openclaw_hf/backup.py

#!/usr/bin/env python3
"""Backup and restore OpenClaw state to/from Hugging Face Dataset."""

from __future__ import annotations

import argparse
import datetime as dt
import fcntl
import hashlib
import json
import os
import re
import shutil
import subprocess
import sys
import tarfile
import tempfile
import time
import urllib.request
import uuid
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any, Optional


def load_env_files():
    """Load environment variables from OpenClaw env files.

    Note: HF Space variables (already in os.environ) take priority over env files.
    This function only fills in missing variables from env files.
    """
    env_files = [
        "/etc/profile.d/openclaw-env.sh",
        "/root/.env.d/openclaw-backup.env",
    ]

    for env_file in env_files:
        if os.path.isfile(env_file):
            try:
                result = subprocess.run(
                    ["bash", "-c", f"source '{env_file}' 2>/dev/null && env"],
                    capture_output=True,
                    text=True,
                    timeout=10,
                )
                if result.returncode == 0:
                    for line in result.stdout.strip().split("\n"):
                        if "=" in line:
                            key, _, value = line.partition("=")
                            if key and key not in os.environ:
                                os.environ[key] = value
            except Exception:
                pass


def env_get_with_hf_priority(name: str, default: str = "") -> str:
    """Get environment variable with HF Space variable taking priority.

    Priority:
    1. Existing os.environ value (HF Space variable)
    2. /etc/profile.d/openclaw-env.sh
    3. /root/.env.d/openclaw-backup.env
    4. default value

    Args:
        name: Environment variable name
        default: Default value if not found anywhere

    Returns:
        The environment variable value
    """
    if name in os.environ and os.environ[name]:
        return os.environ[name]

    env_files = [
        "/etc/profile.d/openclaw-env.sh",
        "/root/.env.d/openclaw-backup.env",
    ]

    for env_file in env_files:
        if os.path.isfile(env_file):
            try:
                result = subprocess.run(
                    ["bash", "-c", f"source '{env_file}' 2>/dev/null && echo ${name}"],
                    capture_output=True,
                    text=True,
                    timeout=10,
                )
                if result.returncode == 0:
                    value = result.stdout.strip()
                    if value:
                        return value
            except Exception:
                pass

    return default

# Set Hugging Face Hub timeout before importing (in seconds)
# This helps avoid timeout errors when downloading large backup files
os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "300")  # 5 minutes
os.environ.setdefault("HF_HUB_ETAG_TIMEOUT", "60")  # 60 seconds timeout for metadata

try:
    from huggingface_hub import HfApi, hf_hub_download
    from huggingface_hub.utils import HfHubHTTPError
except ModuleNotFoundError:  # pragma: no cover - runtime dependency, optional for unit tests
    HfApi = None  # type: ignore[assignment]
    hf_hub_download = None  # type: ignore[assignment]

    class HfHubHTTPError(Exception):  # type: ignore[no-redef]
        """Fallback error when huggingface_hub is not installed."""


def env_bool(name: str, default: bool) -> bool:
    raw = os.getenv(name)
    if raw is None:
        return default
    return raw.strip().lower() in {"1", "true", "yes", "on"}


def utc_now() -> dt.datetime:
    return dt.datetime.now(dt.UTC)


# 元数据版本常量
METADATA_VERSION = "2.1"  # 当前元数据版本
MIN_SUPPORTED_METADATA_VERSION = "1.0"  # 最低支持的元数据版本


def parse_metadata_version(version_str: str) -> tuple[int, int]:
    """解析元数据版本字符串为版本号元组。"""
    try:
        parts = version_str.split(".")
        major = int(parts[0])
        minor = int(parts[1]) if len(parts) > 1 else 0
        return (major, minor)
    except (ValueError, IndexError):
        return (1, 0)  # 默认返回1.0版本


def is_metadata_compatible(version_str: str) -> bool:
    """检查元数据版本是否兼容当前版本。

    兼容性规则：
    - 目标主版本 <= 当前主版本：向后兼容（旧备份可以用新版本恢复）
    - 目标主版本 > 当前主版本：不兼容（新备份不能用旧版本恢复，需升级）
    - 次版本在支持范围内
    """
    current = parse_metadata_version(METADATA_VERSION)
    target = parse_metadata_version(version_str)
    min_supported = parse_metadata_version(MIN_SUPPORTED_METADATA_VERSION)

