feat: call AI Gateway through the Workers AI binding

.github/workflows/deploy.yml

@@ -31,11 +31,9 @@ jobs:
           packageManager: bun
           secrets: |
             PRIVATE_KEY
-            AI_GATEWAY_TOKEN
             INTERNAL_SECRET
             WEBHOOK_SECRET
         env:
           PRIVATE_KEY: ${{ secrets.APP_PRIVATE_KEY }}
-          AI_GATEWAY_TOKEN: ${{ secrets.AI_GATEWAY_TOKEN }}
           INTERNAL_SECRET: ${{ secrets.INTERNAL_SECRET }}
           WEBHOOK_SECRET: ${{ secrets.WEBHOOK_SECRET }}

AGENTS.md

@@ -79,13 +79,15 @@ GitHub webhook
 | ---------------------- | ----------------------------------------------------------------------------------- |
 | `APP_ID`, `WORKER_URL` | `wrangler.jsonc` 의 `vars` (공개 값)                                                |
 | `PRIVATE_KEY`          | GitHub org secret `APP_PRIVATE_KEY`                                                 |
-| `AI_GATEWAY_TOKEN`     | GitHub org secret `AI_GATEWAY_TOKEN`                                                |
+| `AI`                   | `wrangler.jsonc` 의 AI 바인딩. 같은 계정의 AI Gateway `dalestudy` 를 토큰 없이 부른다 |
 | `INTERNAL_SECRET`      | GitHub 저장소 secret `INTERNAL_SECRET`                                              |
 | `WEBHOOK_SECRET`       | GitHub 저장소 secret `WEBHOOK_SECRET`. GitHub App 웹훅 설정의 secret 과 같은 값이다 |
 
 GitHub secret 은 배포할 때마다 Worker secret 으로 들어가므로 `wrangler secret put` 으로 따로 바꾸지 않는다. 다음 배포 때 GitHub 값으로 덮인다.
 로컬에서는 `.dev.vars` 에 같은 이름으로 넣는다. `.dev.vars`, `*.pem` 은 커밋하지 않는다.
 
+OpenAI 키는 Worker 에 두지 않는다. AI Gateway `dalestudy` 의 Provider Keys(BYOK)에 저장돼 있고, 게이트웨이는 BYOK only 라 키가 없으면 Unified Billing 으로 넘어가지 않고 실패한다. AI 바인딩은 `wrangler dev` 에서도 원격 게이트웨이를 부르므로 `wrangler login` 이 dalestudy 계정에 되어 있어야 한다.
+
 ## GitHub App 설정
 
 https://github.com/organizations/DaleStudy/settings/apps/dalestudy

handlers/complexity-analysis.js

@@ -10,7 +10,7 @@
  *  - LLM : actualTime / actualSpace / feedback / suggestion / headerLine 만 책임.
  */
 
-import { aiGatewayHeaders, OPENAI_CHAT_COMPLETIONS_URL } from "../utils/constants.js";
+import { createChatCompletion } from "../utils/openai.js";
 
 // ── 상수 ──────────────────────────────────────────
 
@@ -242,35 +242,24 @@ export function composeSolution(modelSol, originalContent) {
   };
 }
 
-export async function callComplexityAnalysis(fileEntries, gatewayToken) {
+export async function callComplexityAnalysis(fileEntries, ai) {
   const userPrompt = fileEntries
     .map(
       (f) =>
         `${FILE_DELIMITER} ${f.problemName} ${FILE_DELIMITER}\n\`\`\`\n${addLineNumbers(stripComplexityComments(f.content))}\n\`\`\``
     )
     .join("\n\n");
 
-  const response = await fetch(OPENAI_CHAT_COMPLETIONS_URL, {
-    method: "POST",
-    headers: aiGatewayHeaders(gatewayToken),
-    body: JSON.stringify({
-      model: "gpt-5-nano",
-      messages: [
-        { role: "system", content: SYSTEM_PROMPT },
-        { role: "user", content: userPrompt },
-      ],
-      response_format: { type: "json_object" },
-      max_completion_tokens: 5000,
-      reasoning_effort: "minimal",
-    }),
+  const data = await createChatCompletion(ai, {
+    model: "gpt-5-nano",
+    messages: [
+      { role: "system", content: SYSTEM_PROMPT },
+      { role: "user", content: userPrompt },
+    ],
+    response_format: { type: "json_object" },
+    max_completion_tokens: 5000,
+    reasoning_effort: "minimal",
   });
-
-  if (!response.ok) {
-    const error = await response.text();
-    throw new Error(`OpenAI API error: ${error}`);
-  }
-
-  const data = await response.json();
   const content = data.choices[0]?.message?.content;
   if (!content) throw new Error("Empty response from OpenAI");
 

handlers/complexity-analysis.test.js

@@ -11,6 +11,7 @@ import {
   callComplexityAnalysis,
   renderComplexitySection,
 } from "./complexity-analysis.js";
+import { createFakeAi } from "../tests/fakeAi.js";
 
 const PLAIN_SOURCE = "function solution() { return 0; }";
 
