thursday-agent 0.5.0 → 0.7.0

package/scripts/live-session.test.mts

@@ -0,0 +1,938 @@
+import assert from "node:assert/strict";
+import { afterEach, mock, test } from "node:test";
+import { LIVE_CALL } from "../config.ts";
+import type {
+  LiveActivity,
+  LiveReasoning,
+  LiveToolCall,
+  LiveTurn,
+} from "../lib/live/live.session.ts";
+
+let wire: {
+  on: {
+    event(event: Record<string, unknown>): void;
+    dropped(message: string): void;
+  };
+  negotiate(sdp: string): Promise<string>;
+};
+let sent: Record<string, unknown>[] = [];
+let released = false;
+mock.module("../lib/live/live.transport.ts", {
+  namedExports: {
+    createWebRtcTransport: (options: typeof wire) => {
+      wire = options;
+      return {
+        connect: async () => {
+          assert.equal(await wire.negotiate("offer"), "answer");
+          wire.on.event({ type: "session.started" });
+        },
+        send: (event: Record<string, unknown>) => sent.push(event),
+        close: () => {
+          released = true;
+        },
+      };
+    },
+  },
+});
+const { appendChunks, createLiveSession } = await import(
+  "../lib/live/live.session.ts"
+);
+const { acceptedReasoning, createLiveCall } = await import(
+  "../lib/live/live.server.ts"
+);
+
+const sessions: ReturnType<typeof createLiveSession>[] = [];
+afterEach(async () => {
+  for (const session of sessions.splice(0)) {
+    const closed = session.close();
+    wire.on.event({
+      type: "session.closed",
+      reason: "close_requested",
+      usage: { seconds: 1 },
+    });
+    await closed;
+  }
+  mock.restoreAll();
+});
+
+async function connect({
+  runTool = async () => "ok",
+}: {
+  runTool?: (call: LiveToolCall) => Promise<string>;
+} = {}) {
+  sent = [];
+  released = false;
+  const levels = { output: 0 };
+  const turns: LiveTurn[] = [];
+  const warnings: string[] = [];
+  const failures: string[] = [];
+  const activities: LiveActivity[] = [];
+  const reasonings: LiveReasoning[] = [];
+  const closes: { reason: string; seconds: number | null }[] = [];
+  const session = createLiveSession({
+    initialize: async (sdp) => {
+      assert.equal(sdp, "offer");
+      return "answer";
+    },
+    audio: {
+      element: { muted: false } as HTMLAudioElement,
+      listen() {},
+      levels: () => levels,
+    },
+    on: {
+      runTool,
+      reasoning: (part) => reasonings.push(part),
+      turn: (turn) => turns.push(turn),
+      warn: (message) => warnings.push(message),
+      failed: (message) => failures.push(message),
+      activity: (activity) => activities.push(activity),
+      finalized: (close) => closes.push(close),
+    },
+  });
+  sessions.push(session);
+  await session.connect();
+  return {
+    session,
+    levels,
+    turns,
+    warnings,
+    failures,
+    activities,
+    reasonings,
+    closes,
+  };
+}
+
+function nested(event: Record<string, unknown>) {
+  wire.on.event({
+    type: "response.event",
+    delegation_id: "delegation-a",
+    event,
+  });
+}
+const functionCall = (call_id: string) =>
+  nested({
+    type: "response.output_item.done",
+    item: {
+      type: "function_call",
+      id: `item-${call_id}`,
+      call_id,
+      name: "lookup",
+      arguments: "{}",
+    },
+  });
+const count = (type: string) =>
+  sent.filter((event) => event.type === type).length;
+const tick = () => new Promise<void>((resolve) => setImmediate(resolve));
+
+test("the backend waits for every function output and continues once, using the envelope's response", async () => {
+  const pending = new Map<string, (value: string) => void>();
+  const { activities } = await connect({
+    runTool: (call) => new Promise((resolve) => pending.set(call.id, resolve)),
+  });
+  nested({ type: "response.created", response: { id: "r1" } });
+  nested({ type: "response.function_call_arguments.done", arguments: "{}" });
+  functionCall("a");
+  functionCall("b");
+  await tick();
+  assert.equal(pending.size, 2);
+  assert.equal(activities.at(-1)?.working, true);
+  nested({ type: "response.completed", response: { id: "r1", output: [] } });
+  pending.get("a")?.("first");
+  await tick();
