From b24e453482c5d9c59dee264db357b1f0ae3ca04e Mon Sep 17 00:00:00 2001
From: ShiroKSH <kushidashiro@gmail.com>
Date: Sat, 11 Jul 2026 17:17:15 +0300
Subject: [PATCH] fix(sandbox): prevent E2B sync marker race

---
 lib/e2b_helper.py                | 16 +++++----
 tests/unit/test_sandbox_e2b.bats | 60 ++++++++++++++++++++++++++++++++
 2 files changed, 70 insertions(+), 6 deletions(-)

diff --git a/lib/e2b_helper.py b/lib/e2b_helper.py
index 2eead5a8..54486110 100644
--- a/lib/e2b_helper.py
+++ b/lib/e2b_helper.py
@@ -185,21 +185,23 @@ def cmd_download(args):
     _require_sdk()
     sandbox = _connect(args.sandbox_id)
     marker = _sync_marker(args.src)
+    pending_marker = marker + ".pending"
     # Besides the changed files, every download carries a manifest of ALL
     # current workspace files (minus .git) so the host can delete files that
     # disappeared in the sandbox. The sandbox is always Linux/GNU.
-    # The sync marker is NOT advanced here: the host calls ack-download after
-    # it has successfully extracted the archive, so a failure anywhere in
-    # between leaves the changes re-downloadable (at-least-once delivery;
-    # re-extraction is an idempotent overwrite).
+    # Snapshot the scan start before enumerating files, but do not advance the
+    # committed marker until the host acknowledges a successful extraction.
+    # Files written while the archive is being built stay newer than this
+    # pending marker and are therefore included by the next download.
     script = (
-        "cd {src} && "
+        "touch {pending_marker} && cd {src} && "
         "if [ -f {marker} ]; then find . -type f -newer {marker} -print0; else : ; fi > {changed} && "
         "find . -type f ! -path './.git/*' > {manifest_staging} && "
         "tar -czf {staging} --null -T {changed} -C {manifest_dir} {manifest}"
     ).format(
         src=shlex.quote(args.src),
         marker=shlex.quote(marker),
+        pending_marker=shlex.quote(pending_marker),
         changed=shlex.quote(CHANGED_LIST),
         staging=shlex.quote(DOWNLOAD_STAGING),
         manifest_staging=shlex.quote(MANIFEST_STAGING),
@@ -222,8 +224,10 @@ def cmd_ack_download(args):
     _require_sdk()
     sandbox = _connect(args.sandbox_id)
     marker = _sync_marker(args.src)
+    pending_marker = marker + ".pending"
     try:
-        sandbox.commands.run("touch %s && rm -f %s %s %s" % (
+        sandbox.commands.run("if [ -f %s ]; then mv -f %s %s; fi && rm -f %s %s %s" % (
+            shlex.quote(pending_marker), shlex.quote(pending_marker),
             shlex.quote(marker), shlex.quote(DOWNLOAD_STAGING),
             shlex.quote(CHANGED_LIST), shlex.quote(MANIFEST_STAGING)))
     except Exception as exc:
diff --git a/tests/unit/test_sandbox_e2b.bats b/tests/unit/test_sandbox_e2b.bats
index 5e6198a1..acef9a8a 100644
--- a/tests/unit/test_sandbox_e2b.bats
+++ b/tests/unit/test_sandbox_e2b.bats
@@ -654,6 +654,66 @@ EOF
     assert_failure
 }
 
+@test "e2b_helper.py: ack preserves files written after the download snapshot" {
+    local python=python3
+    command -v "$python" > /dev/null || python=python
+    command -v "$python" > /dev/null || skip "python not available"
+    run "$python" - "$PROJECT_ROOT/lib/e2b_helper.py" "$TEST_DIR/remote" <<'PY'
+import importlib.util
+import io
+import shlex
+import subprocess
+import sys
+import tarfile
+from types import SimpleNamespace
+
+helper_path, remote_root = sys.argv[1:]
+spec = importlib.util.spec_from_file_location("e2b_helper", helper_path)
+helper = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(helper)
+
+def bash(command):
+    return subprocess.run(
+        ["bash", "-c", command], check=True, stdout=subprocess.PIPE
+    ).stdout
+
+workspace = remote_root + "/workspace"
+marker = remote_root + "/.ralph_sync_marker"
+bash("mkdir -p {0} && touch {1} && sleep 0.05 && printf early > {0}/early.txt".format(
+    shlex.quote(workspace), shlex.quote(marker)
+))
+
+sandbox = SimpleNamespace(
+    commands=SimpleNamespace(run=lambda command: bash(command)),
+    files=SimpleNamespace(
+        read=lambda path, format=None: bash("cat " + shlex.quote(path))
+    ),
+)
+helper._require_sdk = lambda: None
+helper._connect = lambda _sandbox_id: sandbox
+helper._emit = lambda _value: None
+args = SimpleNamespace(sandbox_id="sandbox", src=workspace)
+
+def download_names():
+    captured = SimpleNamespace(buffer=io.BytesIO())
+    original_stdout = sys.stdout
+    sys.stdout = captured
+    try:
+        helper.cmd_download(args)
+    finally:
+        sys.stdout = original_stdout
+    with tarfile.open(fileobj=io.BytesIO(captured.buffer.getvalue()), mode="r:gz") as archive:
+        return {name[2:] if name.startswith("./") else name for name in archive.getnames()}
+
+assert "early.txt" in download_names()
+bash("sleep 0.05 && printf raced > " + shlex.quote(workspace + "/raced.txt"))
+helper.cmd_ack_download(args)
+helper.cmd_ack_download(args)
+assert "raced.txt" in download_names()
+PY
+    assert_success
+}
+
 @test "upload_project_to_e2b: initializes the synced-files state from the upload list" {
     echo "content" > "$TEST_DIR/tracked.txt"
     _started_sandbox