@@ -452,6 +453,8 @@ function solution() {}`;
 // ── callComplexityAnalysis ────────────────────────
 
 describe("callComplexityAnalysis", () => {
+  const ai = createFakeAi();
+
   beforeEach(() => {
     vi.clearAllMocks();
   });
@@ -471,7 +474,7 @@ describe("callComplexityAnalysis", () => {
       { problemName: "valid-parentheses", content: PLAIN_SOURCE },
     ];
 
-    const results = await callComplexityAnalysis(fileEntries, "fake-key");
+    const results = await callComplexityAnalysis(fileEntries, ai);
 
     expect(results).toHaveLength(2);
     expect(results[0].problemName).toBe("two-sum");
@@ -485,7 +488,7 @@ describe("callComplexityAnalysis", () => {
     await expect(
       callComplexityAnalysis(
         [{ problemName: "two-sum", content: PLAIN_SOURCE }],
-        "fake-key"
+        ai
       )
     ).rejects.toThrow("OpenAI API error");
   });
@@ -498,7 +501,7 @@ describe("callComplexityAnalysis", () => {
     await expect(
       callComplexityAnalysis(
         [{ problemName: "two-sum", content: PLAIN_SOURCE }],
-        "fake-key"
+        ai
       )
     ).rejects.toThrow("Empty response from OpenAI");
   });
@@ -513,7 +516,7 @@ describe("callComplexityAnalysis", () => {
     await expect(
       callComplexityAnalysis(
         [{ problemName: "two-sum", content: PLAIN_SOURCE }],
-        "fake-key"
+        ai
       )
     ).rejects.toThrow("OpenAI returned invalid JSON");
   });
@@ -530,7 +533,7 @@ describe("callComplexityAnalysis", () => {
       { problemName: "longest-prefix", content: PLAIN_SOURCE },
     ];
 
-    const results = await callComplexityAnalysis(fileEntries, "fake-key");
+    const results = await callComplexityAnalysis(fileEntries, ai);
 
     expect(results).toHaveLength(1);
     expect(results[0].problemName).toBe("longest-prefix");
@@ -550,32 +553,39 @@ describe("callComplexityAnalysis", () => {
 function solution() { return 0; }`;
     await callComplexityAnalysis(
       [{ problemName: "two-sum", content: source }],
-      "fake-key"
+      ai
     );
 
     expect(captured).toContain("L1: \nL2: \nL3: function solution()");
     expect(captured).not.toContain("TC: O(n)");
     expect(captured).not.toContain("SC: O(1)");
   });
 
-  it("AI Gateway 인증 헤더와 재시도 헤더를 함께 보낸다", async () => {
-    let captured = null;
-    globalThis.fetch = vi.fn().mockImplementation(async (url, opts) => {
-      captured = opts.headers;
-      return await makeOpenAIResponse([makeSingleSolutionAnalysis("two-sum")]);
-    });
+  it("dalestudy 게이트웨이의 OpenAI chat/completions 로 재시도 헤더와 함께 보낸다", async () => {
+    globalThis.fetch = vi
+      .fn()
+      .mockResolvedValueOnce(
+        await makeOpenAIResponse([makeSingleSolutionAnalysis("two-sum")])
+      );
 
     await callComplexityAnalysis(
       [{ problemName: "two-sum", content: PLAIN_SOURCE }],
-      "fake-key"
+      ai
     );
 
-    expect(captured).toMatchObject({
-      "cf-aig-authorization": "Bearer fake-key",
-      "cf-aig-max-attempts": "3",
-      "cf-aig-retry-delay": "1000",
-      "cf-aig-backoff": "exponential",
+    expect(ai.gateway).toHaveBeenCalledWith("dalestudy");
+    const request = ai.run.mock.calls[0][0];
+    expect(request).toMatchObject({
+      provider: "openai",
+      endpoint: "chat/completions",
+      headers: {
+        "cf-aig-max-attempts": 3,
+        "cf-aig-retry-delay": 1000,
+        "cf-aig-backoff": "exponential",
+      },
     });
+    // 인증은 바인딩이 하므로 게이트웨이 토큰을 싣지 않는다
+    expect(request.headers).not.toHaveProperty("cf-aig-authorization");
   });
 });
 

handlers/internal-dispatch.js

@@ -72,7 +72,7 @@ async function handleTagPatterns(payload, appToken, env) {
     headSha,
     prData,
     appToken,
-    env.AI_GATEWAY_TOKEN
+    env.AI
   );
   return corsResponse({ handler: "tag-patterns", result });
 }
@@ -85,7 +85,7 @@ async function handleLearningStatus(payload, appToken, env) {
     prNumber,
     username,
     appToken,
-    env.AI_GATEWAY_TOKEN
+    env.AI
   );
   return corsResponse({ handler: "learning-status", result });
 }

handlers/internal-dispatch.test.js

@@ -17,6 +17,8 @@ import { tagPatterns } from "./tag-patterns.js";
 import { postLearningStatus } from "./learning-status.js";
 import { generateGitHubAppToken } from "../utils/github.js";
 
