test(ai): step 2 — cover Luna streaming hotfix

Checkpoint: verifies the production streaming request and isolates environment overrides.
Reviewed against testing/hotfix-luna-chat-no-explicit-reasoning.md.

Author: Atharva-Kanherkar

web/src/lib/ai-coach.test.ts

@@ -9,6 +9,7 @@ import {
   normalizeReasoningEffort,
   reasoningEffortForModel,
   runCoach,
+  streamChatCoach,
   type CoachRequest,
 } from "./ai-coach";
 
@@ -21,6 +22,17 @@ function mockOpenAi(content: string, ok = true, status = 200) {
   } as unknown as Response);
 }
 
+function mockOpenAiStream(content: string) {
+  const body = new ReadableStream({
+    start(controller) {
+      const payload = JSON.stringify({ choices: [{ delta: { content } }] });
+      controller.enqueue(new TextEncoder().encode(`data: ${payload}\n\ndata: [DONE]\n\n`));
+      controller.close();
+    },
+  });
+  return vi.fn().mockResolvedValue({ ok: true, status: 200, body } as unknown as Response);
+}
+
 const baseReq: CoachRequest = {
   mode: "explain",
   word: "見る",
@@ -111,6 +123,44 @@ describe("runCoach", () => {
   });
 });
 
+describe("streamChatCoach", () => {
+  const chatReq: CoachRequest = {
+    ...baseReq,
+    mode: "chat",
+    message: "What does this mean?",
+  };
+
+  it("omits reasoning_effort from the default Luna streaming request", async () => {
+    const fetchMock = mockOpenAiStream("It means to look.");
+    vi.stubGlobal("fetch", fetchMock);
+
+    const chunks: string[] = [];
+    for await (const chunk of streamChatCoach("sk-test", DEFAULT_COACH_MODEL, chatReq)) {
+      chunks.push(chunk);
+    }
+
+    expect(chunks).toEqual(["It means to look."]);
+    const body = JSON.parse((fetchMock.mock.calls[0][1] as RequestInit).body as string);
+    expect(body).toMatchObject({ model: "gpt-5.6-luna", stream: true });
+    expect(body.max_completion_tokens).toBeGreaterThan(400);
+    expect(body).not.toHaveProperty("reasoning_effort");
+    expect(body).not.toHaveProperty("temperature");
+    expect(body).not.toHaveProperty("max_tokens");
+  });
+
+  it("preserves an explicit supported streaming effort", async () => {
+    const fetchMock = mockOpenAiStream("It means to look.");
+    vi.stubGlobal("fetch", fetchMock);
+
+    for await (const chunk of streamChatCoach("sk-test", DEFAULT_COACH_MODEL, chatReq, "low")) {
+      expect(chunk).toBe("It means to look.");
+    }
+
+    const body = JSON.parse((fetchMock.mock.calls[0][1] as RequestInit).body as string);
+    expect(body.reasoning_effort).toBe("low");
+  });
+});
+
 describe("reasoning model tuning", () => {
   it("defaults the coach to gpt-5.6-luna and OpenAI's medium effort", () => {
     expect(DEFAULT_COACH_MODEL).toBe("gpt-5.6-luna");

web/src/lib/ai-coach.ts

@@ -72,7 +72,7 @@ export interface ChatResult {
 export type CoachResult = ExplainResult | HooksResult | ChatResult;
 
 // Verified August 2026: gpt-5.6-luna (released 2026-07-09) is $0.20 / 1M input,
-// $1.20 / 1M output, $0.02 / 1M cached input. It is a reasoning model; at the
+// $1.20 / 1M output, $0.02 / 1M cached input. It is a reasoning model, so
 // reasoning tokens are billed as output. The actual cost is workload-dependent
 // and can exceed the backend's $0.002/call assumption, so usage must be
 // monitored. The model is overridable via AI_COACH_MODEL; reasoning effort is

web/src/lib/ai-store.test.ts

@@ -14,11 +14,11 @@ const { currentMonth, getCoachConfig, getUsage, incrementUsage, quotaFor, refund
 
 const MONTH = "2026-08";
 
-describe("usage buckets", () => {
-  beforeEach(() => {
-    vi.unstubAllEnvs();
-  });
+beforeEach(() => {
+  vi.unstubAllEnvs();
+});
 
+describe("usage buckets", () => {
   // The bug this pins: pronunciation audio and smart word picking were metered
   // against the same counter as the coach, so a learner could exhaust the
   // advertised "AI messages" allowance without ever opening the coach.