Skip to main content

Type Alias: NodeExecutorConfig

NodeExecutorConfig = { instructions: string; maxInternalIterations?: number; maxTokens?: number; parallelTools?: boolean; temperature?: number; type: "gmi"; } | { args?: Record<string, unknown>; toolName: string; type: "tool"; } | { extensionId: string; method: string; type: "extension"; } | { autoAccept?: boolean; autoReject?: boolean | string; guardrailOverride?: boolean; judge?: { confidenceThreshold?: number; criteria?: string; model?: string; provider?: string; }; onTimeout?: "accept" | "reject" | "error"; prompt: string; type: "human"; } | { guardrailIds: string[]; onViolation: "block" | "reroute" | "warn" | "sanitize"; rerouteTarget?: string; type: "guardrail"; } | { condition: GraphCondition; type: "router"; } | { graphId: string; inputMapping?: Record<string, string>; outputMapping?: Record<string, string>; type: "subgraph"; } | { type: "voice"; voiceConfig: VoiceNodeConfig; }

Defined in: packages/agentos/src/orchestration/ir/types.ts:162

Describes how the runtime should execute a GraphNode. Each variant maps to a distinct execution strategy.

  • gmi — General Model Invocation: call an LLM with system instructions.
  • tool — Invoke a registered ITool by name, optionally with static args.
  • extension — Call a method on a registered IExtension.
  • human — Suspend execution and surface a prompt to a human operator.
  • guardrail — Run one or more guardrail checks; route or block on violation.
  • router — Pure routing node; evaluates a GraphCondition and emits no output.
  • subgraph — Delegate to another CompiledExecutionGraph with optional field mapping.
  • voice — Run a voice pipeline session with configurable STT/TTS and turn management.

Type Declaration​

{ instructions: string; maxInternalIterations?: number; maxTokens?: number; parallelTools?: boolean; temperature?: number; type: "gmi"; }

instructions​

instructions: string

System-level instructions injected before the user message.

maxInternalIterations?​

optional maxInternalIterations: number

Maximum ReAct loop iterations when executionMode is react_bounded. Defaults to 10.

maxTokens?​

optional maxTokens: number

Hard cap on output tokens for this node's completion.

parallelTools?​

optional parallelTools: boolean

Whether to issue multiple tool calls in a single model turn.

temperature?​

optional temperature: number

Sampling temperature forwarded to the LLM provider.

type​

type: "gmi"

{ args?: Record<string, unknown>; toolName: string; type: "tool"; }

args?​

optional args: Record<string, unknown>

Static arguments merged with runtime-provided arguments before invocation.

toolName​

toolName: string

Registered tool name as it appears in the tool catalogue.

type​

type: "tool"

{ extensionId: string; method: string; type: "extension"; }

extensionId​

extensionId: string

Extension identifier as registered in the capability registry.

method​

method: string

Name of the extension method to call.

type​

type: "extension"

{ autoAccept?: boolean; autoReject?: boolean | string; guardrailOverride?: boolean; judge?: { confidenceThreshold?: number; criteria?: string; model?: string; provider?: string; }; onTimeout?: "accept" | "reject" | "error"; prompt: string; type: "human"; }

autoAccept?​

optional autoAccept: boolean

When true, the node auto-accepts without waiting for human input. Useful for testing.

autoReject?​

optional autoReject: boolean | string

When true or a string, the node auto-rejects. A string value is used as the rejection reason.

guardrailOverride?​

optional guardrailOverride: boolean

Run post-approval guardrails after an automated approval path (autoAccept, judge, or timeout-accept). Defaults to true.

judge?​

optional judge: object

Delegates the approval decision to an LLM judge instead of a human. When the judge's confidence falls below confidenceThreshold, execution falls through to the normal human interrupt.

judge.confidenceThreshold?​

optional confidenceThreshold: number

Confidence threshold (0–1). Below this, fall through to human interrupt. Defaults to 0.7.

judge.criteria?​

optional criteria: string

Custom evaluation criteria/rubric.

judge.model?​

optional model: string

LLM model to use for the judge call. Defaults to 'gpt-4o-mini'.

judge.provider?​

optional provider: string

LLM provider. Defaults to 'openai'.

onTimeout?​

optional onTimeout: "accept" | "reject" | "error"

Behaviour when the node's timeout expires.

  • 'accept' — auto-accept.
  • 'reject' — auto-reject.
  • 'error' — throw a timeout error (default behaviour).

Default​

'error'

prompt​

prompt: string

Message displayed to the human operator while the graph is suspended.

type​

type: "human"

{ guardrailIds: string[]; onViolation: "block" | "reroute" | "warn" | "sanitize"; rerouteTarget?: string; type: "guardrail"; }

guardrailIds​

guardrailIds: string[]

Ordered list of guardrail identifiers to evaluate.

onViolation​

onViolation: "block" | "reroute" | "warn" | "sanitize"

Action taken when any guardrail fires.

rerouteTarget?​

optional rerouteTarget: string

Node id to route to when onViolation is 'reroute'.

type​

type: "guardrail"

{ condition: GraphCondition; type: "router"; }

condition​

condition: GraphCondition

Routing predicate; the returned node-id determines the next edge.

type​

type: "router"

{ graphId: string; inputMapping?: Record<string, string>; outputMapping?: Record<string, string>; type: "subgraph"; }

graphId​

graphId: string

Id of the CompiledExecutionGraph to delegate to.

inputMapping?​

optional inputMapping: Record<string, string>

Maps parent scratch field paths → child input field paths.

outputMapping?​

optional outputMapping: Record<string, string>

Maps child artifacts field paths → parent scratch field paths.

type​

type: "subgraph"

{ type: "voice"; voiceConfig: VoiceNodeConfig; }

type​

type: "voice"

voiceConfig​

voiceConfig: VoiceNodeConfig

Voice pipeline session configuration.