+const FAKE_AI = { gateway: () => {} };
+
 const VALID_SECRET = "test-secret-123";
 
 function makeRequest(pathname, { secret, body } = {}) {
@@ -88,7 +90,7 @@ describe("handleInternalDispatch — 인증", () => {
 });
 
 describe("handleInternalDispatch — 라우팅", () => {
-  const env = { INTERNAL_SECRET: VALID_SECRET, AI_GATEWAY_TOKEN: "fake-gateway-token" };
+  const env = { INTERNAL_SECRET: VALID_SECRET, AI: FAKE_AI };
 
   beforeEach(() => {
     vi.clearAllMocks();
@@ -123,7 +125,7 @@ describe("handleInternalDispatch — 라우팅", () => {
       "abc123",
       prData,
       "fake-token",
-      "fake-gateway-token"
+      FAKE_AI
     );
     expect(postLearningStatus).not.toHaveBeenCalled();
   });
@@ -154,7 +156,7 @@ describe("handleInternalDispatch — 라우팅", () => {
       42,
       "testuser",
       "fake-token",
-      "fake-gateway-token"
+      FAKE_AI
     );
     expect(tagPatterns).not.toHaveBeenCalled();
   });
@@ -200,7 +202,7 @@ describe("handleInternalDispatch — 라우팅", () => {
 });
 
 describe("handleInternalDispatch — 에러 처리", () => {
-  const env = { INTERNAL_SECRET: VALID_SECRET, AI_GATEWAY_TOKEN: "fake-gateway-token" };
+  const env = { INTERNAL_SECRET: VALID_SECRET, AI: FAKE_AI };
 
   beforeEach(() => {
     vi.clearAllMocks();

handlers/learning-status.js

@@ -93,7 +93,7 @@ function buildCategoryProgress(categories, solvedProblems) {
  * @param {number} prNumber
  * @param {string} username
  * @param {string} appToken - GitHub App installation token
- * @param {string} openaiApiKey
+ * @param {Ai} ai - Worker 의 AI 바인딩 (env.AI)
  * @returns {Promise<{ skipped: string }|{ analyzed: number, matched: number }>}
  */
 export async function postLearningStatus(
@@ -102,7 +102,7 @@ export async function postLearningStatus(
   prNumber,
   username,
   appToken,
-  openaiApiKey
+  ai
 ) {
   // 1. 문제 카테고리 메타데이터 조회
   const categories = await fetchProblemCategories(repoOwner, repoName, appToken);
@@ -167,7 +167,7 @@ export async function postLearningStatus(
         fileContent,
         submission.problemName,
         problemInfo,
-        openaiApiKey
+        ai
       );
 
       if (analysis.usage) {

handlers/tag-patterns.js

@@ -35,7 +35,7 @@ const MAX_CHECK_RUN_OUTPUT_LENGTH = 65535;
  * @param {string} headSha - PR head commit SHA
  * @param {object} prData - PR 객체 (draft, labels 포함)
  * @param {string} appToken - GitHub App installation token
- * @param {string} openaiApiKey
+ * @param {Ai} ai - Worker 의 AI 바인딩 (env.AI)
  */
 export async function tagPatterns(
   repoOwner,
@@ -44,7 +44,7 @@ export async function tagPatterns(
   headSha,
   prData,
   appToken,
-  openaiApiKey
+  ai
 ) {
   // 2-1. Skip 조건
   if (prData.draft === true) {
@@ -84,7 +84,7 @@ export async function tagPatterns(
   const fileEntries = await downloadFileEntries(solutionFiles);
 
   // 2-4. 복잡도 분석은 1콜이므로 패턴 루프와 병렬 진행. 실패해도 패턴 분석은 게시.
-  const complexityPromise = callComplexityAnalysis(fileEntries, openaiApiKey)
+  const complexityPromise = callComplexityAnalysis(fileEntries, ai)
     .catch((err) => {
       console.error(`[tagPatterns] complexity analysis failed: ${err.message}`);
       return [];
@@ -97,7 +97,7 @@ export async function tagPatterns(
       const analysis = await generatePatternAnalysis(
         fe.content,
         fe.problemName,
-        openaiApiKey
+        ai
       );
       analyses.push({ fileEntry: fe, analysis });
     } catch (error) {