@alexkroman1/aai-cli 7.0.0 → 8.1.0

package/dist/templates/retail/tools/modify_pending_order_payment.ts

@@ -1,22 +1,20 @@
 import { isToolFailure } from "@alexkroman1/aai";
 import { z } from "zod";
-import { OrderIdField, resolveOrder } from "../resolve.ts";
-import {
-  authenticatedUser,
-  findPaymentMethod,
-  isGiftCard,
-  money,
-  retailTool,
-  setFocus,
-} from "../store.ts";
+import { planPayment } from "../payment.ts";
+import { stageAction } from "../pending.ts";
+import { OrderIdField } from "../resolve.ts";
+import { retailTool, setFocus } from "../store.ts";
 
 export default retailTool({
   name: "modify_pending_order_payment",
   when: "serving",
+  send: { type: "STAGED" },
   description:
-    "Change which payment method a pending order is charged to. The new method must be different " +
-    "from the current one, and a gift card must hold enough to cover the whole order. The original " +
-    "method is refunded. Read the change back and get an explicit yes before calling this.",
+    "STAGE a change to which payment method a pending order is charged to — this does NOT change " +
+    "it. The new method must be different from the current one, and a gift card must hold enough " +
+    "to cover the whole order; the original method is refunded. The change comes back as a " +
+    "sentence to read to the caller; nothing happens until you hear an explicit yes and call " +
+    "confirm_change.",
   inputSchema: z.object({
     order_id: OrderIdField,
     payment_method_id: z
@@ -25,70 +23,10 @@ export default retailTool({
       .describe("The new payment method id, e.g. 'gift_card_0000000'"),
   }),
   execute: (args, state) => {
-    const user = authenticatedUser(state);
-    if (isToolFailure(user)) return user;
-
-    const order = resolveOrder(state, args.order_id);
-    if (isToolFailure(order)) return order;
-    setFocus(state, { orderId: order.order_id });
-
-    if (!order.status.startsWith("pending")) {
-      return {
-        error: `Order ${order.order_id} is ${order.status}, and only a pending order's payment method can be changed.`,
-      };
-    }
-
-    const newMethod = findPaymentMethod(user, args.payment_method_id);
-    if (isToolFailure(newMethod)) return newMethod;
-
-    const original = order.payment_history[0];
-    if (order.payment_history.length !== 1 || original?.transaction_type !== "payment") {
-      return {
-        error: `Order ${order.order_id} does not have exactly one payment on record, so its payment method cannot be changed.`,
-      };
-    }
-    if (original.payment_method_id === args.payment_method_id) {
-      return {
-        error: `Order ${order.order_id} is already paid with ${args.payment_method_id}. The new method must be different.`,
-      };
-    }
-
-    const amount = original.amount;
-    if (isGiftCard(newMethod) && newMethod.balance < amount) {
-      return {
-        error: `Gift card ${args.payment_method_id}'s balance ($${newMethod.balance.toFixed(2)}) does not cover the $${amount.toFixed(2)} order total.`,
-      };
-    }
-
-    order.payment_history.push(
-      {
-        transaction_type: "payment",
-        amount,
-        payment_method_id: args.payment_method_id,
-      },
-      {
-        transaction_type: "refund",
-        amount,
-        payment_method_id: original.payment_method_id,
-      },
-    );
-
-    if (isGiftCard(newMethod)) {
-      newMethod.balance = money(newMethod.balance - amount);
-    }
-    const oldMethod = user.payment_methods[original.payment_method_id];
-    if (oldMethod && isGiftCard(oldMethod)) {
-      oldMethod.balance = money(oldMethod.balance + amount);
-    }
-
-    return {
-      order_id: order.order_id,
-      status: order.status,
-      amount,
-      paid_with: args.payment_method_id,
-      refunded_to: original.payment_method_id,
-      message: `Order ${order.order_id} is now charged to ${args.payment_method_id}, and $${amount.toFixed(2)} is going back to ${original.payment_method_id}.`,
-    };
+    const plan = planPayment(state, args.order_id, args.payment_method_id);
+    if (isToolFailure(plan)) return plan;
+    setFocus(state, { orderId: plan.orderId });
+    return stageAction(state, { kind: "modify_pending_order_payment", plan });
   },
-  summary: (_args, result) => `re-paid ${result.order_id}`,
+  summary: (_args, result) => `staged: ${result.read_back}`,
 });

