AgentLoopInput

Kind: interface

Signature

export interface AgentLoopInput

Properties

• goal (string): Objective, rendered as the first user message.
• priorMessages? (AgentLoopPriorMessage[]): Bounded chat history preceding the current user message.
• latestUserMessage? (string): Current chat message, rendered separately and never truncated or dropped.
• system? (string): Optional base system prompt for the loop.
• providerId (string): Provider identifier resolved by the LLM proxy.
• model (string): Model identifier used for loop turns.
• tools? (string[]): Model-visible tool names requested for the loop. Omit to expose the full signed manifest surface. Names outside the resolved surface are denied.
• maxTurns? (number): Cooperative turn cap. The server-side manifest cap still binds.
• tokenBudget? (number): Cooperative token budget. The server-side manifest budget still binds.
• maxWallClockMs? (number): Cooperative wall-clock budget in milliseconds, measured with checkpointed kernel.now() reads so replay stays deterministic. The runtime's subprocess walltime and server-side backstop still bind.
• temperature? (number): Optional sampling temperature.
• maxTokens? (number): Optional per-turn completion token ceiling. Omitted loops conservatively use 128,000.
• maxInputTokensPerTurn? (number): Optional per-turn provider-input ceiling. The transcript is deterministically bounded before each action; omitted loops use 128,000.
• reasoningEffort? (AgentLoopReasoningEffort): Optional reasoning effort hint.
• toolChoice? (AgentLoopToolChoice): Optional initial tool selection directive.