// REFUND AGENT — an LLM decides, the language decides what it may do.
//
// The model proposes approve / deny / escalate as JSON; its output becomes a
// sum type (malformed → Escalate, never a crash). The hard rules are not the
// model's to bend: a refund above the order amount or after 30 days is
// refused by guards on the edge to `paid`; the daily payout cap is a memory
// invariant; a refund above 20,000 cents needs a human (`approve()`), which
// tests script with `mock approve`. Proven: paid is unreachable without
// decided; nothing after denied. Proven too: the cost bound, because every
// think() has a literal max_tokens and one provider round (max_rounds).
//
//   soma check  refund_agent.cell      → cost: 'tokens' bound proven
//   soma verify refund_agent.cell
//   soma test   refund_agent.cell      → offline: think() and approve() are scripted

cell type Decision {
    variants {
        Approve
        Deny { reason: String }
        Escalate { reason: String }
    }
}

cell agent RefundAgent {
    face {
        signal seed_order(order_id: String, amount: Int, days_since_delivery: Int) -> Int
        signal decide(request_id: String, order_id: String, amount: Int, text: String) -> Map
        signal status(request_id: String) -> String
        signal paid_today() -> Int
        tool lookup_order(order_id: String) -> Map "Order amount and days since delivery"
    }
    cost {
        tokens: 300
    }
    memory {
        orders: Map<String, Map> [persistent]
        paid_by_day: Map<String, Int> [persistent]
        invariant paid_by_day >= 0 && paid_by_day <= 100000   // daily cap, whatever the model says
    }
    state request {
        initial: received
        received -> decided
        decided -> paid { guard { amount <= order.amount && order.days <= 30 } }
        decided -> denied
        decided -> escalated
        * -> failed except [paid, denied, escalated]
    }

    on seed_order(order_id: String, amount: Int, days_since_delivery: Int) {
        orders.set(order_id, map("amount", amount, "days", days_since_delivery))
        return amount
    }
    on lookup_order(order_id: String) {
        let o = orders.get(order_id)
        if o == () { fail("not_found", "order {order_id}") }
        return o
    }
    on paid_today() { return paid_by_day.get(today()) ?? 0 }
    on status(request_id: String) { return get_status(request_id) }

    on decide(request_id: String, order_id: String, amount: Int, text: String) {
        let order = lookup_order(order_id)
        set_budget(2000)
        let r = try {
            think(
                "Refund request for order {order_id}: {text}. Amount asked: {amount}. Reply with JSON {\"decision\": \"approve\"|\"deny\"|\"escalate\", \"reason\": \"...\"}",
                "You are a support agent. Be strict.",
                map("max_tokens", 300, "max_rounds", 1)
            )
        }
        if r.error != () {
            transition(request_id, "failed")
            return map("state", "failed", "reason", r.detail)
        }
        transition(request_id, "decided")
        let decision = _parse(r.value)
        return match decision {
            Approve -> _pay(request_id, order_id, amount, order)
            Deny { reason } -> {
                transition(request_id, "denied")
                map("state", "denied", "reason", reason)
            }
            Escalate { reason } -> {
                transition(request_id, "escalated")
                map("state", "escalated", "reason", reason)
            }
        }
    }

    // Approval says approve; the rules decide. The guard on decided -> paid
    // reads `amount` and `order` from this handler; a refused guard is
    // caught, rolled back, and the case is escalated instead.
    on _pay(request_id: String, order_id: String, amount: Int, order: Map) {
        if amount > 20000 && !approve("refund {amount} on {order_id}") {
            transition(request_id, "escalated")
            return map("state", "escalated", "reason", "human declined")
        }
        let p = try {
            transition(request_id, "paid")
            paid_by_day.set(today(), paid_today() + amount)      // the daily cap lives here
        }
        if p.error != () {
            transition(request_id, "escalated")
            return map("state", "escalated", "reason", p.detail)
        }
        return map("state", "paid", "amount", amount)
    }

    on _parse(raw: String) {
        let j = try { from_json(raw) }
        if j.error != () || type_of(j.value) != "Map" { return Escalate { reason: "unreadable decision" } }
        let d = j.value
        return match d.decision ?? "" {
            "approve" -> Approve
            "deny" -> Deny { reason: d.reason ?? "" }
            _ -> Escalate { reason: d.reason ?? "unknown decision" }
        }
    }
}

cell test RefundTests {
    rules {
        assert seed_order("o1", 5000, 3) == 5000
        assert seed_order("o2", 50000, 3) == 50000
        assert seed_order("o3", 5000, 45) == 5000
        // the model approves, deny and escalate — scripted per test
        mock think """{"decision": "approve", "reason": "ok"}"""
        assert decide("r1", "o1", 4000, "broken").state == "paid"
        assert status("r1") == "paid"
        mock think """{"decision": "deny", "reason": "used"}"""
        assert decide("r2", "o1", 4000, "meh").state == "denied"
        mock think "garbage from the model"
        assert decide("r3", "o1", 10, "?").state == "escalated"
        // the model says approve for MORE than the order: the guard refuses, the case escalates
        mock think """{"decision": "approve", "reason": "generous"}"""
        assert decide("r4", "o1", 9000, "want more").state == "escalated"
        assert status("r4") == "escalated"
        assert paid_today() == 4000
        // 30-day rule
        mock think """{"decision": "approve"}"""
        assert decide("r5", "o3", 100, "late").state == "escalated"
        // human gate above 20,000: declined, then accepted
        mock think """{"decision": "approve"}"""
        mock approve false
        assert decide("r6", "o2", 30000, "big").state == "escalated"
        mock think """{"decision": "approve"}"""
        mock approve true
        assert decide("r7", "o2", 30000, "big").state == "paid"
        assert paid_today() == 34000
        // the daily cap: 34000 + 70000 > 100000 → escalated, cap intact
        assert seed_order("o4", 100000, 1) == 100000
        mock think """{"decision": "approve"}"""
        mock approve true
        assert decide("r8", "o4", 70000, "huge").state == "escalated"
        assert paid_today() == 34000
        // an LLM failure is a failed request, not a crash
        mock think error "timeout"
        assert decide("r9", "o1", 10, "x").state == "failed"
        assert status("r9") == "failed"
    }
}