feat(file-sync): sync remote changes back to host on teardown

Add sync_back() to FileSyncManager — on sandbox cleanup, downloads
the remote .hermes/ directory as a tar archive, diffs against SHA-256
hashes of what was originally pushed, and applies only changed files.

- SHA-256 content hashing on push for accurate change detection
- Retry with exponential backoff (3 attempts, 2s/4s/8s)
- SIGINT deferred during sync-back to prevent partial writes
- fcntl.flock serialization for concurrent gateway sandboxes
- Last-write-wins conflict resolution with logged warnings
- New files created on remote are pulled back via path inference
- Backend implementations: SSH (tar cf over pipe), Modal (exec tar
  cf, read stdout), Daytona (exec tar cf, SDK download_file)
- Wired into cleanup() for all three backends (runs before
  ControlMaster close / sandbox terminate / sandbox stop)

28 new tests (10 FSM core + 18 backend-specific), 72 total passing.
This commit is contained in:
alt-glitch 2026-04-11 17:32:27 -07:00
parent 27eeea0555
commit 37c478cf2f
6 changed files with 1023 additions and 0 deletions

View file

@ -22,6 +22,7 @@ from tools.environments.base import (
_save_json_store,
)
from tools.environments.file_sync import (
BulkDownloadFn,
FileSyncManager,
iter_sync_files,
quoted_mkdir_command,
@ -269,6 +270,7 @@ class ModalEnvironment(BaseEnvironment):
upload_fn=self._modal_upload,
delete_fn=self._modal_delete,
bulk_upload_fn=self._modal_bulk_upload,
bulk_download_fn=self._modal_bulk_download,
)
self._sync_manager.sync(force=True)
self.init_session()
@ -347,6 +349,23 @@ class ModalEnvironment(BaseEnvironment):
self._worker.run_coroutine(_bulk(), timeout=120)
def _modal_bulk_download(self, dest: Path) -> None:
"""Download remote .hermes/ as a tar archive."""
async def _download():
proc = await self._sandbox.exec.aio(
"bash", "-c", "tar cf - -C /root/.hermes ."
)
data = await proc.stdout.read.aio()
exit_code = await proc.wait.aio()
if exit_code != 0:
raise RuntimeError(f"Modal bulk download failed (exit {exit_code})")
return data
tar_bytes = self._worker.run_coroutine(_download(), timeout=120)
if isinstance(tar_bytes, str):
tar_bytes = tar_bytes.encode()
dest.write_bytes(tar_bytes)
def _modal_delete(self, remote_paths: list[str]) -> None:
"""Batch-delete remote files via exec."""
rm_cmd = quoted_rm_command(remote_paths)
@ -404,6 +423,10 @@ class ModalEnvironment(BaseEnvironment):
if self._sandbox is None:
return
if self._sync_manager:
logger.info("Modal: syncing files from sandbox...")
self._sync_manager.sync_back()
if self._persistent:
try:
async def _snapshot():