{
  "note": "Recorded planner responses. Keys are purpose plus the exact prompt text.",
  "entries": [
    {
      "purpose": "plan",
      "prompt": "You plan research for a cited report.\nSplit the question into at most 4 facets.\nReply with JSON only: {\"facets\": [{\"sub_question\": \"...\", \"keywords\": [\"...\"]}]}\nQuestion: Are command denylists enough to sandbox an agent?",
      "response": "{\"facets\": [{\"label\": \"What denylists catch\", \"sub_question\": \"What do command denylists block?\", \"keywords\": [\"denylist\", \"command\", \"blocked\"]}, {\"label\": \"How they fail\", \"sub_question\": \"How can an agent get around a denylist?\", \"keywords\": [\"interpreter\", \"script\", \"bypass\"]}, {\"label\": \"What to add\", \"sub_question\": \"What isolation should sit under a denylist?\", \"keywords\": [\"isolation\", \"least\", \"privilege\"]}]}"
    },
    {
      "purpose": "plan",
      "prompt": "You plan research for a cited report.\nSplit the question into at most 4 facets.\nReply with JSON only: {\"facets\": [{\"sub_question\": \"...\", \"keywords\": [\"...\"]}]}\nQuestion: What is the stateful sandbox controller for?",
      "response": "Sure! Here are some facets: 1. what it is 2. why it matters"
    }
  ]
}