+  assert.equal(count("response.item.create"), 1);
+  assert.equal(count("response.create"), 0);
+  pending.get("b")?.("second");
+  await tick();
+  assert.equal(count("response.item.create"), 2);
+  assert.equal(count("response.create"), 1);
+  // A repeated terminal snapshot and a repeated item run nothing twice
+  nested({ type: "response.completed", response: { id: "r1" } });
+  functionCall("a");
+  await tick();
+  assert.equal(count("response.create"), 1);
+  assert.equal(count("response.item.create"), 2);
+});
+
+test("an incomplete response that asked for tools is continued once, and a second in a row only warns", async () => {
+  const { warnings } = await connect();
+  nested({ type: "response.created", response: { id: "r1" } });
+  functionCall("a");
+  nested({ type: "response.incomplete", response: { id: "r1" } });
+  await tick();
+  assert.equal(count("response.item.create"), 1);
+  assert.equal(count("response.create"), 1);
+  assert.equal(warnings.length, 0);
+  // The continuation runs out again: its output still goes in, and nothing continues it
+  nested({ type: "response.created", response: { id: "r2" } });
+  functionCall("b");
+  nested({ type: "response.incomplete", response: { id: "r2" } });
+  await tick();
+  assert.equal(count("response.item.create"), 2);
+  assert.equal(count("response.create"), 1);
+  assert.equal(warnings.length, 1);
+  // A failed response is never continued
+  nested({ type: "response.created", response: { id: "r3" } });
+  functionCall("c");
+  nested({ type: "response.failed", response: { id: "r3" } });
+  await tick();
+  assert.equal(count("response.item.create"), 3);
+  assert.equal(count("response.create"), 1);
+});
+
+test("a call completed before one the output cap cut off still has its result continued", async () => {
+  let ran = 0;
+  await connect({
+    runTool: async () => {
+      ran += 1;
+      return "ok";
+    },
+  });
+  nested({ type: "response.created", response: { id: "r1" } });
+  functionCall("a");
+  nested({
+    type: "response.output_item.done",
+    item: {
+      type: "function_call",
+      id: "item-cut",
+      call_id: "cut",
+      name: "delegate",
+      arguments: '{"bot":"Ana',
+      status: "incomplete",
+    },
+  });
+  await tick();
+  await tick();
+  assert.equal(ran, 1);
+  assert.equal(count("response.item.create"), 1);
+  assert.equal(count("response.create"), 1);
+  // The continuation is cut off the same way: its result goes in, nothing continues it
+  nested({ type: "response.created", response: { id: "r2" } });
+  functionCall("b");
+  nested({
+    type: "response.output_item.done",
+    item: {
+      type: "function_call",
+      id: "item-cut-again",
+      call_id: "cut-again",
+      name: "delegate",
+      arguments: '{"bot":"Ana',
+      status: "incomplete",
+    },
+  });
+  await tick();
+  await tick();
+  assert.equal(ran, 2);
+  assert.equal(count("response.item.create"), 2);
+  assert.equal(count("response.create"), 1);
+});
+
+test("a finished reasoning summary part is reported once, whole, with its place in the call", async () => {
+  const { reasonings } = await connect();
+  nested({ type: "response.created", response: { id: "r1" } });
+  const part = (event: Record<string, unknown>) =>
+    nested({ item_id: "rs_1", output_index: 0, ...event });
+  part({
+    type: "response.reasoning_summary_text.delta",
+    summary_index: 0,
+    delta: "**Comparing",
+  });
+  part({
+    type: "response.reasoning_summary_text.done",
+    summary_index: 0,
+    text: "**Comparing markets**\n\nRates first.",
+  });
+  part({
+    type: "response.reasoning_summary_text.done",
+    summary_index: 1,
+    text: "**Handing over**",
+  });
+  await tick();
+  assert.deepEqual(
+    reasonings.map(({ id, text }) => [id, text]),
+    [
+      ["rs_1:0", "**Comparing markets**\n\nRates first."],
+      ["rs_1:1", "**Handing over**"],
+    ],
+  );
+  assert.ok(reasonings.every((part) => part.seq >= 0));
+});
+
+test("a function call cut off by the output cap is never run, and its response stops counting as work", async () => {
+  let ran = 0;
+  const { activities, turns } = await connect({