    # 目标主版本大于当前主版本：不兼容（需要升级软件）
    if target[0] > current[0]:
        return False

    # 目标主版本小于当前主版本：向后兼容（旧备份可用新版本恢复）
    # 但仍需检查次版本是否在支持范围内
    if target[0] < current[0]:
        return min_supported[1] <= target[1]

    # 主版本相同：次版本必须在支持范围内
    return min_supported[1] <= target[1] <= current[1]


@dataclass
class BackupConfig:
    dataset_repo: str
    state_dir: Path
    backup_source_dir: Path
    work_dir: Path
    metadata_dir: Path  # Directory for backup metadata (last-backup-metadata.json)
    repo_type: str
    path_prefix: str
    private: bool
    # Multi-dataset configuration for flexible restore/backup scenarios
    restore_dataset_repo: Optional[str] = None  # Dataset to restore from (defaults to dataset_repo)
    backup_npm_enabled: bool = True  # Whether to backup npm directory
    restore_npm_enabled: bool = True  # Whether to restore npm directory
    # Note: Container restart/rebuild always triggers restore. This is by design.
    incremental: bool = True  # 默认开启增量备份
    incremental_interval_minutes: int = 10  # 增量备份间隔（分钟）
    compression_level: int = 6
    split_size: str = ""  # e.g., "500M" for 500MB volumes
    size_warning_mb: int = 1500  # Warn if backup exceeds this size
    # Dynamic backup strategy settings
    dynamic_backup: bool = True  # 默认开启动态备份策略
    dynamic_small_threshold_mb: int = 500
    dynamic_medium_threshold_mb: int = 2000
    dynamic_high_change_rate: int = 10
    dynamic_low_change_rate: int = 2
    dynamic_min_changed_files: int = 5  # Minimum changed files to consider backup worthwhile
    dynamic_min_changed_size_kb: int = 100  # Minimum total size of changed files (KB)
    # Full backup strategy settings
    full_backup_interval_hours: int = 2  # Force full backup after this many hours
    max_incremental_backups: int = 10  # Force full backup after this many incremental backups
    root_config_dir: Path = Path("/root/.config").resolve()
    root_codex_dir: Path = Path("/root/.codex").resolve()
    root_claude_dir: Path = Path("/root/.claude").resolve()
    root_agents_dir: Path = Path("/root/.agents").resolve()
    root_ssh_dir: Path = Path("/root/.ssh").resolve()
    root_env_dir: Path = Path("/root/.env.d").resolve()
    root_npm_dir: Path = Path("/root/.npm").resolve()
    root_lark_cli_dir: Path = Path("/root/.lark-cli").resolve()
    extra_dirs: dict[str, Path] = field(default_factory=dict)
    extra_files: dict[str, Path] = field(default_factory=dict)
    keep_count: int = 48
    # Local backup retention settings
    keep_local_backup: bool = False  # Whether to keep a local copy after upload
    local_backup_dir: Optional[Path] = None  # Directory to store local backups (None = use work_dir)
    keep_local_count: int = 5  # Number of local backups to keep
    # Encryption settings
    encryption_enabled: bool = False  # Whether to encrypt backups
    encryption_password: Optional[str] = None  # Password for encryption (from HF secret)

    @classmethod
    def from_env(cls) -> "BackupConfig":
        # Load environment from files first (fills in missing vars only)
        load_env_files()

        # HF Space variables take priority over env files
        dataset_repo = env_get_with_hf_priority("OPENCLAW_BACKUP_DATASET_REPO")
        if not dataset_repo:
            raise ValueError("OPENCLAW_BACKUP_DATASET_REPO is not set")

        # Multi-dataset configuration
        # HF Space variable takes priority, fallback to dataset_repo
        restore_dataset_repo = env_get_with_hf_priority("OPENCLAW_RESTORE_DATASET_REPO", dataset_repo)
        backup_npm_enabled = env_get_with_hf_priority("OPENCLAW_BACKUP_NPM_ENABLED", "true").lower() in ("true", "1", "yes")
        restore_npm_enabled = env_get_with_hf_priority("OPENCLAW_RESTORE_NPM_ENABLED", "true").lower() in ("true", "1", "yes")
        # Note: Container restart/rebuild always triggers restore. This is by design.