package/dist/templates/retail/tools/modify_user_address.ts

@@ -1,29 +1,26 @@
 import { isToolFailure } from "@alexkroman1/aai";
 import { z } from "zod";
-import { AddressFields, formatAddress, toAddress } from "../address.ts";
-import { requireOwnUser, retailTool } from "../store.ts";
+import { AddressFields, planUserAddress } from "../address.ts";
+import { stageAction } from "../pending.ts";
+import { retailTool } from "../store.ts";
 
 export default retailTool({
   name: "modify_user_address",
   when: "serving",
+  send: { type: "STAGED" },
   description:
-    "Change the customer's default address, used for future orders. Read the new address back and " +
-    "get an explicit yes before calling this. This does not change the address on any existing " +
-    "order — use modify_pending_order_address for that.",
+    "STAGE a change to the customer's default address, used for future orders — this does NOT " +
+    "change it. The new address comes back as a sentence to read to the caller; nothing happens " +
+    "until you hear an explicit yes and call confirm_change. This does not change the address on " +
+    "any existing order — use modify_pending_order_address for that.",
   inputSchema: z.object({
     user_id: z.string().max(100).describe("The user id, e.g. 'sara_doe_496'"),
     ...AddressFields,
   }),
   execute: (args, state) => {
-    const user = requireOwnUser(state, args.user_id);
-    if (isToolFailure(user)) return user;
-
-    user.address = toAddress(args);
-    return {
-      user_id: user.user_id,
-      address: user.address,
-      message: `Default address updated to ${formatAddress(user.address)}. Existing orders keep their own shipping addresses.`,
-    };
+    const plan = planUserAddress(state, args.user_id, args);
+    if (isToolFailure(plan)) return plan;
+    return stageAction(state, { kind: "modify_user_address", plan });
   },
-  summary: (_args, result) => `re-addressed ${result.user_id}`,
+  summary: (_args, result) => `staged: ${result.read_back}`,
 });

package/dist/templates/retail/tools/return_delivered_order_items.ts

@@ -1,23 +1,21 @@
 import { isToolFailure } from "@alexkroman1/aai";
 import { z } from "zod";
-import { OrderIdField, resolveOrder } from "../resolve.ts";
-import {
-  authenticatedUser,
-  findPaymentMethod,
-  isGiftCard,
-  retailTool,
-  setFocus,
-} from "../store.ts";
+import { stageAction } from "../pending.ts";
+import { OrderIdField } from "../resolve.ts";
+import { planReturn } from "../returns.ts";
+import { retailTool, setFocus } from "../store.ts";
 