+    runTool: async () => {
+      ran += 1;
+      return "ok";
+    },
+  });
+  nested({ type: "response.created", response: { id: "r1" } });
+  await tick();
+  assert.equal(activities.at(-1)?.working, true);
+  nested({
+    type: "response.output_item.done",
+    item: {
+      type: "function_call",
+      id: "item-cut",
+      call_id: "cut",
+      name: "delegate",
+      arguments: '{"bot":"Analyst',
+      status: "incomplete",
+    },
+  });
+  // What Live sends instead of a terminal event for that response
+  wire.on.event({
+    type: "error",
+    error: { message: "Responses handoff incomplete." },
+  });
+  await tick();
+  assert.equal(ran, 0);
+  assert.equal(turns.filter((turn) => turn.role === "tool").length, 0);
+  assert.equal(count("response.item.create"), 0);
+  assert.equal(count("response.create"), 0);
+  assert.equal(activities.at(-1)?.working, false);
+});
+
+test("captions keep exact fragments through overlap and late delivery, and never show backend text", async (context) => {
+  context.mock.timers.enable({ apis: ["setTimeout"] });
+  const { session, turns } = await connect();
+  const fragment = (
+    type: string,
+    id: string,
+    start: number,
+    end: number,
+    delta: string,
+  ) =>
+    wire.on.event({ type, event_id: id, start_ms: start, end_ms: end, delta });
+  fragment("session.input_transcript.delta", "u1", 100, 200, "Please ");
+  fragment("session.output_transcript.delta", "a1", 150, 250, "Sure.");
+  fragment("session.input_transcript.delta", "u3", 400, 500, "it.");
+  fragment("session.input_transcript.delta", "u2", 250, 350, "find ");
+  fragment("session.input_transcript.delta", "u2", 250, 350, "find ");
+  nested({ type: "response.output_text.delta", delta: "private backend text" });
+  context.mock.timers.tick(LIVE_CALL.transcriptSaveMs);
+
+  const saved = turns.filter((turn) => turn.done);
+  assert.deepEqual(
+    saved.map((turn) => turn.text),
+    ["Please find it.", "Sure."],
+  );
+  assert.equal(saved[0].seq, 100);
+  assert.equal(saved[1].seq, 150);
+  assert.deepEqual(saved[0].fragments, [
+    { start: 100, end: 200, text: "Please " },
+    { start: 250, end: 350, text: "find " },
+    { start: 400, end: 500, text: "it." },
+  ]);
+
+  // A fragment after the checkpoint revises the same group, originals included
+  fragment("session.input_transcript.delta", "u4", 520, 600, " Now.");
+  context.mock.timers.tick(LIVE_CALL.transcriptSaveMs);
+  const revised = turns
+    .filter((turn) => turn.done && turn.id === saved[0].id)
+    .at(-1);
+  assert.equal(revised?.text, "Please find it. Now.");
+  assert.equal(revised?.fragments?.length, 4);
+  assert.equal(
+    turns.some((turn) => turn.text.includes("private backend text")),
+    false,
+  );
+
+  const close = session.close();
+  assert.equal(released, false);
+  wire.on.event({
+    type: "session.closed",
+    reason: "close_requested",
+    usage: { seconds: 12 },
+  });
+  await close;
+  assert.equal(released, true);
+});
+
+test("closing never starts a late tool", async () => {
+  let executed = 0;
+  const { session, failures } = await connect({
+    runTool: async () => {
+      executed++;
+      return "ok";
+    },
+  });
+  const closing = session.close();
+  nested({ type: "response.created", response: { id: "r" } });
+  functionCall("late");
+  wire.on.event({ type: "session.closed", reason: "close_requested" });
+  await closing;
+  assert.equal(executed, 0);
+  assert.equal(failures.length, 0);
+});
+
+test("closing waits for session.closed and reports the reason and billed seconds", async () => {
+  const { session, closes, failures } = await connect();
+  const closing = session.close();
+  assert.equal(sent.at(-1)?.type, "session.close");
+  assert.equal(released, false);
+  wire.on.event({
+    type: "session.closed",
+    reason: "close_requested",
+    usage: { seconds: 42 },
+  });
+  await closing;
+  assert.deepEqual(closes, [{ reason: "close_requested", seconds: 42 }]);
+  assert.equal(failures.length, 0);
+  assert.equal(released, true);
+});
+
+test("a session the provider closes fails the call and still reports usage", async () => {
