Avoid renderer OOM when adding large PDFs to the workbench

frontend/editor/src/core/contexts/file/fileActions.ts

@@ -13,7 +13,7 @@ import {
   ProcessedFileMetadata,
 } from "@app/types/fileContext";
 import { FileId, ToolOperation } from "@app/types/file";
-import { generateThumbnailWithMetadata } from "@app/utils/thumbnailUtils";
+import { generateThumbnailPairWithMetadata } from "@app/utils/thumbnailUtils";
 import { FileLifecycleManager } from "@app/contexts/file/lifecycle";
 import { buildQuickKeySet } from "@app/contexts/file/fileSelectors";
 import { StirlingFile } from "@app/types/fileContext";
@@ -125,11 +125,20 @@ export async function generateProcessedFileMetadata(
   }
 
   try {
-    // Generate unrotated thumbnails for PageEditor (rotation applied via CSS)
-    const unrotatedResult = await generateThumbnailWithMetadata(file, false);
+    // One parse produces both variants: unrotated thumbnails for PageEditor
+    // (rotation applied via CSS) and the rotated one for file manager display.
+    const { unrotated: unrotatedResult, rotated: rotatedResult } =
+      await generateThumbnailPairWithMetadata(file);
 
-    // Generate rotated thumbnail for file manager display
-    const rotatedResult = await generateThumbnailWithMetadata(file, true);
+    // Large PDF whose linearized-prefix attempt failed: report "no metadata"
+    // (the tolerated failure shape) rather than a bogus zero-page document.
+    if (
+      !unrotatedResult.thumbnail &&
+      unrotatedResult.pageCount === 0 &&
+      !unrotatedResult.isEncrypted
+    ) {
+      return undefined;
+    }
 
     const processedFile = createProcessedFile(
       unrotatedResult.pageCount,

frontend/editor/src/core/services/fileStorage.ts

@@ -23,7 +23,9 @@ import {
 const THUMBNAIL_TTL_MS = 30 * 24 * 60 * 60 * 1000; // 30 days
 
 export interface StoredStirlingFileRecord extends BaseFileMetadata {
-  data: ArrayBuffer;
+  // Blob since the large-file OOM fix (stored by reference, no JS-side copy);
+  // ArrayBuffer records predate it and are still readable.
+  data: ArrayBuffer | Blob;
   fileId: FileId; // Matches runtime StirlingFile.fileId exactly
   quickKey: string; // Matches runtime StirlingFile.quickKey exactly
   thumbnail?: string;
@@ -118,7 +120,6 @@ class FileStorageService {
     stub: StirlingFileStub,
   ): Promise<void> {
     const db = await this.getDatabase();
-    const arrayBuffer = await stirlingFile.arrayBuffer();
 
     const record: StoredStirlingFileRecord = {
       id: stirlingFile.fileId,
@@ -129,7 +130,9 @@ class FileStorageService {
       size: stirlingFile.size,
       lastModified: stirlingFile.lastModified,
       createdAt: stub.createdAt,
-      data: arrayBuffer,
+      // Store the File (a Blob) itself: IndexedDB persists it by reference and
+      // streams to disk, so multi-GB files never materialize in JS memory.
+      data: stirlingFile,
       thumbnail: stub.thumbnailUrl,
       thumbnailStoredAt: stub.thumbnailUrl ? Date.now() : undefined,
       isLeaf: stub.isLeaf ?? true,

frontend/editor/src/core/utils/thumbnailUtils.ts

@@ -31,6 +31,14 @@ export function calculateScaleFromFileSize(fileSize: number): number {
 /** PDFium error code 4 = password required (encrypted PDF). */
 const PDFIUM_ERR_PASSWORD = 4;
 
+/** PDFs at or above this size never get a full-buffer client-side parse
+ * (renderer OOM) - only the linearized-prefix attempt below. */
+export const LARGE_PDF_PARSE_LIMIT = 100 * 1024 * 1024;
+
+/** Linearized PDFs keep page 1 + hint tables in the first bytes, so a small
+ * prefix is often enough to render a thumbnail without reading the file. */
+const LINEARIZED_PREFIX_BYTES = 2 * 1024 * 1024;
+
 interface PdfiumRenderResult {
   thumbnail: string;
   pageCount: number;
@@ -112,6 +120,72 @@ async function renderPdfThumbnailPdfium(
   }
 }
 
