Function Dispatch From a Plain-Language Command

{
  "model": "jev-latest",
  "state": {
    "command": "{{user-command}}"
  },
  "questions": {
    "__tool__": {
      "type": "choice",
      "instructions": "What is the user asking the assistant to do?",
      "criteria": {
        "{{function-a}}": {"what": "{{function-a-purpose}}", "examples": ["{{function-a-example}}"]},
        "{{function-b}}": {"what": "{{function-b-purpose}}", "examples": ["{{function-b-example}}"]},
        "none": {"what": "Not a request for any listed action"}
      }
    },
    "{{function-a}}__{{enum-argument}}": {
      "type": "choice",
      "instructions": "{{enum-argument-question}}",
      "criteria": {
        "{{enum-value-1}}": "{{enum-value-1-meaning}}",
        "{{enum-value-2}}": "{{enum-value-2-meaning}}",
        "not_stated": "The user did not say, use the default"
      }
    },
    "{{function-a}}__{{flag-argument}}": {
      "type": "noul",
      "instructions": "{{flag-argument-question}}"
    }
  }
}

How to build it: for every function, add one question per argument that takes a value from a fixed list, a set of values, or a boolean. Phrase each argument question as a meaning, never as a parameter name; "Which resolution?" gives the model nothing to match, "Does the user want the full-size file rather than a preview?" does. Include a not_stated option so optional arguments fall back to defaults. Dispatch in code: pick __tool__.choice, read only that function's argument answers, and treat the lowest confidence among them as the confidence of the whole call. Free-text arguments, like a message body, are not for this model; take them from the command string in code or ask an LLM.