 export default retailTool({
   name: "return_delivered_order_items",
   when: "serving",
+  send: { type: "STAGED" },
   description:
-    "Request a return of items from a delivered order. Only a 'delivered' order can be returned, " +
-    "and only once. The refund must go to the order's ORIGINAL payment method or to one of the " +
-    "customer's gift cards. Confirm the exact item list and the refund destination with an " +
-    "explicit yes before calling this. The customer gets an email explaining how to send the " +
-    "items back.",
+    "STAGE a return of items from a delivered order — this does NOT request anything. Only a " +
+    "'delivered' order can be returned, and only once. The refund must go to the order's ORIGINAL " +
+    "payment method or to one of the customer's gift cards. The exact item list and the refund " +
+    "destination come back as a sentence to read to the caller; nothing happens until you hear an " +
+    "explicit yes and call confirm_change. The customer then gets an email explaining how to send " +
+    "the items back.",
   inputSchema: z.object({
     order_id: OrderIdField,
     item_ids: z
@@ -30,63 +28,10 @@ export default retailTool({
       .describe("Where the refund goes — the original method, or one of their gift cards"),
   }),
   execute: (args, state) => {
-    const user = authenticatedUser(state);
-    if (isToolFailure(user)) return user;
-
-    const order = resolveOrder(state, args.order_id);
-    if (isToolFailure(order)) return order;
-    setFocus(state, { orderId: order.order_id });
-
-    if (order.status !== "delivered") {
-      return {
-        error: `Order ${order.order_id} is ${order.status}. Only a delivered order can be returned, and only once.`,
-      };
-    }
-
-    const method = findPaymentMethod(user, args.payment_method_id);
-    if (isToolFailure(method)) return method;
-
-    const originalMethodId = order.payment_history[0]?.payment_method_id;
-    if (!isGiftCard(method) && args.payment_method_id !== originalMethodId) {
-      return {
-        error: `A refund must go to the original payment method (${originalMethodId}) or to a gift card. ${args.payment_method_id} is neither.`,
-      };
-    }
-
-    if (args.item_ids.length === 0) {
-      return { error: "No items were listed to return." };
-    }
-    const held = new Map<string, number>();
-    for (const item of order.items) {
-      held.set(item.item_id, (held.get(item.item_id) ?? 0) + 1);
-    }
-    const asked = new Map<string, number>();
-    for (const itemId of args.item_ids) {
-      asked.set(itemId, (asked.get(itemId) ?? 0) + 1);
-    }
-    for (const [itemId, count] of asked) {
-      const available = held.get(itemId) ?? 0;
-      if (count > available) {
-        return {
-          error: `Order ${order.order_id} holds ${available} of item ${itemId}, but ${count} were listed for return.`,
-        };
-      }
-    }
-
-    order.status = "return requested";
-    order.return_items = [...args.item_ids].sort();
-    order.return_payment_method_id = args.payment_method_id;
-
-    const names = order.return_items
-      .map((id) => order.items.find((item) => item.item_id === id)?.name ?? id)
-      .join(", ");
-    return {
-      order_id: order.order_id,
-      status: order.status,
-      return_items: order.return_items,
-      refund_to: args.payment_method_id,
-      message: `Return requested on ${order.order_id} for ${names}. The customer will get an email with return instructions, and the refund goes to ${args.payment_method_id} once the items arrive.`,
-    };
+    const plan = planReturn(state, args.order_id, args.item_ids, args.payment_method_id);
+    if (isToolFailure(plan)) return plan;
+    setFocus(state, { orderId: plan.orderId });
+    return stageAction(state, { kind: "return_delivered_order_items", plan });
   },
-  summary: (_args, result) => `return requested on ${result.order_id}`,
+  summary: (_args, result) => `staged: ${result.read_back}`,
 });

package/dist/templates/retail/tools/transfer_to_human_agents.ts

@@ -5,8 +5,10 @@ export default retailTool({
   name: "transfer_to_human_agents",
   description:
     "Hand the caller to a human agent. Use this ONLY when the caller explicitly asks for a human, " +
-    "or when their request cannot be handled with the other tools and the policy. Call this FIRST, " +
-    "then say 'You are being transferred to a human agent. Please hold on.' and nothing else.",
+    "or when their request cannot be handled with the other tools and the policy. The caller is " +
+    "transferred by THIS CALL and by nothing else: call it before you say anything about a " +
+    "transfer, and never say 'You are being transferred' until it has answered. Once it has, say " +
+    "'You are being transferred to a human agent. Please hold on.' and nothing else.",
   inputSchema: z.object({
     summary: z.string().max(2000).describe("A short summary of the caller's issue for the human"),
   }),