+  const { closes, failures } = await connect();
+  wire.on.event({ type: "session.closed", reason: "idle_timeout" });
+  assert.deepEqual(closes, [{ reason: "idle_timeout", seconds: null }]);
+  assert.match(failures[0], /idle_timeout/);
+  assert.equal(released, true);
+});
+
+test("a close with no confirmation times out, says so, and releases media", async (context) => {
+  context.mock.timers.enable({ apis: ["setTimeout", "setInterval"] });
+  const { session, warnings, closes } = await connect();
+  const closing = session.close();
+  context.mock.timers.tick(LIVE_CALL.closeMs);
+  await closing;
+  assert.equal(released, true);
+  assert.equal(closes.length, 0);
+  assert.match(warnings[0], /final session usage/);
+});
+
+test("activity keeps reporting while she speaks, and names the tools while they run", async (context) => {
+  context.mock.timers.enable({ apis: ["setInterval"] });
+  const pending: ((value: string) => void)[] = [];
+  const { activities, levels } = await connect({
+    runTool: () => new Promise((resolve) => pending.push(resolve)),
+  });
+  levels.output = 0.2;
+  const before = activities.length;
+  context.mock.timers.tick(100);
+  context.mock.timers.tick(100);
+  context.mock.timers.tick(100);
+  assert.ok(
+    activities.slice(before).filter((activity) => activity.speaking).length >=
+      3,
+  );
+
+  nested({ type: "response.created", response: { id: "r1" } });
+  functionCall("a");
+  await tick();
+  assert.equal(activities.at(-1)?.working, true);
+  assert.deepEqual(activities.at(-1)?.tools, ["lookup"]);
+  pending[0]("done");
+  await tick();
+});
+
+test("updates go out one at a time by kind, settle on their own acknowledgement, and are never replayed", async (context) => {
+  context.mock.timers.enable({ apis: ["setTimeout"] });
+  const { session, warnings } = await connect();
+  const first = session.append("instructions", "Greet the user.");
+  const second = session.append("commentary", "Scout found three flights.");
+  assert.equal(sent.length, 1);
+  assert.equal(sent[0].type, "session.instructions.append");
+  assert.equal(sent[0].delegation_id, null);
+  assert.equal(sent[0].content, "Greet the user.");
+
+  context.mock.timers.tick(LIVE_CALL.appendMs);
+  assert.equal(await first, false);
+  assert.match(warnings[0], /delivery is unknown/);
+  assert.equal(sent.length, 2);
+  assert.equal(sent[1].type, "session.commentary.append");
+
+  // A late acknowledgement for the first sends nothing more
+  wire.on.event({
+    type: "session.instructions.appended",
+    client_event_id: sent[0].event_id,
+  });
+  assert.equal(sent.length, 2);
+  wire.on.event({
+    type: "session.commentary.appended",
+    client_event_id: sent[1].event_id,
+  });
+  assert.equal(await second, true);
+  assert.equal(sent.length, 2);
+});
+
+test("a long update goes chunk by chunk, and a rejected chunk drops the rest of it", async () => {
+  const { session, warnings } = await connect();
+  const long = "word ".repeat(200);
+  const whole = session.append("commentary", long);
+  const parts: string[] = [];
+  while (sent.length > parts.length) {
+    const chunk = sent[parts.length];
+    parts.push(String(chunk.content));
+    wire.on.event({
+      type: "session.commentary.appended",
+      client_event_id: chunk.event_id,
+    });
+  }
+  assert.equal(await whole, true);
+  assert.ok(parts.length > 1);
+  assert.equal(parts.join(""), long);
+
+  const before = sent.length;
+  const rejected = session.append("commentary", long);
+  wire.on.event({
+    type: "error",
+    error: {
+      message: "Update refused",
+      client_event_id: sent[before].event_id,
+    },
+  });
+  assert.equal(await rejected, false);
+  assert.equal(sent.length, before + 1);
+  assert.match(warnings.at(-1) ?? "", /Update refused/);
+
+  const next = session.append(
+    "thinking",
+    "The user stopped a thread on screen.",
+  );
+  assert.equal(sent.at(-1)?.type, "session.thinking.append");
+  wire.on.event({
+    type: "session.thinking.appended",
+    client_event_id: sent.at(-1)?.event_id,
+  });
+  assert.equal(await next, true);
+});
+
+test("update chunks stay within the byte bound for any script and keep words whole", () => {
