fix: stop duplicating text in Gemini Web streamed responses (#7163) (#7198)

changelog.d/fixes/7163-gemini-web-duplicated-text.md

@@ -0,0 +1 @@
+- fix(sse): stop duplicating text in Gemini Web streamed responses (#7163)

open-sse/executors/gemini-web.ts

@@ -112,12 +112,15 @@ function parseCookies(raw: string): Array<{ name: string; value: string }> {
  *   [["wrb.fr", null, "<JSON string>"]]
  *
  * The JSON string contains nested array: inner[4][0][1] = ["text chunks"].
- * We concatenate text from every wrb.fr line because Gemini can split one
- * assistant answer across multiple StreamGenerate chunks.
+ * Each wrb.fr line is a CUMULATIVE snapshot of the whole answer generated so
+ * far (not an independent delta), so we keep only the text from the LAST
+ * frame that yields non-empty text instead of concatenating every frame —
+ * concatenating would reproduce the same growing text with each snapshot
+ * (see #7163).
  */
 export function parseStreamResponse(raw: string): string {
   const lines = raw.split("\n");
-  const textChunks: string[] = [];
+  let lastText = "";
 
   for (const rawLine of lines) {
     const line = rawLine.trim();
@@ -133,12 +136,12 @@ export function parseStreamResponse(raw: string): string {
       const responseArray = inner?.[4]?.[0]?.[1];
       if (!Array.isArray(responseArray)) continue;
       const text = responseArray.filter((c: unknown) => typeof c === "string").join("");
-      if (text) textChunks.push(text);
+      if (text) lastText = text;
     } catch {
       // Skip unparseable lines
     }
   }
-  return textChunks.join("");
+  return lastText;
 }
 
 function readCredentialString(value: unknown): string {

tests/unit/gemini-web.test.ts

@@ -266,15 +266,20 @@ test("#2832: GeminiWebExecutor catch block sanitizes Playwright launch errors (i
 
 // ─── StreamGenerate parsing ─────────────────────────────────────────────────
 
-test("parseStreamResponse concatenates Gemini Web text from multiple wrb.fr chunks", () => {
+test("parseStreamResponse keeps only the final cumulative StreamGenerate snapshot (no duplication) — regression for #7163", () => {
   const makeChunk = (text: string) => {
     const inner = new Array(80).fill(null);
     inner[4] = [[null, [text]]];
     return `[["wrb.fr", null, ${JSON.stringify(JSON.stringify(inner))}]]`;
   };
 
-  const raw = `)]}'\n10\n${makeChunk("First ")}\n5\n${makeChunk("chunk")}`;
-  assert.equal(parseStreamResponse(raw), "First chunk");
+  // Gemini's StreamGenerate frames are CUMULATIVE snapshots: each later frame
+  // repeats the full answer generated so far, not just the new characters.
+  const frame1 = "Hello!";
+  const frame2 = "Hello! How can I";
+  const frame3 = "Hello! How can I help you out today?";
+  const raw = `)]}'\n10\n${makeChunk(frame1)}\n5\n${makeChunk(frame2)}\n5\n${makeChunk(frame3)}`;
+  assert.equal(parseStreamResponse(raw), frame3);
 });
 
 test("parseStreamResponse ignores wrb.fr lines whose first entry is not an array", () => {