package/dist/templates/simple/agent.eval.test.ts

@@ -0,0 +1,54 @@
+// An EVAL: does the agent actually behave? Run it with `aai eval`.
+//
+// A test asserts about the config (see agent.test.ts — it never calls a model).
+// An eval drives the real thing: a real session, the real tool executor, the
+// real event stream, with only the microphone and the speaker faked.
+//
+// `describeEval` picks the model for you and says which it picked:
+//
+//   * with a provider key — a LIVE model. This spends tokens, takes a few
+//     seconds a case, and is a NOISY instrument: a model is probabilistic, so
+//     one failure is a question, not a verdict. Re-run before believing either
+//     answer.
+//   * without one — a SCRIPTED model answering each case's `stubReply`. The
+//     agent, the session and this file all really run, so what it proves is
+//     that the wiring works. It proves nothing about what the agent SAYS.
+//
+// What no eval here can see: anything below the audio boundary — where the
+// agent decides you stopped talking, how it handles being interrupted, whether
+// two sentences merged into one turn. Those need real paced audio.
+import { describeEval } from "@alexkroman1/aai-runtime/eval/vitest";
+import { expect } from "vitest";
+import agentDef from "./agent.ts";
+
+describeEval(agentDef, (test) => {
+  test(
+    "answers a question in its own voice",
+    async ({ session }) => {
+      // `say()` hands back THAT turn, so the claim is about the reply to this
+      // question — not about everything said so far, which already includes
+      // the agent's greeting.
+      const turn = await session.say("What is the capital of France?");
+
+      expect(turn.completed).toBe(true);
+      expect(turn.text).toMatch(/paris/i);
+      // This agent has no tools, so reaching for one would be a real finding.
+      expect(turn.toolCalls).toEqual([]);
+    },
+    { stubReply: "Paris is the capital of France." },
+  );
+
+  test(
+    "keeps the thread across two turns",
+    async ({ session }) => {
+      await session.say("My name is Sam.");
+      const turn = await session.say("What did I say my name was?");
+
+      expect(turn.text).toMatch(/sam/i);
+      expect(session.events().some((e) => e.type === "error.reported")).toBe(false);
+    },
+    // One reply per turn: the second is the one under test, and a stub that
+    // answered the first would fail the case it is supposed to let run.
+    { stubReply: ["Nice to meet you, Sam.", "You said your name was Sam."] },
+  );
+});

