@heybeaux/sonder-adapter-awm

Sonder adapter for AWM (Agent Workflow Memory). Fills event.intent on every SonderEvent and scores LeWM predictions against actual step outcomes for calibration.

Install

npm install @heybeaux/sonder-adapter-awm @heybeaux/sonder-core

What it fills

event.intent = {
  action:               string;   // the action/step type being taken
  step_trace_id:        string;   // AWM StepTrace reference
  skipped:              boolean;  // true if AWM recommended skipping (high confidence)
  skip_reason?:         string;   // reason for skip
  constraint_injected:  boolean;  // true if approval gate pre-injected constraints
}

Usage

import { AwmAdapter } from '@heybeaux/sonder-adapter-awm';
import { createRuntime } from '@heybeaux/sonder-sdk';

let currentIntent = null;

const runtime = createRuntime({
  adapters: [
    new AwmAdapter({
      getCurrentIntent: () => currentIntent,

      // Called after events that have both a step_trace_id and a governance contract_id.
      // Use this to record step outcomes and calibrate LeWM predictions.
      onStepOutcome: (traceId, outcome, event) => {
        console.log(`AWM scored trace ${traceId}: ${outcome}`);
        // Record in your frequency model here
      },
    }),
  ],
});

// Set intent before each step:
currentIntent = {
  action:              'draft_post',
  step_trace_id:       'trace:draft-001',
  skipped:             false,
  constraint_injected: true,
};

Config

interface AWMAdapterConfig {
  // Return the current intent snapshot. Return null if none — event.intent will be zeroed.
  getCurrentIntent(): AWMIntentSnapshot | null;

  // Optional. Called after events with both a step_trace_id and governance.contract_id.
  // Use this to score LeWM predictions against actual outcomes.
  onStepOutcome?(
    stepTraceId: string,
    outcome:     'pass' | 'revise' | 'fail',
    event:       SonderEvent,
  ): void;
}

interface AWMIntentSnapshot {
  action:               string;
  step_trace_id:        string;
  skipped:              boolean;
  skip_reason?:         string;
  constraint_injected:  boolean;
}

AWM ↔ LeWM feedback loop

AWM is the calibration layer — it tracks historical step frequencies and scores LeWM's predictions against actual outcomes. When onStepOutcome() fires, record the result in your frequency model to update how much weight to place on LeWM's structural predictions.

LeWM is the hypothesis generator — it maintains Beta distribution beliefs that AWM's calibration informs over time.

Both callbacks are optional — the observe loop only activates when you supply them.

Compliance

event.intent.* surfaces what the agent intended to do, whether it was skipped for efficiency, and whether approval constraints were pre-injected — supporting auditability requirements under EU AI Act Art. 12.

License

MIT