DriftEngineSystems
Intelligence
88 symbols, imported from @driftengine/ai.
Explained in Agents and behaviour.
Classes
AgentSession- One agent: a floor that always runs, and one slot holding what comes next.
AuthoritativeAgent- An agent whose model decisions are taken in one place.
Budget- A ceiling on what an agent may spend, enforced here rather than asked for in a prompt.
CommandLog- What was accepted, and when — beside
TickTracerather than inside it. DeterministicProviderLatencyEstimator- A p90 that revises as it measures, and says so before it knows.
RateWindows- Rate state, owned by whoever owns the agent.
ReplaySession- Replay accepted commands at their original boundaries. Calls no provider, ever.
ToolRegistryUtilityPolicy- The default floor: score every option, take the highest.
Interfaces
AgentPolicyAgentSessionOptionsAiCommandAiManifest- What an agent is, written down.
AiPreemption- An observation that ended the current intent early.
AiProviderAiProviderCapabilities- What a provider says it can do.
AiProviderConfigAiProviderResultAiRequestAiSessionAiSessionOptionsAiUsage- What an agent has spent.
AssembledContextAssembledContextLike- Context assembled for one request.
AuthoritativeAgentOptionsBudgetLimitsContextProvider- A read-only snapshot the model is told about.
ContextSectionDecisionChannel- How a decision crosses. The consumer's, because packing and transport are theirs.
DeterministicScript- A provider that answers after a stated number of ticks, or never.
DevelopmentManifestEntityContextOptionsExecutionPolicyIntentLocalProviderConfig- A provider running on this device, accepted by running it.
NavigateArgsNavigateResultNavigationAdapter- What the consumer answers, because only the consumer knows what an agent is.
NavigationBridgeOptionsObservationPolicyContextPolicyOptionProxyProviderConfig- A remote provider reached through the consumer's own endpoint.
RealtimeOptions- A realtime transport, emitting the same events every other session emits.
ReplaySourceToolDefinitionWis the consumer's world view, and this package never constrains it.
Functions
admitToolCall- Whether a proposed tool call may be accepted.
applyCommand- Run an accepted command at a tick boundary, revalidating on the way in.
assembleContext- Never send the scene graph.
chargeUsage- Tokens only. Who aborted, and why, is something the session knows and an event does not.
continuationPreamble- Tell the model when its answer will be used.
createAiProvider- Accept a provider only if it declares what was asked of it, and say so either way.
createLocalProvidercreateProxyProvider- Returns a provider directly, where the local path returns a
Promise<AiProviderResult>. createRealtimeSession- Open a realtime session, or refuse before connecting.
createUsagedescribeAgentdescribeForDevelopment- The manifest a coding assistant reads.
entityContext- A context provider that samples a query into a list of references.
entityRef- A stable, printable name for an entity, and the only form a model ever sees.
entityTool- A tool addressing an entity, whose guard checks the generation and not only the index.
hasKnownExtentloopbackDecisionChannel- A channel with both ends in this process, for a test or a single-process host.
navigationBridge- Build the navigation tools for a consumer's world.
nextState- The next state, or
nullwhen the transition is illegal from here. noteAbort- Record an abort, whatever caused it.
notePreemption- Record that the abort being taken is a preemption.
parseEntityRef- Read a reference back, or
nullwhen it is not one. reachableBy- Whether a route exists for this agent right now, without writing one.
requireStructuredOutput- Ask for output shaped like a schema, or refuse before spending anything.
validateArgs- Check a model's arguments against a tool's schema.
validateIntent- Whether an intent is internally coherent.
Types
AdmissionAgentRoleAgentState- Four states, and a second request that has no edge to arrive on.
AgentTransitionAiEventApplyOutcomeIntentCheckLogEntryRealtimeResultStructuredResultToolSchema- What a model may be asked to produce.
ValidationResultWhileBusy
Constants
AI_NETWORK_AUTHORITY- The sentinel's own name, exported so the documentation gate fires when this file arrives.
UNKNOWN_EXTENT- An intent whose extent is unknown. The watermark reads it as "ask immediately".