        # Encryption configuration
        encryption_enabled = env_bool("OPENCLAW_BACKUP_ENCRYPTION_ENABLED", False)
        encryption_password = os.getenv("OPENCLAW_BACKUP_ENCRYPTION_PASSWORD") or None

        state_dir = Path(os.getenv("OPENCLAW_STATE_DIR", "/root/.openclaw")).resolve()
        backup_source_raw = (os.getenv("OPENCLAW_BACKUP_SOURCE_DIR") or "").strip()
        if backup_source_raw:
            backup_source_dir = Path(backup_source_raw).resolve()
        else:
            backup_source_dir = state_dir
        root_config_dir = Path(
            (os.getenv("OPENCLAW_BACKUP_ROOT_CONFIG_DIR") or "/root/.config").strip() or "/root/.config"
        ).resolve()
        root_codex_dir = Path(
            (os.getenv("OPENCLAW_BACKUP_ROOT_CODEX_DIR") or "/root/.codex").strip() or "/root/.codex"
        ).resolve()
        root_claude_dir = Path(
            (os.getenv("OPENCLAW_BACKUP_ROOT_CLAUDE_DIR") or "/root/.claude").strip() or "/root/.claude"
        ).resolve()
        root_agents_dir = Path(
            (os.getenv("OPENCLAW_BACKUP_ROOT_AGENTS_DIR") or "/root/.agents").strip() or "/root/.agents"
        ).resolve()
        root_ssh_dir = Path((os.getenv("OPENCLAW_BACKUP_ROOT_SSH_DIR") or "/root/.ssh").strip() or "/root/.ssh").resolve()
        root_env_dir = Path((os.getenv("OPENCLAW_BACKUP_ROOT_ENV_DIR") or "/root/.env.d").strip() or "/root/.env.d").resolve()
        root_npm_dir = Path((os.getenv("OPENCLAW_BACKUP_ROOT_NPM_DIR") or "/root/.npm").strip() or "/root/.npm").resolve()

        root_lark_cli_dir = Path(
            (os.getenv("OPENCLAW_BACKUP_ROOT_LARK_CLI_DIR") or "/root/.lark-cli").strip() or "/root/.lark-cli"
        ).resolve()
        work_dir = Path(os.getenv("OPENCLAW_BACKUP_WORK_DIR", "/tmp/openclaw-backup")).resolve()
        metadata_dir = Path(os.getenv("OPENCLAW_BACKUP_METADATA_DIR", "/root/.backup-info")).resolve()
        repo_type = (os.getenv("OPENCLAW_BACKUP_REPO_TYPE") or "dataset").strip() or "dataset"
        path_prefix = (os.getenv("OPENCLAW_BACKUP_PATH_PREFIX") or "backups").strip("/")
        private = env_bool("OPENCLAW_BACKUP_PRIVATE", True)
        keep_count = 48
        keep_count_raw = (os.getenv("OPENCLAW_BACKUP_KEEP_COUNT") or "").strip()
        if keep_count_raw:
            try:
                parsed_keep_count = int(keep_count_raw)
                if parsed_keep_count >= 1:
                    keep_count = parsed_keep_count
                else:
                    print(
                        f"invalid OPENCLAW_BACKUP_KEEP_COUNT={keep_count_raw!r} (must be >= 1), using default 48",
                        file=sys.stderr,
                    )
            except ValueError:
                print(
                    f"invalid OPENCLAW_BACKUP_KEEP_COUNT={keep_count_raw!r} (not a number), using default 48",
                    file=sys.stderr,
                )

