curl --request POST \
  --url https://api.copass.id/api/v1/query/sandboxes/{sandbox_id}/discover \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data @- <<EOF
{
  "history": [
    {
      "content": "I'm evaluating Copass for our team.",
      "role": "user"
    },
    {
      "content": "Happy to help — what are you most curious about?",
      "role": "assistant"
    }
  ],
  "query": "What is the Copass Knowledge Graph?",
  "reference_date": "2026-04-16"
}
EOF

{
  "count": 1,
  "header": "## Discover: Relevant Context\n\nCandidate context from the user's knowledge graph that isn't in your current window. The items below are labels, not content — treat them as an index you must open before citing them.\n\n**How to use this response:**\n1. Scan for items that look relevant to the user's current question.\n2. Before answering from any item, call `interpret` with its `canonical_ids` tuple to read the actual content. A title alone is not evidence — do not describe or rely on an item without interpreting it first. Batch multiple items into a single `interpret` call.\n3. Use `search` as a fallback when nothing in the menu looks obviously relevant but the question still implies the graph holds the answer.\n4. Skip this menu entirely when the user's question is already answerable from conversation or code in your context.",
  "items": [
    {
      "canonical_ids": [
        "454494b2-6753-50b6-b659-6e4cf6a13391",
        "9c4c60c9-c404-5b9e-a69c-345593a44631"
      ],
      "id": "item_7c3f",
      "score": 0.87,
      "summary": "Olane CLI queries the Copass knowledge graph."
    }
  ],
  "next_steps": "If any item above plausibly relates to the user's question, call `interpret` on those canonical_ids before responding — titles are not evidence. `interpret` accepts multiple tuples in one call, so batch them: `interpret(items=[[<canonical_ids from item 1>], [<canonical_ids from item 2>], ...])`. Fall back to `search` only when no item looks relevant but the question still seems to need graph context. Skip both when the conversation already has what you need.",
  "query": "how does the CLI query the knowledge graph?",
  "sandbox_id": "sbx_demo"
}