grok-bot-jev

import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    intent: { type: 'choice', instructions: 'What kind of work does this request mainly need?',
      criteria: { chat: 'Short answer or conversation; no tools needed', lookup: 'Known file/API/status check; mostly deterministic', research: 'Need web/search and synthesis', browser: 'Need interactive browser clicks', coding: 'Edit code / tests / repo work', write: 'Draft long prose/email/doc', account: 'Send, publish, pay, delete, change permissions' } },
    reuse_cache: { type: 'boolean', instructions: 'Is there already a fresh enough cached result that should be reused instead of doing new heavy work?' },
    needs_subagent: { type: 'boolean', instructions: 'Does this clearly need an extra specialized bot (research/browser/coding) beyond one Grok Bot turn?' },
    stop_retry: { type: 'boolean', instructions: 'Given prior_error and same_error_count, should we STOP retrying the same approach?' },
    complexity: { type: 'score', instructions: 'How much Grok Bot effort is justified?',
      criteria: ['Trivial — one step or cached', 'Normal — short tool use', 'Heavy — multi-step research/browser/coding'] },
  },
});