        # Parse extra directories from OPENCLAW_BACKUP_EXTRA_DIRS
        # Format: "archive_name:/path/to/dir,archive_name2:/path/to/dir2"
        extra_dirs: dict[str, Path] = {}
        extra_dirs_raw = (os.getenv("OPENCLAW_BACKUP_EXTRA_DIRS") or "").strip()
        if extra_dirs_raw:
            for item in extra_dirs_raw.split(","):
                item = item.strip()
                if not item:
                    continue
                if ":" not in item:
                    print(
                        f"invalid extra dir entry (missing ':'): {item!r}, skipping",
                        file=sys.stderr,
                    )
                    continue
                archive_name, dir_path = item.split(":", 1)
                archive_name = archive_name.strip()
                dir_path = dir_path.strip()
                if not archive_name or not dir_path:
                    print(
                        f"invalid extra dir entry (empty name or path): {item!r}, skipping",
                        file=sys.stderr,
                    )
                    continue
                extra_dirs[archive_name] = Path(dir_path).resolve()

        # Parse extra files from OPENCLAW_BACKUP_EXTRA_FILES
        # Format: "archive_name:/path/to/file,archive_name2:/path/to/file2"
        extra_files: dict[str, Path] = {}
        extra_files_raw = (os.getenv("OPENCLAW_BACKUP_EXTRA_FILES") or "").strip()
        if extra_files_raw:
            for item in extra_files_raw.split(","):
                item = item.strip()
                if not item:
                    continue
                if ":" not in item:
                    print(
                        f"invalid extra file entry (missing ':'): {item!r}, skipping",
                        file=sys.stderr,
                    )
                    continue
                archive_name, file_path = item.split(":", 1)
                archive_name = archive_name.strip()
                file_path = file_path.strip()
                if not archive_name or not file_path:
                    print(
                        f"invalid extra file entry (empty name or path): {item!r}, skipping",
                        file=sys.stderr,
                    )
                    continue
                extra_files[archive_name] = Path(file_path).resolve()

        return cls(
            dataset_repo=dataset_repo,
            state_dir=state_dir,
            backup_source_dir=backup_source_dir,
            work_dir=work_dir,
            metadata_dir=metadata_dir,
            repo_type=repo_type,
            path_prefix=path_prefix,
            private=private,
            incremental=env_bool("OPENCLAW_INCREMENTAL_BACKUP", True),
            incremental_interval_minutes=int(os.getenv("OPENCLAW_INCREMENTAL_INTERVAL_MINUTES", "5")),
            compression_level=int(os.getenv("OPENCLAW_BACKUP_COMPRESSION_LEVEL", "6")),
            split_size=(os.getenv("OPENCLAW_BACKUP_SPLIT_SIZE") or "").strip(),
            size_warning_mb=int(os.getenv("OPENCLAW_BACKUP_SIZE_WARNING_MB", "1500")),
            # Dynamic backup strategy
            dynamic_backup=env_bool("OPENCLAW_DYNAMIC_BACKUP", True),
            dynamic_small_threshold_mb=int(os.getenv("OPENCLAW_DYNAMIC_SMALL_THRESHOLD_MB", "500")),
            dynamic_medium_threshold_mb=int(os.getenv("OPENCLAW_DYNAMIC_MEDIUM_THRESHOLD_MB", "2000")),
            dynamic_high_change_rate=int(os.getenv("OPENCLAW_DYNAMIC_HIGH_CHANGE_RATE", "10")),
            dynamic_low_change_rate=int(os.getenv("OPENCLAW_DYNAMIC_LOW_CHANGE_RATE", "2")),
            dynamic_min_changed_files=int(os.getenv("OPENCLAW_DYNAMIC_MIN_CHANGED_FILES", "5")),
            dynamic_min_changed_size_kb=int(os.getenv("OPENCLAW_DYNAMIC_MIN_CHANGED_SIZE_KB", "100")),
            full_backup_interval_hours=int(os.getenv("OPENCLAW_FULL_BACKUP_INTERVAL_HOURS", "2")),
            max_incremental_backups=int(os.getenv("OPENCLAW_MAX_INCREMENTAL_BACKUPS", "10")),
            root_config_dir=root_config_dir,
            root_codex_dir=root_codex_dir,
            root_claude_dir=root_claude_dir,
            root_agents_dir=root_agents_dir,
            root_ssh_dir=root_ssh_dir,
            root_env_dir=root_env_dir,
            root_npm_dir=root_npm_dir,
            root_lark_cli_dir=root_lark_cli_dir,
            extra_dirs=extra_dirs,
            extra_files=extra_files,
            keep_count=keep_count,
            # Multi-dataset configuration
            restore_dataset_repo=restore_dataset_repo,
            backup_npm_enabled=backup_npm_enabled,
            restore_npm_enabled=restore_npm_enabled,
            # Encryption settings
            encryption_enabled=encryption_enabled,
            encryption_password=encryption_password,
        )