package/dist/templates/solo-rpg/agent.eval.test.ts

@@ -0,0 +1,320 @@
+/// <reference types="vite/client" />
+
+// An EVAL: does the story machine actually hold? Run it with `aai eval`.
+//
+// `agent.test.ts` drives each tool directly against a context it made itself,
+// which is where the dice arithmetic and the consequence tables belong. What it
+// cannot reach is the thing this template really is: a campaign and a POSITION
+// that both have to survive a turn boundary, in one session, with a model
+// deciding what to call. Every case here is one of those:
+//
+//   * the campaign and the flow both outlive the turn that created them,
+//   * a standing roll is spent when the scene moves on,
+//   * and an emptied pair of tracks really ends the story — the `after` hook
+//     writes `gameOver`, the tool reports it in the same call, and the final
+//     state refuses everything afterwards.
+//
+// Two modes, announced by `describeEval` on every run:
+//
+//   * with ASSEMBLYAI_API_KEY — a LIVE model, which has to choose these tools
+//     for itself from what the player said.
+//   * without one — a SCRIPTED model whose tool calls REALLY EXECUTE, so the
+//     campaign really changes and the flow really moves. That is a genuine
+//     multi-turn state test with no model in it.
+//
+// What no eval here can see: anything below the audio boundary — endpointing,
+// barge-in, a sentence split across two turns. Those need real paced audio.
+
+import { withSystemPrompt } from "@alexkroman1/aai/manifest";
+import { withDiscoveredTools } from "@alexkroman1/aai/testing";
+import { type EvalTurn, toolResultIn } from "@alexkroman1/aai-runtime/eval";
+import { describeEval } from "@alexkroman1/aai-runtime/eval/vitest";
+import { expect } from "vitest";
+import { z } from "zod";
+import authoredAgent from "./agent.ts";
+import systemPrompt from "./system-prompt.md?raw";
+
+/**
+ * The def a DEPLOYED agent runs, assembled the way the build assembles it: the
+ * authored export, plus what `tools/` declares, plus `system-prompt.md`.
+ *
+ * The prompt is not optional here the way it is in a config test. It is the only
+ * thing that tells a live model to set the whole game up in ONE call, and an
+ * eval run against the framework default prompt would measure an agent nobody
+ * deployed.
+ */
+const agentDef = withSystemPrompt(
+  withDiscoveredTools(authoredAgent, import.meta.glob("./tools/*.ts", { eager: true })),
+  systemPrompt,
+);
+
+/**
+ * What each tool this file drives answers, off the wire.
+ *
+ * `tool.completed` carries a tool result as a JSON STRING, so a case either
+ * casts it or validates it — and a cast is silent exactly when a tool's result
+ * changed shape underneath the eval, which is the regression an eval exists to
+ * catch. `toolResultIn` takes a schema for that reason, and these name only the
+ * fields asserted below: a gated tool's `{ state, done }` position, plus the
+ * `result` it nests its own answer under.
+ */
+const Setup = z.object({
+  state: z.string(),
+  initialized: z.boolean(),
+  playerName: z.string(),
+});
+const Status = z.object({
+  state: z.string(),
+  done: z.boolean(),
+  gameOver: z.boolean(),
+  playerName: z.string(),
+});
+const Roll = z.object({
+  state: z.string(),
+  result: z.object({ actionDice: z.array(z.number()), challengeDice: z.array(z.number()) }),
+});
+const Settled = z.object({
+  state: z.string(),
+  done: z.boolean(),
+  result: z.object({ gameOver: z.boolean() }),
+});
+const Refusal = z.object({ error: z.string() });
+
+/**
+ * What the ONE call to `name` answered on this turn.
+ *
+ * `toolResultIn` is the SDK's reader: it throws rather than returning
+ * undefined, and names what the agent called instead — "it called something
+ * else" is the finding, and a case that read `undefined` off a missing call
+ * would assert against nothing. A turn that called it TWICE is refused too,
+ * rather than silently answered with the first.
+ */
+const answerOf = <T>(turn: EvalTurn, name: string, schema: z.ZodType<T>): T =>
+  toolResultIn(turn.toolCalls, name, schema);
+
+/**
+ * What EVERY call to `name` answered on this turn, in call order.
+ *
+ * The plural form is for a claim about calls that may legitimately not have
+ * happened — a refusal the narrator may have pre-empted by not calling at all.
+ * `toolResultIn` over a ONE-CALL list per call: the name is that call's own, so
+ * the reader's "no such call" and "two calls" throws are unreachable and what
+ * is left is the parse, the schema, and the "never completed" failure.
+ */
+const answersOf = <T>(turn: EvalTurn, name: string, schema: z.ZodType<T>): T[] =>
+  turn.toolCalls
+    .filter((c) => c.name === name)
+    .map((call) => toolResultIn([call], call.name, schema));
