ask<T>

const answer = try ask<Type>("the prompt, with {values}")

type Verdict {
    /// true if the review is positive
    positive: Bool
    /// one short reason
    reason: String
}

agent Reader {
    can Ai, Console

    on review(text: String) ! Error {
        const v = try ask<Verdict>("Is this review positive? {text}")
        print("positive: {v.positive}")
    }
}

fn main() {
    const bot = Reader.start()
    bot.review("Loved it, fast delivery") catch e => print("failed: {e}")
}

positive: true[dry run] ask("Is this review positive? Loved it, fast delivery") -> Verdict(positive: true, reason: " fry ey")

• ask needs the Ai capability, so it is in the agent's can.
• Kavach builds a schema from Verdict and sends it with the prompt. The /// comments describe each field to the model.
• The answer is checked against Verdict. A wrong answer is asked again, twice by default, then fails.

This output comes from kavach run --dry-run: no model is called and nothing costs. ask gets a made-up Verdict, the same on every run, and the last line shows the simulated call. Without --dry-run, the model the code chooses answers: see Choosing the model.

Limits and sizes

Every limit in Kavach, on one page: Limits and Sizes.