class OpenClawBackup:
    _ARCHIVE_NAME_RE = re.compile(r"^openclaw-backup-(\d{8}-\d{6})\.tar\.gz(?:\.enc)?(?:\.part-[a-z]{2})?$")

    def __init__(self, config: BackupConfig, api: Optional[Any] = None, token: Optional[str] = None) -> None:
        self.config = config
        self.token = self._resolve_token(token)
        config.metadata_dir.mkdir(parents=True, exist_ok=True)
        self._meta_cache: dict[str, dict] = {}
        self._meta_file_cache: dict[str, Path] = {}
        if api is not None:
            self.api = api
            return
        if HfApi is None:
            raise ModuleNotFoundError(
                "huggingface_hub is required. Install with: pip install 'huggingface_hub[cli]'"
            )
        self.api = HfApi(token=self.token)

    def _resolve_token(self, explicit_token: Optional[str] = None) -> Optional[str]:
        """Resolve HF token with proper fallback hierarchy.

        Priority:
        1. Explicitly passed token (from constructor)
        2. HF_TOKEN from os.environ (HF Space variables, highest priority)
        3. HUGGINGFACE_HUB_TOKEN from os.environ
        4. HF_TOKEN from env files (/etc/profile.d/openclaw-env.sh, /root/.env.d/openclaw-backup.env)
        5. HUGGINGFACE_HUB_TOKEN from env files
        6. None (no token available)
        """
        if explicit_token:
            return explicit_token

        for env_var in ("HF_TOKEN", "HUGGINGFACE_HUB_TOKEN"):
            token = os.getenv(env_var)
            if token:
                return token

        env_files = [
            "/etc/profile.d/openclaw-env.sh",
            "/root/.env.d/openclaw-backup.env",
        ]
        for env_file in env_files:
            if os.path.isfile(env_file):
                try:
                    result = subprocess.run(
                        ["bash", "-c", f"source '{env_file}' 2>/dev/null && env"],
                        capture_output=True,
                        text=True,
                        timeout=10,
                    )
                    if result.returncode == 0:
                        for line in result.stdout.strip().split("\n"):
                            if "=" in line:
                                key, _, value = line.partition("=")
                                if key in ("HF_TOKEN", "HUGGINGFACE_HUB_TOKEN") and value:
                                    return value
                except Exception:
                    pass
        return None

    @staticmethod
    def _calculate_checksum(file_path: Path, algorithm: str = "sha256") -> str:
        """Calculate checksum for a file.
        
        Enterprise-grade data integrity verification.
        """
        hash_obj = hashlib.new(algorithm)
        with open(file_path, "rb") as f:
            for chunk in iter(lambda: f.read(8192), b""):
                hash_obj.update(chunk)
        return hash_obj.hexdigest()

    def _verify_backup_integrity(self, archive_path: Path, is_split: bool = False) -> dict:
        """Verify backup archive integrity.
        
        For split volumes, only checks existence and size (cannot validate tar structure of partial file).
        
        Returns a dictionary with verification results.
        """
        result = {
            "valid": False,
            "checksum": None,
            "size": 0,
            "error": None,
        }
        
        try:
            # Check file exists and is readable
            if not archive_path.exists():
                result["error"] = "Archive does not exist"
                return result
            
            # Check file size
            result["size"] = archive_path.stat().st_size
            if result["size"] == 0:
                result["error"] = "Archive is empty"
                return result
            
            # Calculate checksum
            result["checksum"] = self._calculate_checksum(archive_path)
            
            # For split volumes, skip tar validation (partial file)
            if not is_split:
                # Verify tar archive can be opened
                with tarfile.open(archive_path, "r:gz") as tar:
                    # Try to list members (this validates the archive structure)
                    members = tar.getmembers()
                    if not members:
                        result["error"] = "Archive is empty (no members)"
                        return result