+  const encoder = new TextEncoder();
+  // Three UTF-8 bytes each, with spaces between words
+  const symbols = "\u2600\u2601\u2602\u2603 ".repeat(80);
+  const emoji = "🙂".repeat(300);
+  const english = "The flight leaves at nine. ".repeat(60);
+  for (const text of [symbols, emoji, english]) {
+    const chunks = appendChunks(text);
+    assert.ok(chunks.length > 1);
+    assert.equal(chunks.join(""), text);
+    for (const chunk of chunks) assert.ok(encoder.encode(chunk).length <= 480);
+  }
+  for (const chunk of appendChunks(english).slice(0, -1)) {
+    assert.match(chunk, /\s$/);
+  }
+  assert.deepEqual(appendChunks(""), []);
+});
+
+const backend = (overrides: Record<string, unknown> = {}) => ({
+  model: "gpt-5.6-luna",
+  instructions: "Tools only",
+  tools: [
+    {
+      name: "lookup",
+      description: "Look up an item",
+      parameters: { type: "object" as const, properties: {} },
+    },
+  ],
+  reasoning: null,
+  webSearch: false,
+  ...overrides,
+});
+
+function captureFetch() {
+  const requests: Record<string, any>[] = [];
+  mock.method(globalThis, "fetch", async (url: string, init: RequestInit) => {
+    assert.equal(url, "https://api.openai.com/v1/live/sessions");
+    requests.push(JSON.parse(init.body as string));
+    return Response.json(
+      {
+        session: { id: "live-example" },
+        transport: { type: "webrtc", sdp: "answer" },
+      },
+      { status: 201 },
+    );
+  });
+  return requests;
+}
+
+test("startup sends voice and backend apart, seeds no history, and keeps the key on the server", async () => {
+  const requests = captureFetch();
+  const connection = await createLiveCall({
+    apiKey: "test-key",
+    sdp: "offer",
+    voice: "marin",
+    instructions: "Conversation only",
+    backend: backend(),
+  });
+  const [request] = requests;
+  assert.equal(request.session.model, "gpt-live-1");
+  assert.equal(request.session.instructions, "Conversation only");
+  // Earlier calls are the backend's to read; the voice opens on nothing but its instructions
+  assert.equal("input" in request.session, false);
+  assert.equal(request.session.audio.output.voice, "marin");
+  assert.equal(request.session.store, false);
+  assert.equal(request.session.delegation.type, "responses");
+  const responses = request.session.delegation.responses;
+  assert.equal(responses.model, "gpt-5.6-luna");
+  assert.equal(responses.instructions, "Tools only");
+  assert.equal(responses.tools.length, 1);
+  assert.equal("strict" in responses.tools[0], false);
+  assert.equal("reasoning" in responses, false);
+  assert.equal(
+    responses.tools.some(
+      (tool: { type: string }) => tool.type === "web_search",
+    ),
+    false,
+  );
+  assert.equal(JSON.stringify(connection).includes("test-key"), false);
+});
+
+test("reasoning and web search are sent only when given", async () => {
+  const requests = captureFetch();
+  await createLiveCall({
+    apiKey: "test-key",
+    sdp: "offer",
+    voice: "cedar",
+    instructions: "Talk",
+    backend: backend({
+      model: "gpt-4.1",
+      reasoning: { effort: "low", summary: "auto" },
+      webSearch: true,
+    }),
+  });
+  const responses = requests[0].session.delegation.responses;
+  assert.equal(responses.model, "gpt-4.1");
+  assert.deepEqual(responses.reasoning, { effort: "low", summary: "auto" });
+  assert.deepEqual(responses.tools.at(-1), { type: "web_search" });
+});
+
+test("a provider refusal reaches the caller unchanged", async () => {
+  mock.method(globalThis, "fetch", async () =>
+    Response.json(
+      { error: { message: "This project cannot access gpt-live-1" } },
+      { status: 403 },
+    ),
+  );
+  await assert.rejects(
+    createLiveCall({
+      apiKey: "test",
+      sdp: "offer",
+      voice: "marin",
+      instructions: "Talk",
+      backend: backend(),
+    }),
+    /This project cannot access gpt-live-1/,
+  );
+});
+
+/** A token-count endpoint whose models refuse the reasoning settings they are given. */
+function refusingFetch(refuses: Record<string, string[]>) {
+  const asked: string[] = [];