+
+/**
+ * Every field `setup_character` requires, for the SCRIPTED runs.
+ *
+ * A live model generates these itself from one sentence — that is the
+ * template's one-turn setup — so this exists only so a keyless run reaches
+ * `playing` and the later turns have a campaign to be about.
+ */
+const SERA = {
+  genre: "dark_fantasy",
+  tone: "melancholic",
+  archetype: "outsider_loner",
+  playerName: "Sera",
+  characterConcept: "A burned-out lamplighter who keeps the last road lit",
+  settingDescription:
+    "The Ashen Hollow, a valley of cold chimneys where the lamps have been going out one by one.",
+  startingLocation: "The Lamplighter's Shed",
+  locationDesc: "A shed of oil cans and broken wicks at the edge of the hollow.",
+  timeOfDay: "late_evening",
+  openingSituation: "The last lamp on the north road has gone dark, and something is moving on it.",
+  npc1Name: "Old Ivo",
+  npc1Desc: "A wick-cutter who remembers when the hollow was lit end to end",
+  npc1Disposition: "neutral",
+  npc1Agenda: "Keep the shed's oil ledger balanced, whatever it costs",
+  threatClockName: "The Dark Road",
+  threatClockDesc: "The hollow goes fully dark and the road is lost",
+} as const;
+
+describeEval(agentDef, (test) => {
+  test(
+    "the character and the story's position both outlive the setup turn",
+    async ({ session }) => {
+      const opening = await session.say(
+        "Dark fantasy, melancholic. My character is Sera, a burned-out lamplighter. Begin.",
+      );
+      const setup = answerOf(opening, "setup_character", Setup);
+      // `setup_character` drives the flow itself (`reset` then `SETUP`) and
+      // reports the position it landed in. `awaitingSetup` here would mean the
+      // roll tools are still refusing with a full character sheet in the slot.
+      expect(setup.state).toBe("playing.awaitingRoll");
+      expect(setup.initialized).toBe(true);
+      expect(setup.playerName).toMatch(/sera/i);
+
+      // A SECOND turn, which is the only place either half can be shown to have
+      // survived: the campaign lives in a session slot and the position in a
+      // dialog beside it, and `check_state` is the one tool that reports both.
+      const status = await session.say(
+        "Check the state and remind me who I am and where the story is.",
+      );
+      const now = answerOf(status, "check_state", Status);
+      expect(now.playerName).toBe(setup.playerName);
+      expect(now.state).toBe("playing.awaitingRoll");
+      expect(now.gameOver).toBe(false);
+      expect(status.completed).toBe(true);
+    },
+    {
+      stubReply: [
+        { tool: "setup_character", args: SERA },
+        "The last lamp on the north road has gone out. What do you do, Sera?",
+        { tool: "check_state" },
+        "You are Sera, a burned-out lamplighter, and the story is waiting on your next move.",
+      ],
+    },
+  );
+
+  test(
+    "a roll leaves a standing roll, and moving the scene on spends it",
+    async ({ session }) => {
+      await session.say(
+        "Dark fantasy, melancholic. My character is Sera, a burned-out lamplighter. Begin.",
+      );
+
+      const climb = await session.say("I climb the storm-lashed tower wall to reach the lamp.");
+      const rolled = answerOf(climb, "action_roll", Roll);
+      // The dice come from code, never from the narrator — two action dice and
+      // two challenge dice, on every roll.
+      expect(rolled.result.actionDice).toHaveLength(2);
+      expect(rolled.result.challengeDice).toHaveLength(2);
+      // `rollResolved` is what keeps `burn_momentum` reachable. A roll that left
+      // the flow in `awaitingRoll` would close the burn window instantly.
+      expect(rolled.state).toBe("playing.rollResolved");
+
+      const moved = await session.say(
+        "Nothing risky now. Just record that I am at the Drowned Steps and log the scene.",
+      );
+      const settled = answerOf(moved, "update_state", Settled);
+      // `update_state` sends SETTLED: the scene has moved on, so the standing
+      // roll is spent and the burn window is shut.
+      expect(settled.state).toBe("playing.awaitingRoll");
+      expect(settled.result.gameOver).toBe(false);
+    },
+    {
+      stubReply: [
+        { tool: "setup_character", args: SERA },
+        "The last lamp on the north road has gone out. What do you do, Sera?",
+        {
+          tool: "action_roll",
+          args: {
+            move: "face_danger",
+            stat: "edge",
+            position: "risky",
+            effect: "standard",
+            purpose: "climb the storm-lashed tower wall",
+          },
+        },
+        "You haul yourself up the wet stone, one handhold at a time.",
+        {