                    # Debug: print first few members
                    print(f"    Archive contains {len(members)} entries")
                    for m in members[:10]:
                        print(f"      - {m.name}")
                    if len(members) > 10:
                        print(f"      ... and {len(members) - 10} more entries")

                    # Archive structure validation passed
                    # Note: valid archive roots include predefined ones (openclaw-state,
                    # root-config, etc.) plus custom ones from extra_dirs/extra_files env vars.
                    # Since custom names are user-defined, we only check that:
                    # 1. Archive is a valid tar file (proven by being able to list members)
                    # 2. Archive is not empty (checked above)
                    # The specific root directory names are validated by the backup source
                    # configuration, not here.
            
            result["valid"] = True
            
        except tarfile.TarError as e:
            result["error"] = f"Invalid tar archive: {e}"
        except Exception as e:
            result["error"] = f"Verification failed: {e}"
        
        return result

    def ensure_dataset_repo(self) -> None:
        """Ensure the backup target repository exists.

        """
        target_repo = self.config.dataset_repo
        self.api.create_repo(
            repo_id=target_repo,
            repo_type=self.config.repo_type,
            private=self.config.private,
            exist_ok=True,
        )

    def _check_repo_exists(self, repo_id: str, repo_type: str = "dataset") -> bool:
        """Check if a repository exists.

        Args:
            repo_id: Repository ID to check
            repo_type: Repository type (dataset, model, space)

        Returns:
            True if repository exists, False otherwise
        """
        try:
            self.api.list_repo_files(
                repo_id=repo_id,
                repo_type=repo_type,
            )
            return True
        except Exception as e:
            error_str = str(e).lower()
            if "not found" in error_str or "404" in error_str or "does not exist" in error_str:
                return False
            raise

    def _get_restore_repo(self) -> str:
        """Get the repository to restore from, with fallback logic.

        If OPENCLAW_RESTORE_DATASET_REPO is set but doesn't exist,
        falls back to OPENCLAW_BACKUP_DATASET_REPO.

        Returns:
            Repository ID to restore from
        """
        restore_repo = self.config.restore_dataset_repo
        backup_repo = self.config.dataset_repo

        if restore_repo and restore_repo != backup_repo:
            if self._check_repo_exists(restore_repo, self.config.repo_type):
                print(f"Using restore dataset: {restore_repo}")
                return restore_repo
            else:
                print(f"Warning: Restore dataset '{restore_repo}' does not exist or is not accessible")
                print(f"Falling back to backup dataset: {backup_repo}")
                return backup_repo

        print(f"Using backup dataset for restore: {backup_repo}")
        return backup_repo

    def _calculate_change_rate(self) -> float:
        """Calculate file change rate (files per minute) since last backup."""
        rate, _ = self._calculate_change_rate_and_count()
        return rate

    def _calculate_change_rate_and_count(self) -> tuple[float, int]:
        """Calculate file change rate (files per minute) and count since last backup.
        
        Returns:
            Tuple of (change_rate, changed_files_count)
        """
        last_backup_time = self._get_last_backup_time()
        if last_backup_time is None:
            return (float('inf'), 0)  # First backup, treat as high change rate
        
        minutes_since_last = (utc_now() - last_backup_time).total_seconds() / 60
        if minutes_since_last < 1:
            minutes_since_last = 1  # Avoid division by zero
        
        # Count recently modified files
        changed_files = 0
        for directory in [self.config.backup_source_dir] + list(self._extra_root_dirs().values()):
            if directory.exists():
                changed_files += len(self._get_changed_files(directory, last_backup_time))
        
        return (changed_files / minutes_since_last, changed_files)

    def _estimate_backup_size_mb(self) -> float:
        """Estimate the size of the next backup in MB."""
        total_size = 0
        