+/**
+ * Render both thumbnail variants (upright + rotation-baked) from a single
+ * document open - halves the parse and memory cost of the add-files path.
+ */
+async function renderPdfThumbnailPairPdfium(
+  data: ArrayBuffer,
+  scale: number,
+  collectAllPagesMetadata: boolean,
+): Promise<{ unrotated: PdfiumRenderResult; rotated: PdfiumRenderResult }> {
+  const m = await getPdfiumModule();
+  let docPtr: number;
+  try {
+    docPtr = await openRawDocumentSafe(data);
+  } catch (error) {
+    if (
+      error instanceof Error &&
+      new RegExp(`error ${PDFIUM_ERR_PASSWORD}`).test(error.message)
+    ) {
+      const encrypted: PdfiumRenderResult = {
+        thumbnail: "",
+        pageCount: 1,
+        pageRotations: [],
+        pageDimensions: [],
+        isEncrypted: true,
+      };
+      return { unrotated: encrypted, rotated: { ...encrypted } };
+    }
+    throw error;
+  }
+
+  try {
+    const pageCount = m.FPDF_GetPageCount(docPtr);
+    const unrotatedThumb = await renderPdfiumPageDataUrl(docPtr, 0, scale, {
+      applyRotation: false,
+    });
+    const rotatedThumb = await renderPdfiumPageDataUrl(docPtr, 0, scale, {
+      applyRotation: true,
+    });
+    if (!unrotatedThumb || !rotatedThumb) {
+      throw new Error("PDFium: failed to render page 0");
+    }
+
+    const firstMeta = await readPdfiumPageMetadata(docPtr, 0);
+    const pageRotations: number[] = [firstMeta?.rotation ?? 0];
+    const pageDimensions: Array<{ width: number; height: number }> = [
+      { width: firstMeta?.width ?? 0, height: firstMeta?.height ?? 0 },
+    ];
+    if (collectAllPagesMetadata) {
+      for (let i = 1; i < pageCount; i++) {
+        const meta = await readPdfiumPageMetadata(docPtr, i);
+        if (!meta) continue;
+        pageRotations[i] = meta.rotation;
+        pageDimensions[i] = { width: meta.width, height: meta.height };
+      }
+    }
+
+    const base = { pageCount, pageRotations, pageDimensions };
+    return {
+      unrotated: { thumbnail: unrotatedThumb, ...base },
+      rotated: { thumbnail: rotatedThumb, ...base },
+    };
+  } finally {
+    await closeRawDocument(docPtr);
+  }
+}
+
 async function generatePDFThumbnail(
   arrayBuffer: ArrayBuffer,
   scale: number,
@@ -133,7 +207,7 @@ async function generatePDFThumbnail(
  */
 export async function generateThumbnailForFile(file: File): Promise<string> {
   // Very large PDFs skip thumbnail generation — SVG icon shown in UI instead
-  if (file.size >= 100 * 1024 * 1024) {
+  if (file.size >= LARGE_PDF_PARSE_LIMIT) {
     return "";
   }
 
@@ -152,8 +226,7 @@ export async function generateThumbnailForFile(file: File): Promise<string> {
     const scale = calculateScaleFromFileSize(file.size);
 
     // Only read first 2MB for thumbnail generation to save memory
-    const chunkSize = 2 * 1024 * 1024; // 2MB
-    const chunk = file.slice(0, Math.min(chunkSize, file.size));
+    const chunk = file.slice(0, Math.min(LINEARIZED_PREFIX_BYTES, file.size));
     const arrayBuffer = await chunk.arrayBuffer();
 
     try {
@@ -193,6 +266,31 @@ export async function generateThumbnailWithMetadata(
 
   const scale = calculateScaleFromFileSize(file.size);
 
+  // Never full-parse huge PDFs client-side - the renderer process OOMs long
+  // before system RAM runs out. The prefix succeeds for linearized PDFs.
+  if (file.size >= LARGE_PDF_PARSE_LIMIT) {
+    try {
+      const chunk = await file.slice(0, LINEARIZED_PREFIX_BYTES).arrayBuffer();
+      const result = await renderPdfThumbnailPdfium(
+        chunk,
+        scale,
+        applyRotation,
+        false,
+      );
+      if (result.isEncrypted) {
+        return { thumbnail: "", pageCount: 1, isEncrypted: true };
+      }
+      return {
+        thumbnail: result.thumbnail,
+        pageCount: result.pageCount,
+        pageRotations: result.pageRotations,
+        pageDimensions: result.pageDimensions,
+      };
+    } catch {
+      return { thumbnail: "", pageCount: 0 };
+    }
+  }
+
   try {
     const arrayBuffer = await file.arrayBuffer();
     // Always read per-page rotation: PageEditor renders thumbnails upright and
@@ -222,3 +320,41 @@ export async function generateThumbnailWithMetadata(
     return { thumbnail: "", pageCount: 1 };
   }
 }
+
+/**
+ * Both thumbnail variants + page metadata from ONE full parse instead of two.
+ * Large PDFs only get the linearized-prefix attempt; if that fails, both
+ * variants are empty placeholders and page metadata is omitted.
+ */
+export async function generateThumbnailPairWithMetadata(file: File): Promise<{
+  unrotated: ThumbnailWithMetadata;
+  rotated: ThumbnailWithMetadata;
+}> {
+  const scale = calculateScaleFromFileSize(file.size);
+  try {
+    const isLarge = file.size >= LARGE_PDF_PARSE_LIMIT;
+    const buffer = isLarge
+      ? await file.slice(0, LINEARIZED_PREFIX_BYTES).arrayBuffer()
+      : await file.arrayBuffer();
+    const pair = await renderPdfThumbnailPairPdfium(buffer, scale, !isLarge);
+
+    const toPublic = (r: PdfiumRenderResult): ThumbnailWithMetadata =>
+      r.isEncrypted
+        ? { thumbnail: "", pageCount: 1, isEncrypted: true }
+        : {
+            thumbnail: r.thumbnail,
+            pageCount: r.pageCount,
+            pageRotations: r.pageRotations,
+            pageDimensions: r.pageDimensions,
+          };
+    return {
+      unrotated: toPublic(pair.unrotated),
+      rotated: toPublic(pair.rotated),
+    };
+  } catch {
+    return {
+      unrotated: { thumbnail: "", pageCount: 0 },
+      rotated: { thumbnail: "", pageCount: 0 },
+    };
+  }
+}