/** * Reliability execution step — a Jensen-owned decode/validate/execute/record * pipeline over raw model output. * * This is the authority boundary in code form: raw output is decoded into an * AgentAction, validated against the tool registry/boundary/policy, and only * then executed. Evidence is recorded by Jensen from the observed outcome. * * The deterministic Reliability Suite drives the kernel through this function * with a scripted model and tool harness. The LIVE agent loop integrates with * the same decoder/validator/evidence/gate primitives through * {@link ReliabilitySessionBridge} (wired by ReliabilitySessionController as * Agent beforeToolCall/afterToolCall/onTurnEnd hooks) rather than through this * function, so the live loop is not duplicated and adversarial behavior cannot * bypass the kernel by taking a different code path. */ import { type ActionValidatorContext } from "./action-validator.js"; import type { MissionRuntime } from "./mission-runtime.js"; import type { ActionValidationFailure, AgentAction, MissionEvidence, ToolCallAction } from "./types.js"; export interface ReliabilityExecutionContext extends ActionValidatorContext { /** Execute a validated tool action and return the observed outcome. */ executeTool(action: ToolCallAction): Promise; } export type ReliabilityStepResult = { kind: "executed"; action: ToolCallAction; evidence: MissionEvidence; } | { kind: "validation_rejected"; failure: ActionValidationFailure; } | { kind: "decode_rejected"; failure: ActionValidationFailure; } | { kind: "non_tool"; action: Exclude; }; /** * Decode, validate, and (if valid) execute a single raw model output as a * Jensen-owned action. Never executes an invalid action. Never throws for * invalid model output — failures are returned as structured results. */ export declare function executeAgentAction(runtime: MissionRuntime, raw: unknown, ctx: ReliabilityExecutionContext): Promise; //# sourceMappingURL=execution.d.ts.map