+  mock.method(globalThis, "fetch", async (url: string, init: RequestInit) => {
+    assert.equal(url, "https://api.openai.com/v1/responses/input_tokens");
+    const body = JSON.parse(init.body as string);
+    asked.push(`${body.model} ${JSON.stringify(body.reasoning)}`);
+    const refused = (refuses[body.model] ?? []).find(
+      (setting) => setting in body.reasoning,
+    );
+    return refused
+      ? Response.json(
+          {
+            error: {
+              message: `Unsupported parameter: 'reasoning.${refused}' is not supported with this model.`,
+              param: `reasoning.${refused}`,
+              code: "unsupported_parameter",
+            },
+          },
+          { status: 400 },
+        )
+      : Response.json({ input_tokens: 7 });
+  });
+  return asked;
+}
+
+test("a reasoning setting the backend model refuses is dropped before the call, and the answer is kept", async () => {
+  const asked = refusingFetch({ "gpt-4.1": ["effort"], legacy: ["summary"] });
+  const check = (model: string, effort: string | null) =>
+    acceptedReasoning({ apiKey: "test", model, effort });
+
+  assert.deepEqual(await check("gpt-4.1", "low"), { summary: "auto" });
+  assert.deepEqual(await check("gpt-4.1", "low"), { summary: "auto" });
+  assert.deepEqual(await check("gpt-5.6-luna", "low"), {
+    effort: "low",
+    summary: "auto",
+  });
+  assert.deepEqual(await check("gpt-5.6-luna", null), { summary: "auto" });
+  assert.equal(await check("gpt-4.1", "none"), null);
+  assert.deepEqual(await check("legacy", "high"), { effort: "high" });
+  assert.deepEqual(asked, [
+    'gpt-4.1 {"effort":"low","summary":"auto"}',
+    'gpt-4.1 {"summary":"auto"}',
+    'gpt-5.6-luna {"effort":"low","summary":"auto"}',
+    'gpt-5.6-luna {"summary":"auto"}',
+    'gpt-4.1 {"effort":"none"}',
+    'legacy {"effort":"high","summary":"auto"}',
+    'legacy {"effort":"high"}',
+  ]);
+});
+
+test("any other answer keeps the chosen settings and is asked again next call", async () => {
+  let asked = 0;
+  mock.method(globalThis, "fetch", async () => {
+    asked += 1;
+    return asked === 1
+      ? Response.json(
+          { error: { message: "Incorrect API key provided", param: null } },
+          { status: 401 },
+        )
+      : Promise.reject(new TypeError("fetch failed"));
+  });
+  const check = () =>
+    acceptedReasoning({ apiKey: "test", model: "gpt-5.6-sol", effort: "high" });
+
+  assert.deepEqual(await check(), { effort: "high", summary: "auto" });
+  assert.deepEqual(await check(), { effort: "high", summary: "auto" });
+  assert.equal(asked, 2);
+});
+
+test("stored settings keep OpenAI choices and the shared instruction, drop a Grok voice, and recover field by field", async () => {
+  const { LIVE_DEFAULTS, migrateLiveSettings } = await import(
+    "../features/ai/live.schema.ts"
+  );
+  const openai = migrateLiveSettings({
+    model: {
+      provider: "openai",
+      voice: "cedar",
+      model: "gpt-live-1",
+      backendModel: "gpt-5.6-sol",
+    },
+    systemPrompt: "Call me Sam.",
+    captionView: "sides",
+  });
+  assert.equal(openai.voice, "cedar");
+  assert.equal(openai.backendModel, "gpt-5.6-sol");
+  assert.equal(openai.voicePrompt, "Call me Sam.");
+  assert.equal(openai.backendPrompt, "Call me Sam.");
+  assert.equal(openai.captionView, "sides");
+  assert.equal(openai.reasoningEffort, LIVE_DEFAULTS.reasoningEffort);
+  assert.equal(openai.webSearch, false);
+  assert.equal("model" in openai, false);
+  assert.equal("systemPrompt" in openai, false);
+
+  const grok = migrateLiveSettings({
+    model: { provider: "xai", voice: "Ara", model: "grok-voice" },
+  });
+  assert.equal(grok.voice, LIVE_DEFAULTS.voice);
+  assert.equal(grok.backendModel, LIVE_DEFAULTS.backendModel);
+
+  const broken = migrateLiveSettings({
+    voice: 42,
+    backendModel: "gpt-4.1",
+    reasoningEffort: "extreme",
+    webSearch: true,
+  });
+  assert.equal(broken.voice, LIVE_DEFAULTS.voice);
+  assert.equal(broken.backendModel, "gpt-4.1");
+  assert.equal(broken.reasoningEffort, LIVE_DEFAULTS.reasoningEffort);
+  assert.equal(broken.webSearch, true);
+
