{
  "skill_name": "cursor-sdk",
  "evals": [
    {
      "id": 1,
      "name": "agent-create-vs-prompt",
      "prompt": "Apply the cursor-sdk skill: a user wants to run a one-off code task. Should they use Agent.create or Agent.prompt? Show the correct call.",
      "expected_output": "Output distinguishes Agent.prompt (one-off, fire-and-forget) from Agent.create (stateful/resumable) and shows the correct one-off call.",
      "assertions": [
        "The output recommends Agent.prompt for a one-off task",
        "The code sample matches the current Cursor SDK signature",
        "The output does not confuse local and cloud runtime defaults"
      ]
    },
    {
      "id": 2,
      "name": "resume-after-interrupt",
      "prompt": "Use the cursor-sdk skill to resume an agent session after an interrupt, preserving prior directives.",
      "expected_output": "Output uses Agent.resume with the prior session id, and warns that interrupt-chained resumes silently drop directives so a fresh consolidated-scope subagent is safer.",
      "assertions": [
        "The output references Agent.resume and the session id",
        "The output warns about directives being dropped on interrupt-chain resume",
        "The output recommends firing a fresh subagent with consolidated scope"
      ]
    },
    {
      "id": 3,
      "name": "capability-gating",
      "prompt": "Apply the cursor-sdk skill to write code that only invokes a reviewer when the current model supports it.",
      "expected_output": "Output gates reviewer invocation behind run.supports(...) capability checking rather than assuming the capability exists.",
      "assertions": [
        "The output uses run.supports(...) or equivalent capability gating",
        "The output does not hard-assume skipReviewerRequest exists on every model",
        "The code degrades gracefully when the capability is absent"
      ]
    }
  ]
}