AgentSessionOptions

interface AgentSessionOptions<W = unknown>

import type { AgentSessionOptions } from '@driftengine/ai';

Properties

Name	Type	Description
`agentId`readonly	`string`
`policy`readonly	`AgentPolicy`
`provider`readonlyoptional	`AiProvider`	Absent is a supported configuration, not a degraded one. The floor is enough.
`budget`readonlyoptional	`Budget`
`model`readonlyoptional	`string`
`tools`readonlyoptional	`ToolRegistry<W>`	Tools, and the admission guards they declared. Absent admits everything.
`world`readonlyoptional	`W`	The consumer's world, handed to a guard unchanged.
`whileBusy`readonlyoptional	`WhileBusy`	What an observation arriving while a request is out does. More `queue` and `replace pending` are deliberately absent. Both described what to do with a second request, and there is never one — the state machine has no edge on which it could be issued.
`maxPendingObservations`readonlyoptional	`number`
`dedupeWindowMs`readonlyoptional	`number`
`log`readonlyoptional	`CommandLog`	Where accepted model commands are recorded, so a run can be replayed exactly.
`maxIntentMs`readonlyoptional	`number`	A ceiling on how long any one intent may hold the slot, in milliseconds. More An intent with an unknown extent ends when the consumer says so, which is the contract — the consumer is the thing that knows whether the walk finished. But a consumer that forgets leaves the agent on one intent forever, which is the blocked agent arriving through the back door. Off by default, because a default here would be a guess about behaviours this package cannot see, and a wrong one would cut short a legitimately long action. What it costs when set: a genuinely long intent is reclaimed by the floor. What would make it wrong: a consumer whose intents legitimately outlast any figure worth writing down, which should set nothing and call `complete`.
`interruptProvider`readonlyoptional	`AiProvider`	A second, faster provider asked only when something interrupts. More §46's tiering: a small or local model answers "something happened, react now" in well under the strong model's latency, while the strong one is still composing the next intent. Additive — absent, preemption behaves exactly as it does without it. What it costs: a second provider to configure and a second bill. What would make it wrong: an interrupt model slow enough that the floor beats it to the answer, at which point it is buying nothing the floor did not already give.