diff --git a/4-Infrastructure/AGENTS.md b/4-Infrastructure/AGENTS.md
index b82381e0..b98946bb 100644
--- a/4-Infrastructure/AGENTS.md
+++ b/4-Infrastructure/AGENTS.md
@@ -313,7 +313,7 @@ python3 4-Infrastructure/storage/storage_agent.py --loop --interval 900
 
 ## Current Stack-Solidification Anchors
 
-- `4-Infrastructure/shim/mcp_drive_gccl_compress.py` — MCP-only Google Drive folder compressor: downloads a Drive folder via `@piotr-agier/google-drive-mcp`, GCCL-delta-compresses contents into a `.tar.zst` archive with JSON manifest, uploads archive + manifest back to Drive, optionally trashes the original. No rclone.
+- `4-Infrastructure/shim/mcp_drive_gccl_compress.py` — MCP-only Google Drive folder compressor: downloads a Drive folder via `@piotr-agier/google-drive-mcp`, GCCL-delta-compresses contents into a `.tar.zst` archive with JSON manifest, uploads archive + manifest back to Drive, optionally trashes the original. Optimized to skip existing local files and stream large files (>100MB) directly to prevent OOM. No rclone.
 - `4-Infrastructure/shim/rrc_arxiv_kernel_refine.py` — RRC arXiv kernel refinement: title+abstract keyword search against arxiv_papers for unmatched equations
 - `4-Infrastructure/infra/lean_lsp_mcp_wrapper.py` — Python wrapper for `lean-lsp-mcp` to fix the schema of `lean_diagnostic_messages` for compatibility with strict validators like Moonshot/Kimi API.
 - `4-Infrastructure/shim/stack_solidification_audit.py`
diff --git a/4-Infrastructure/shim/mcp_drive_gccl_compress.py b/4-Infrastructure/shim/mcp_drive_gccl_compress.py
index 19e5a6fc..78fe10e8 100644
--- a/4-Infrastructure/shim/mcp_drive_gccl_compress.py
+++ b/4-Infrastructure/shim/mcp_drive_gccl_compress.py
@@ -205,8 +205,22 @@ def gccl_compress_file(
     src_path: Path,
     reference_path: Optional[Path] = None,
     threshold_q16: int = 32768,
+    max_gccl_size: int = 100 * 1024 * 1024,
 ) -> Tuple[bytes, Dict[str, Any]]:
-    """Compress a single file with GCCL. Optionally use reference for delta."""
+    """Compress a single file with GCCL. Optionally use reference for delta.
+    Files larger than max_gccl_size are stored raw to avoid memory issues."""
+    original_size = src_path.stat().st_size
+
+    if original_size > max_gccl_size:
+        return b"", {
+            "mode": "raw-oversize",
+            "reference": str(reference_path) if reference_path else None,
+            "original_size": original_size,
+            "compressed_size": original_size,
+            "ratio": 1.0,
+            "decision": "OVERSIZE",
+        }
+
     data = src_path.read_bytes()
     reference = reference_path.read_bytes() if reference_path else None
 
@@ -265,8 +279,11 @@ def _download_folder_recursive(
         if item["type"] == "folder":
             _download_folder_recursive(client, item["id"], local_path, files, rename_map)
         else:
-            print(f"  downloading {safe_name} ...")
-            client.download_file(item["id"], local_path)
+            if local_path.exists():
+                print(f"  [skip] {safe_name} already exists locally")
+            else:
+                print(f"  downloading {safe_name} ...")
+                client.download_file(item["id"], local_path)
             files.append({
                 "drive_id": item["id"],
                 "name": safe_name,
@@ -299,6 +316,7 @@ def compress_folder(
     local_root: Path,
     archive_path: Path,
     threshold_q16: int = 32768,
+    max_gccl_size: int = 100 * 1024 * 1024,
 ) -> Dict[str, Any]:
     """GCCL-compress downloaded files and pack into a .tar.zst archive."""
     refs = build_reference_index(files, local_root)
@@ -316,13 +334,16 @@ def compress_folder(
             rel = src.relative_to(local_root)
             ref = refs.get(f["local_path"])
 
-            compressed_bytes, meta = gccl_compress_file(src, ref, threshold_q16)
+            compressed_bytes, meta = gccl_compress_file(src, ref, threshold_q16, max_gccl_size)
 
-            arc_name = f"{rel}.gccl" if meta["mode"] != "raw" else str(rel)
-            info = tarfile.TarInfo(name=arc_name)
-            info.size = len(compressed_bytes)
-            info.mtime = time.time()
-            tar.addfile(info, io.BytesIO(compressed_bytes))
+            arc_name = f"{rel}.gccl" if meta["mode"] not in ("raw", "raw-oversize") else str(rel)
+            if meta["mode"] in ("raw", "raw-oversize"):
+                tar.add(src, arcname=arc_name)
+            else:
+                info = tarfile.TarInfo(name=arc_name)
+                info.size = len(compressed_bytes)
+                info.mtime = time.time()
+                tar.addfile(info, io.BytesIO(compressed_bytes))
 
             manifest["files"].append({
                 "path": str(rel),
@@ -367,6 +388,7 @@ def main() -> int:
     parser.add_argument("--upload-parent", help="Drive folder ID to upload archive into")
     parser.add_argument("--trash-original", action="store_true", help="Move original folder to trash after upload")
     parser.add_argument("--threshold-q16", type=int, default=32768, help="GCCL threshold in Q16.16 (default 32768 = 0.5)")
+    parser.add_argument("--max-gccl-size", type=int, default=100*1024*1024, help="Max bytes for GCCL compression; larger files stored raw (default 100 MiB)")
     parser.add_argument("--dry-run", action="store_true", help="Download and compress but do not upload or trash")
     args = parser.parse_args()
 
@@ -393,7 +415,7 @@ def main() -> int:
             return 1
 
         print(f"[*] GCCL-compressing into {archive_path}")
-        manifest = compress_folder(files, local_clone, archive_path, args.threshold_q16)
+        manifest = compress_folder(files, local_clone, archive_path, args.threshold_q16, args.max_gccl_size)
         print(f"[+] Archive: {archive_path} ({archive_path.stat().st_size} bytes)")
         print(f"[+] Manifest: {manifest_path}")
         print(f"[+] SI compression ratio: {manifest['summary']['si_compression_ratio']:.2f}")