+  // A stored choice survives a change of default: auto stays auto
+  assert.equal(
+    migrateLiveSettings({ reasoningEffort: null }).reasoningEffort,
+    null,
+  );
+
+  assert.deepEqual(migrateLiveSettings(null), LIVE_DEFAULTS);
+});
+
+test("both call prompts open as one Thursday: the voice gets a two-line delegation policy and memory and no earlier calls, the backend asks before handing work over, and every call has an opening", async () => {
+  let profileFacts = 200;
+  let samFacts = 2;
+  const botMock = mock.module("../features/bot/bot.query.ts", {
+    namedExports: {
+      listJobBots: async () => [
+        { name: "Scout", description: "Finds things out on the web" },
+      ],
+      readBotMemoryOn: async () => true,
+    },
+  });
+  const skillsMock = mock.module("../features/skills/skills.discover.ts", {
+    namedExports: {
+      loadSkills: async () => [
+        { name: "browser", description: "Any browser.", path: "/browser" },
+      ],
+    },
+  });
+  const connectedMock = mock.module("../features/ai/tools/connected.ts", {
+    namedExports: {
+      listConnectedToolNames: async () => [
+        { server: "studio", name: "generate_image" },
+        { server: "github", name: "create_issue" },
+      ],
+    },
+  });
+  const workspaceMock = mock.module("../features/workspace/workspace.ts", {
+    namedExports: { openWorkspace: async () => ({ cwd: "/workspace" }) },
+  });
+  const memoryMock = mock.module("../features/memory/memory.query.ts", {
+    namedExports: {
+      listAlwaysLoaded: async () => [
+        { id: 1, path: "profile", text: "Prefer brief replies." },
+      ],
+      readNotes: async () => ({
+        notes: [
+          {
+            path: "profile",
+            facts: Array.from({ length: profileFacts }, (_, index) => ({
+              id: index + 1,
+              text: index === 0 ? "Prefer brief replies." : `Fact ${index}`,
+            })),
+          },
+        ],
+      }),
+      listNoteIndex: async () => [
+        {
+          path: "people/sam",
+          description: "Their brother",
+          aliases: ["Sam"],
+          factCount: samFacts,
+          lastSeenAt: new Date(),
+        },
+      ],
+    },
+  });
+  const callMock = mock.module("../features/thursday/thursday.query.ts", {
+    namedExports: {
+      listRecentTurns: async () => [
+        {
+          callId: "c1",
+          startedAt: new Date("2026-09-13T10:00:00Z"),
+          turns: [
+            { role: "user", tool: null, text: "Book the dentist.", seq: 1 },
+            { role: "tool", tool: "delegate", text: '{"bot":"Scout"}', seq: 2 },
+            { role: "assistant", tool: null, text: "Scout has it.", seq: 3 },
+          ],
+        },
+      ],
+      readCallSkillsOn: async () => false,
+    },
+  });
+  const threadMock = mock.module("../features/bot/thread.query.ts", {
+    namedExports: { listCallJobs: async () => [] },
+  });
+  try {
+    const { loadLivePrompt } = await import(
+      "../features/ai/prompts/live.prompt.ts"
+    );
+    const { loadThursdayPrompt } = await import(
+      "../features/ai/prompts/thursday.prompt.ts"
+    );
+    const on = await loadLivePrompt({
+      voicePrompt: "Use a calm voice.",
+      locale: "ko-KR",
+    });
+    assert.match(on.text, /modeled on Friday, the AI in \*Iron Man\*/);
+    assert.match(on.text, /Prefer brief replies/);
+    assert.match(
+      on.text,
+      /\n\n## Always\n\nIMPORTANT — always follow this: [^\n]+\n\nBackchannel policy: Use moderate backchannels\. .*\n\nInterruption policy: Stop speaking when the user interrupts\. Listen to what they say\.\n\nDelegation policy:\nDelegate to the backend everything [^\n]+\nWhatever they tell you about themselves[^\n]+unless it is already written below\.\nAnswer from what the backend returns[^\n]+\n\n## What you know about them\n/,
+    );
+    // Only the ending rule carries the stamp
+    assert.equal(on.text.split("IMPORTANT").length, 2);
+    assert.match(
+      on.text,
+      /When the user speaks at length, acknowledge now and then/,
+    );
+    // What the backend can do, and when to hand over, are no longer the voice's to read
+    assert.equal(
