@rdk-moss/teaching

Teach-while-solve annotation layer for the Moss agent runtime.

Dependency surface: this package depends on @rdk-moss/agent only through its two stable public subpaths — @rdk-moss/agent/core (types) and @rdk-moss/agent/safety (sanitizeSecrets). It never imports agent internals, so agent-internal refactors cannot break the teaching layer.

Features

• Pre-execution Annotations: Explain tool intent before execution
• Post-execution Reflection: Verify results and suggest next steps
• Dry-run Summaries: Preview risky operations before execution
• Confidence Assessment: Evaluate operation success and rollback options
• Failure Cards: Structured error analysis with actionable recovery steps

Installation

npm install @rdk-moss/teaching

Usage

import { createTeachingHooks } from '@rdk-moss/teaching';

const { onBeforeToolExec, onToolResult } = createTeachingHooks({
  depth: 'concise', // 'off' | 'concise' | 'detailed'
  llmProvider: provider,
  modelId: 'claude-sonnet-4-20250514',
  deviceLabel: 'RDK X5',
  familyIsRdk: true,
  runId: 'run-123',
  sessionKey: 'session-456',
  teachingConfirmRequested: false,
  teachingConfirmInteractive: false,
  emitTeachingMeta: (meta) => console.log('Teaching meta:', meta),
  waitTeachingConfirm: async () => true,
  classifyPlanMutation: (toolName) => toolName.startsWith('device_exec'),
});

// Register with agent
agent.registerHook('onBeforeToolExec', onBeforeToolExec);
agent.registerHook('onToolResult', onToolResult);

API

• TeachingLayer: Main teaching annotation layer
• TeachingAnnotationCollector: Collect and format annotations
• TeachingToolDigest: Generate tool execution digests

Teaching Depth Levels

• off: No teaching annotations
• concise: Brief explanations for mutations only
• detailed: Comprehensive explanations for all operations

License

MIT