fix: stop inferring file extraction from Body/RawBody mismatch (#46454)

src/auto-reply/reply/followup-media.ts

@@ -292,27 +292,21 @@ export async function applyDeferredMediaUnderstandingToQueuedRun(
     return;
   }
 
-  const referenceBody = mediaContext.RawBody ?? mediaContext.Body;
-  // Prefer RawBody-vs-Body comparison when RawBody exists. If RawBody is
-  // missing, any real <file>...</file> block plus file-like attachments means
-  // extraction already ran, even if the stored name came from Content-Disposition
-  // instead of the attachment path/url basename.
-  if (!mediaContext.DeferredFileBlocksExtracted && hasAnyFileAttachments(mediaContext)) {
-    const rawBodyMissing = typeof mediaContext.RawBody !== "string";
-    if (mediaContext.Body !== referenceBody) {
-      mediaContext.DeferredFileBlocksExtracted = true;
-    } else if (
-      rawBodyMissing &&
-      (Boolean(mediaContext.MediaUnderstanding?.length) ||
-        bodyContainsExtractedFileBlock(mediaContext.Body))
-    ) {
-      mediaContext.DeferredFileBlocksExtracted = true;
-    }
-  }
   if (mediaContext.MediaUnderstanding?.length) {
     mediaContext.DeferredMediaApplied = true;
     return;
   }
+  // Treat followup file extraction as already applied only when we have explicit
+  // evidence: the queue snapshot already flagged it or Body already contains a
+  // real extracted <file>...</file> block. Body/RawBody mismatches are not
+  // reliable because some channels wrap Body with envelope metadata.
+  if (
+    !mediaContext.DeferredFileBlocksExtracted &&
+    hasAnyFileAttachments(mediaContext) &&
+    bodyContainsExtractedFileBlock(mediaContext.Body)
+  ) {
+    mediaContext.DeferredFileBlocksExtracted = true;
+  }
 
   if (mediaContext.DeferredFileBlocksExtracted && hasOnlyFileLikeAttachments(mediaContext)) {
     mediaContext.DeferredMediaApplied = true;

src/auto-reply/reply/followup-runner.test.ts

@@ -1231,6 +1231,54 @@ describe("createFollowupRunner media understanding", () => {
     expect(agentCall?.prompt?.match(/<file\b/g)).toHaveLength(1);
   });
 
+  it("does not infer file extraction from wrapped Body/RawBody mismatches", async () => {
+    const fileBlock = '<file name="report.pdf" mime="application/pdf">\nreport content\n</file>';
+    applyMediaUnderstandingMock.mockImplementationOnce(
+      async (params: { ctx: Record<string, unknown> }) => {
+        params.ctx.Body = `summarize this\n\n${fileBlock}`;
+        return {
+          outputs: [],
+          decisions: [],
+          appliedImage: false,
+          appliedAudio: false,
+          appliedVideo: false,
+          appliedFile: true,
+        };
+      },
+    );
+    runEmbeddedPiAgentMock.mockResolvedValueOnce({
+      payloads: [{ text: "file processed" }],
+      meta: {},
+    });
+
+    const runner = createFollowupRunner({
+      opts: { onBlockReply: vi.fn(async () => {}) },
+      typing: createMockTypingController(),
+      typingMode: "instant",
+      defaultModel: "anthropic/claude-opus-4-5",
+    });
+
+    await runner(
+      createQueuedRun({
+        prompt: "[media attached: /tmp/report.pdf]\nLine: Alice\nsummarize this",
+        mediaContext: {
+          Body: "Line: Alice\nsummarize this",
+          RawBody: "summarize this",
+          MediaPaths: ["/tmp/report.pdf"],
+          MediaTypes: ["application/pdf"],
+        },
+      }),
+    );
+
+    expect(applyMediaUnderstandingMock).toHaveBeenCalledTimes(1);
+    const agentCall = runEmbeddedPiAgentMock.mock.calls.at(-1)?.[0] as {
+      prompt?: string;
+    };
+    expect(agentCall?.prompt).toContain("Line: Alice");
+    expect(agentCall?.prompt).toContain(fileBlock);
+    expect(agentCall?.prompt?.match(/<file\s+name="report\.pdf"/g)).toHaveLength(1);
+  });
+
   it("preserves non-audio media lines when only audio is transcribed", async () => {
     applyMediaUnderstandingMock.mockImplementationOnce(
       async (params: { ctx: Record<string, unknown> }) => {

src/auto-reply/reply/queue/types.ts

@@ -52,9 +52,9 @@ export type FollowupMediaContext = {
   DeferredMediaApplied?: boolean;
   /**
    * Set when file extraction has already been applied to Body (either in the
-   * primary path or by a previous deferred-media run).  Checked instead of
-   * scanning body text for `<file` patterns to avoid false-positives on user
-   * messages that contain literal XML-like text.
+   * primary path or by a previous deferred-media run). This avoids re-running
+   * file extraction when Body already contains real extracted `<file>...</file>`
+   * blocks.
    */
   DeferredFileBlocksExtracted?: boolean;
 };