+          tool: "update_state",
+          args: { location: "The Drowned Steps", logEntry: "Descended to the drowned steps" },
+        },
+        "You take the drowned steps down, and the water closes over your boots.",
+      ],
+    },
+  );
+
+  test(
+    "both tracks empty ends the story, and the ending sticks",
+    async ({ session }) => {
+      await session.say(
+        "Dark fantasy, melancholic. My character is Sera, a burned-out lamplighter. Begin.",
+      );
+
+      const down = await session.say(
+        "## Correction. In that last scene I lost everything: my health is zero and my " +
+          "spirit is zero. Sync the state to match.",
+      );
+      const downed = answerOf(down, "update_state", Settled);
+      // The flag is DERIVED where it is reported. `gameOver` is written by the
+      // slot's `after` hook, which runs only once this body has returned — so a
+      // tool that copied `state.gameOver` into its own result would report the
+      // value from BEFORE the update that emptied the tracks, and would report
+      // `false` right here.
+      expect(downed.result.gameOver).toBe(true);
+      // And the flag is what the transition reads: `sendFrom` turns it into
+      // DOWNED, whose target is a `final` state.
+      expect(downed.state).toBe("gameOver");
+      expect(downed.done).toBe(true);
+
+      // And the ending sticks. `gameOver` is `final`, so nothing delivered to
+      // the flow can move it — a later turn still reads a finished story, which
+      // is what an `on: { SETUP }` on that state used to make look untrue.
+      const asked = await session.say("Check the state. Is the story over?");
+      const ended = answerOf(asked, "check_state", Status);
+      expect(ended.state).toBe("gameOver");
+      expect(ended.done).toBe(true);
+      expect(ended.gameOver).toBe(true);
+    },
+    {
+      stubReply: [
+        { tool: "setup_character", args: SERA },
+        "The last lamp on the north road has gone out. What do you do, Sera?",
+        { tool: "update_state", args: { health: 0, spirit: 0, logEntry: "Sera's lamp goes out" } },
+        "Your hands stop shaking, because there is nothing left in them.",
+        { tool: "check_state" },
+        "The story is over. Say the word and we will begin another.",
+      ],
+    },
+  );
+
+  test(
+    "an ended story refuses to be played",
+    async ({ session }) => {
+      await session.say(
+        "Dark fantasy, melancholic. My character is Sera, a burned-out lamplighter. Begin.",
+      );
+      const down = await session.say(
+        "## Correction. In that last scene I lost everything: my health is zero and my " +
+          "spirit is zero. Sync the state to match.",
+      );
+      expect(answerOf(down, "update_state", Settled).state).toBe("gameOver");
+
+      const after = await session.say("I refuse to die. Roll to fight on.");
+      const attempts = answersOf(after, "action_roll", Refusal);
+      // The gate is only OBSERVABLE if something calls the gated tool, and the
+      // script is what guarantees that — hence the exact count. `gameOver` was
+      // once a flag nothing acted on, so a player with both tracks empty could
+      // keep rolling for as long as they liked; what must never happen is a roll
+      // that RESOLVES.
+      expect(attempts).toHaveLength(1);
+      for (const attempt of attempts) {
+        expect(attempt.error).toMatch(/not available yet/i);
+        expect(attempt.error).toContain("gameOver");
+      }
+      expect(after.completed).toBe(true);
+    },
+    // Scripted only, and `{ scripted: true }` rather than a loop that accepts
+    // zero attempts: the narrator's own instruction on `gameOver` tells it not
+    // to roll, so a live model correctly declines and the refusal is never
+    // provoked — which used to leave this claim asserting nothing on the runs
+    // that cost money. The honest live version of "it declined" is the case
+    // above, which reads the position rather than a refusal.
+    {
+      scripted: true,
+      stubReply: [
+        { tool: "setup_character", args: SERA },
+        "The last lamp on the north road has gone out. What do you do, Sera?",
+        { tool: "update_state", args: { health: 0, spirit: 0, logEntry: "Sera's lamp goes out" } },
+        "Your hands stop shaking, because there is nothing left in them.",
+        {
+          tool: "action_roll",
+          args: {
+            move: "endure_harm",
+            stat: "iron",
+            position: "desperate",
+            effect: "limited",
+            purpose: "fight on with nothing left",
+          },
+        },
+        "There is nothing left to roll for. The hollow goes dark.",
+      ],
+    },
+  );
+});