fix(media-understanding): strip repeated placeholders (#96431)

lin-hongkuan · vincentkoc · web-flow · commit f57a30289d32 · 2026-06-25T01:25:08.000+08:00
Merged via squash. Prepared head SHA: 4b00486 Co-authored-by: lin-hongkuan <234943746+lin-hongkuan@users.noreply.github.com> Co-authored-by: vincentkoc <25068+vincentkoc@users.noreply.github.com> Reviewed-by: @vincentkoc
diff --git a/packages/media-understanding-common/src/format.test.ts b/packages/media-understanding-common/src/format.test.ts
@@ -48,6 +48,36 @@ describe("formatMediaUnderstandingBody", () => {
     expect(body).toBe("[Audio]\nUser text:\ncaption here\nTranscript:\ntranscribed");
   });
 
+  it("strips repeated leading media placeholders from user text", () => {
+    const body = formatMediaUnderstandingBody({
+      body: "<media:image> <media:audio> caption here",
+      outputs: [
+        {
+          kind: "audio.transcription",
+          attachmentIndex: 0,
+          text: "transcribed",
+          provider: "groq",
+        },
+      ],
+    });
+    expect(body).toBe("[Audio]\nUser text:\ncaption here\nTranscript:\ntranscribed");
+  });
+
+  it("treats repeated media placeholders without captions as synthetic text", () => {
+    const body = formatMediaUnderstandingBody({
+      body: "<media:image> <media:audio>",
+      outputs: [
+        {
+          kind: "image.description",
+          attachmentIndex: 0,
+          text: "a chart",
+          provider: "openai",
+        },
+      ],
+    });
+    expect(body).toBe("[Image]\nDescription:\na chart");
+  });
+
   it("keeps user text once when multiple outputs exist", () => {
     const body = formatMediaUnderstandingBody({
       body: "caption here",
diff --git a/packages/media-understanding-common/src/format.ts b/packages/media-understanding-common/src/format.ts
@@ -1,8 +1,9 @@
 // Media Understanding Common helper module supports format behavior.
 import type { MediaUnderstandingOutput } from "./types.js";
 
-const MEDIA_PLACEHOLDER_RE = /^<media:[^>]+>(\s*\([^)]*\))?$/i;
-const MEDIA_PLACEHOLDER_TOKEN_RE = /^<media:[^>]+>(\s*\([^)]*\))?\s*/i;
+const MEDIA_PLACEHOLDER_TOKEN = String.raw`<media:[^>]+>(?:\s*\([^)]*\))?`;
+const MEDIA_PLACEHOLDER_RE = new RegExp(String.raw`^(?:${MEDIA_PLACEHOLDER_TOKEN}\s*)+$`, "i");
+const MEDIA_PLACEHOLDER_TOKEN_RE = new RegExp(String.raw`^(?:${MEDIA_PLACEHOLDER_TOKEN}\s*)+`, "i");
 
 /** Extracts user-authored text while ignoring synthetic media placeholder tokens. */
 export function extractMediaUserText(body?: string): string | undefined {
