Token-2022, transfer hooks and other extensions are rejected.
One session at a time. Registering a new session key replaces the previous one. Sessions last at most 30 days.
At most six assets and six venues per policy.
Reference prices are Pyth sponsored feeds. They update on the feed's own cadence (observed: about 50 s on mainnet, slower on devnet). Markets without a defensible reference (for example devnet test tokens) cannot be traded under a policy that requires reference protection; the UI shows the protection as unsupported and the policy must switch it off explicitly.
Slippage tolerance is not price impact. The reference floor bounds how far the minimum output may sit below the reference; the preview shows the estimated impact separately, and a large trade can still fail at the venue.
The drawdown circuit breaker is a stop, not a floor. It blocks further eligible trading once equity is observed below the threshold. It needs fresh prices for every policy asset, marks itself stale after deposits/withdrawals/basis changes, and cannot undo finalized trades.
Equity and P&L are shown only when every held asset has a fresh reference price; otherwise the dashboard says "Unavailable" with the reason. Fees are shown as the difference between quoted and received output when both exist.
The quote is advisory. The Orca SDK quote (or the spot fallback) is used to pick min_out; the chain enforces the minimum atomically. A worse-than-expected pool state makes the trade fail, never overpay.
The hosted worker is trusted with the session signing material under the local-file signer (encrypted at rest; the host and the key together control it). Encrypting a key in a file or database does not remove server trust.
The remote-http signer adapter is untested against a real signing service.
The LLM providers (Anthropic structured output, OpenAI-compatible JSON schema) are integrated but were not exercised against live API keys during development; the scripted provider is what the local flows use.
Devnet and mainnet deployments require operator steps; see Setup. Mainnet readiness has not been assessed and the program is unaudited.
The strategy is narrow by design: target-band rebalancing between approved assets. The model may only take the computed candidate (or a smaller amount) or hold.