diff --git a/AGENTS.md b/AGENTS.md
index e1d4de1..aed8e74 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -487,6 +487,10 @@ describe("checkWeeks", () => {
    - GitHub 인증 로직 (`generateGitHubAppToken`, `createJWT` 등)은 모든 기능에서 공통으로 사용
    - 새 기능 추가 시 기존 유틸리티 함수 활용
 
+7. **코멘트 숨김 마커 직렬화 포맷 변경**
+   - 코멘트에 `<!-- xxx-data: ... -->` 형태로 숨겨 저장하는 데이터의 직렬화 포맷(객체↔배열 등)을 바꿀 때는 **정규식·문서 주석·테스트를 같은 PR에서 함께 갱신**
+   - 파싱이 `Array.isArray` 같은 방어 코드로 빈 값에 fallback하면 회귀가 조용히 묻혀 디버깅이 어려워짐
+
 ## 관련 문서
 
 - [Cloudflare Workers Docs](https://developers.cloudflare.com/workers/)
diff --git a/tests/learningComment.test.js b/tests/learningComment.test.js
new file mode 100644
index 0000000..eb74a3d
--- /dev/null
+++ b/tests/learningComment.test.js
@@ -0,0 +1,179 @@
+import { describe, it, expect, beforeEach } from "bun:test";
+
+import { upsertLearningStatusComment } from "../utils/learningComment.js";
+
+const REPO_OWNER = "DaleStudy";
+const REPO_NAME = "leetcode-study";
+const PR_NUMBER = 42;
+const APP_TOKEN = "fake-app-token";
+const COMMENT_MARKER = "<!-- dalestudy-learning-status -->";
+
+function ok(body) {
+  return Promise.resolve({
+    ok: true,
+    status: 200,
+    statusText: "OK",
+    json: () => Promise.resolve(body),
+  });
+}
+
+function parseBody(call) {
+  return JSON.parse(call.init.body).body;
+}
+
+describe("upsertLearningStatusComment — usage history accumulation", () => {
+  let originalFetch;
+
+  beforeEach(() => {
+    originalFetch = globalThis.fetch;
+  });
+
+  it("posts a new comment with a single usage row when no prior comment exists", async () => {
+    const calls = [];
+    globalThis.fetch = (url, init = {}) => {
+      calls.push({ url: String(url), init });
+      if (init.method === "POST") return ok({});
+      return ok([]); // no existing comments
+    };
+
+    await upsertLearningStatusComment(
+      REPO_OWNER,
+      REPO_NAME,
+      PR_NUMBER,
+      `${COMMENT_MARKER}\n## body`,
+      APP_TOKEN,
+      { prompt_tokens: 100, completion_tokens: 50 }
+    );
+
+    const post = calls.find((c) => c.init.method === "POST");
+    expect(post).toBeDefined();
+
+    const body = parseBody(post);
+    expect(body).toContain("🔢 API 사용량 (gpt-4.1-nano)");
+    expect(body).toContain("| #1 | 100 | 50 | 150 |");
+    // single-row history => no totals row
+    expect(body).not.toContain("**합계**");
+    // hidden marker stores an array
+    expect(body).toMatch(/<!-- usage-data: \[\{"prompt":100,"completion":50\}\] -->/);
+
+    globalThis.fetch = originalFetch;
+  });
+
+  it("accumulates usage rows across PR updates when prior usage marker is present", async () => {
+    const previousBody = [
+      COMMENT_MARKER,
+      "## existing body",
+      "",
+      `<!-- usage-data: [{"prompt":100,"completion":50},{"prompt":200,"completion":80}] -->`,
+    ].join("\n");
+
+    const calls = [];
+    globalThis.fetch = (url, init = {}) => {
+      const u = String(url);
+      calls.push({ url: u, init });
+      if (u.includes(`/issues/${PR_NUMBER}/comments`) && (!init.method || init.method === "GET")) {
+        return ok([
+          {
+            id: 999,
+            user: { type: "Bot" },
+            body: previousBody,
+          },
+        ]);
+      }
+      if (init.method === "PATCH") return ok({});
+      return ok([]);
+    };
+
+    await upsertLearningStatusComment(
+      REPO_OWNER,
+      REPO_NAME,
+      PR_NUMBER,
+      `${COMMENT_MARKER}\n## new body`,
+      APP_TOKEN,
+      { prompt_tokens: 300, completion_tokens: 120 }
+    );
+
+    const patch = calls.find((c) => c.init.method === "PATCH");
+    expect(patch).toBeDefined();
+    expect(patch.url).toContain("/issues/comments/999");