+      /Backend tools|Delegate to the backend when|What bots can reach for|- Web:/.test(
+        on.text,
+      ),
+      false,
+    );
+    assert.match(on.text, /- people\/sam — Their brother \(2\) "Sam"/);
+    assert.match(on.text, /What is in these notes, the backend recalls\./);
+    // The roster, threads and earlier calls are the backend's to read
+    assert.equal(on.text.includes("Scout"), false);
+    assert.equal(
+      /thread|\bseen\b|## Earlier calls|dentist|works beside you/.test(on.text),
+      false,
+    );
+    assert.equal("input" in on, false);
+    assert.equal(
+      /ko-KR|browser's setting|language they use/.test(on.text),
+      false,
+    );
+    assert.equal(/memory_|generate_|load_skill|`/.test(on.text), false);
+    // The one tool name the voice holds: the ending rule, where it makes ending a thing to do
+    assert.equal(on.text.split("end_call").length, 2);
+    assert.match(
+      on.text,
+      /immediately, without thinking, use the end_call tool\./,
+    );
+    assert.equal(on.text.endsWith("Use a calm voice."), true);
+    assert.match(on.opening, /The call has just started/);
+
+    // One Thursday: the backend opens with the voice's own words and is never told it is a part
+    const backend = await loadThursdayPrompt(null);
+    const withoutClock = (text: string) =>
+      text.replace(/\*\*Now\*\*: [^\n]+/, "");
+    const identity = on.text.slice(0, on.text.indexOf("\n\n## Always"));
+    const ending = /IMPORTANT — always follow this: [^\n]+/.exec(on.text)?.[0];
+    assert.equal(
+      withoutClock(backend).startsWith(
+        `${withoutClock(identity)}\n\n${ending}\n\n`,
+      ),
+      true,
+    );
+    assert.match(
+      backend,
+      /\*\*Keep memory clean as you write\.\*\* A fact that repeats, narrows or changes one already in the note replaces it/,
+    );
+    for (const heading of [
+      "## Memory",
+      "## Background work",
+      "## This computer",
+      "## Return the result",
+      "## Earlier calls",
+    ])
+      assert.equal(backend.includes(`\n${heading}\n`), true, heading);
+    assert.equal(backend.includes("## Voice conversation context"), false);
+    assert.equal(/backend of Thursday|voice model/.test(backend), false);
+    // The end_call name sits in the ending rule only
+    assert.equal(backend.split("end_call").length, 2);
+    assert.match(backend, /\*\*Ask once before handing work over\.\*\*/);
+    assert.match(backend, /Each bot keeps its own memory from job to job/);
+    assert.match(backend, /they come from bots, not the user/);
+    assert.match(backend, /Book the dentist\./);
+    assert.equal(backend.includes("grown past what it holds well"), false);
+
+    // Tidying memory is the backend's: the voice neither reads it nor opens with it
+    samFacts = 60;
+    const heavy = await loadLivePrompt({ locale: null });
+    assert.equal(/tidying|grown past/.test(heavy.text), false);
+    assert.match(heavy.opening, /The call has just started/);
+    assert.match(
+      await loadThursdayPrompt(null),
+      /grown past what it holds well \(people\/sam\): say so once in what you return/,
+    );
+    samFacts = 2;
+
+    profileFacts = 0;
+    const first = await loadLivePrompt({ locale: "ko-KR" });
+    assert.match(first.text, /## First call/);
+    // The one call that opens with nothing: she says who she is, then learns who they are
+    assert.match(first.opening ?? "", /say who you are/);
+    assert.match(first.opening ?? "", /ask what to call them/);
+    assert.match(first.opening ?? "", /ko-KR/);
+    assert.match(first.text, /what they do, how old they are, where they live/);
+
+    // A call the page placed says so, ahead of even the first-call opening
+    const rung = await loadLivePrompt({
+      locale: "ko-KR",
+      calledBack: true,
+    });
+    assert.match(rung.opening, /You placed this call/);
+    assert.match(rung.opening, /say that is why you called/);
+  } finally {
+    memoryMock.restore();
+    callMock.restore();
+    botMock.restore();
+    skillsMock.restore();
+    connectedMock.restore();
+    workspaceMock.restore();
+    threadMock.restore();
+  }
+});