+
+    const body = parseBody(patch);
+    // all three calls present, in order
+    expect(body).toContain("| #1 | 100 | 50 | 150 |");
+    expect(body).toContain("| #2 | 200 | 80 | 280 |");
+    expect(body).toContain("| #3 | 300 | 120 | 420 |");
+    // totals row appears once history.length > 1
+    expect(body).toContain("| **합계** | **600** | **250** | **850** |");
+    // marker is rewritten with the full array
+    expect(body).toMatch(
+      /<!-- usage-data: \[\{"prompt":100,"completion":50\},\{"prompt":200,"completion":80\},\{"prompt":300,"completion":120\}\] -->/
+    );
+
+    globalThis.fetch = originalFetch;
+  });
+
+  it("falls back to a single-row history when the prior marker is malformed", async () => {
+    // legacy / corrupt marker (object instead of array) — should not crash, just reset
+    const previousBody = [
+      COMMENT_MARKER,
+      "## existing body",
+      "",
+      `<!-- usage-data: {"prompt":100,"completion":50} -->`,
+    ].join("\n");
+
+    const calls = [];
+    globalThis.fetch = (url, init = {}) => {
+      const u = String(url);
+      calls.push({ url: u, init });
+      if (u.includes(`/issues/${PR_NUMBER}/comments`) && (!init.method || init.method === "GET")) {
+        return ok([
+          { id: 1, user: { type: "Bot" }, body: previousBody },
+        ]);
+      }
+      if (init.method === "PATCH") return ok({});
+      return ok([]);
+    };
+
+    await upsertLearningStatusComment(
+      REPO_OWNER,
+      REPO_NAME,
+      PR_NUMBER,
+      `${COMMENT_MARKER}\n## new body`,
+      APP_TOKEN,
+      { prompt_tokens: 999, completion_tokens: 111 }
+    );
+
+    const patch = calls.find((c) => c.init.method === "PATCH");
+    const body = parseBody(patch);
+
+    expect(body).toContain("| #1 | 999 | 111 | 1,110 |");
+    expect(body).not.toContain("**합계**");
+
+    globalThis.fetch = originalFetch;
+  });
+
+  it("omits the usage section entirely when no usage is provided", async () => {
+    const calls = [];
+    globalThis.fetch = (url, init = {}) => {
+      calls.push({ url: String(url), init });
+      if (init.method === "POST") return ok({});
+      return ok([]);
+    };
+
+    await upsertLearningStatusComment(
+      REPO_OWNER,
+      REPO_NAME,
+      PR_NUMBER,
+      `${COMMENT_MARKER}\n## body`,
+      APP_TOKEN
+      // no usage
+    );
+
+    const post = calls.find((c) => c.init.method === "POST");
+    const body = parseBody(post);
+    expect(body).not.toContain("🔢 API 사용량");
+    expect(body).not.toContain("usage-data:");
+
+    globalThis.fetch = originalFetch;
+  });
+});
diff --git a/utils/learningComment.js b/utils/learningComment.js
index 28b34f1..4dfad34 100644
--- a/utils/learningComment.js
+++ b/utils/learningComment.js
@@ -10,10 +10,15 @@ import { getGitHubHeaders } from "./github.js";
 const COMMENT_MARKER = "<!-- dalestudy-learning-status -->";
 
 /**
- * Hidden marker for embedding cumulative usage data in the comment.
- * Format: <!-- usage-data: {"prompt":N,"completion":N,"requests":N} -->
+ * Hidden marker for embedding per-request usage history in the comment.
+ * Format: <!-- usage-data: [{"prompt":N,"completion":N}, ...] -->
+ *
+ * The capture group must match the array — earlier versions of this regex
+ * matched only `{...}` and silently captured the first object inside the
+ * array, which made `parseUsageFromComment` always return `[]` and broke
+ * cumulative aggregation across PR updates.
  */
-const USAGE_DATA_RE = /<!-- usage-data: ({.*?}) -->/;
+const USAGE_DATA_RE = /<!-- usage-data: (\[.*?\]) -->/;
 
 /** gpt-4.1-nano pricing (USD per token) */
 const INPUT_COST_PER_TOKEN = 0.10 / 1_000_000;