        if not self.config.incremental or self._should_do_full_backup(use_cached=True):
            # Full backup - calculate total size
            for directory in [self.config.backup_source_dir] + list(self._extra_root_dirs().values()):
                if directory.exists():
                    for item in directory.rglob("*"):
                        if item.is_file():
                            try:
                                total_size += item.stat().st_size
                            except OSError:
                                continue
        else:
            # Incremental backup - only changed files
            last_backup_time = self._get_last_backup_time()
            if last_backup_time:
                for directory in [self.config.backup_source_dir] + list(self._extra_root_dirs().values()):
                    if directory.exists():
                        for file_path in self._get_changed_files(directory, last_backup_time):
                            try:
                                total_size += file_path.stat().st_size
                            except OSError:
                                continue
        
        return total_size / (1024 * 1024)

    def _apply_dynamic_strategy(self, force_full_backup: bool = False) -> dict[str, any]:
        """Apply dynamic backup strategy based on file size and change rate.
        
        Args:
            force_full_backup: If True, skip checking for incremental backup eligibility
                              (backup type decision already made by caller).
        """
        if not self.config.dynamic_backup:
            return {
                "compression_level": self.config.compression_level,
                "split_size": self.config.split_size,
                "skip_backup": False,
                "reason": "Dynamic backup disabled",
            }
        
        # Get current metrics
        estimated_size_mb = self._estimate_backup_size_mb()
        estimated_size_kb = estimated_size_mb * 1024
        change_rate, changed_files = self._calculate_change_rate_and_count()
        
        print(f"[Dynamic Strategy] Estimated size: {estimated_size_mb:.1f}MB, Change rate: {change_rate:.1f} files/min, Changed files: {changed_files}")
        
        # Determine file size category
        if estimated_size_mb < self.config.dynamic_small_threshold_mb:
            size_category = "small"
            compression = 3  # Fast compression for small files
            split = ""
        elif estimated_size_mb < self.config.dynamic_medium_threshold_mb:
            size_category = "medium"
            compression = 6  # Balanced for medium files
            split = "500M" if estimated_size_mb > 1000 else ""
        else:
            size_category = "large"
            compression = 9  # Maximum compression for large files
            split = "500M"
        
        # Adjust based on change rate
        skip_reason = None

        if change_rate > self.config.dynamic_high_change_rate:
            # High change rate - prioritize speed over compression
            compression = max(1, compression - 2)
        elif change_rate < self.config.dynamic_low_change_rate:
            # Low change rate - can use higher compression
            compression = min(9, compression + 1)
            # Skip if very low change rate AND (fewer changed files OR smaller total size) than minimum threshold
            # But still perform backup if it's a scheduled full backup
            if changed_files < self.config.dynamic_min_changed_files and estimated_size_kb < self.config.dynamic_min_changed_size_kb and not force_full_backup:
                current_skipped = self._load_skipped_count() + 1
                self._save_skipped_count(current_skipped)

                max_skipped_incrementals = max(3, self.config.max_incremental_backups // 3)
                if current_skipped >= max_skipped_incrementals:
                    print(f"[Dynamic Strategy] Too many skipped incrementals ({current_skipped}), forcing backup to maintain chain continuity")
                    self._clear_skipped_count()
                else:
                    skip_reason = f"Very low change rate ({change_rate:.1f} files/min, {changed_files} files, {estimated_size_kb:.0f}KB), skipping incremental backup ({current_skipped}/{max_skipped_incrementals})"
                    print(f"[Dynamic Strategy] {skip_reason}")
                    return {
                        "compression_level": compression,
                        "split_size": split,
                        "skip_backup": True,
                        "reason": skip_reason,
                    }
        else:
            self._clear_skipped_count()
        
        # Adjust compression based on size category
        if size_category == "large" and estimated_size_mb > 3000:
            # Very large files - always use max compression
            compression = 9
        
        strategy = {
            "compression_level": compression,
            "split_size": split,
            "skip_backup": False,
            "reason": f"Size: {size_category} ({estimated_size_mb:.1f}MB), Change rate: {change_rate:.1f} files/min",
        }
        
        print(f"[Dynamic Strategy] Applied: compression={compression}, split={split or 'none'}")
        
        return strategy

    def _extra_root_dirs(self) -> dict[str, Path]:
        dirs = {
            "root-config": self.config.root_config_dir,
            "root-codex": self.config.root_codex_dir,
            "root-claude": self.config.root_claude_dir,
            "root-agents": self.config.root_agents_dir,
            "root-ssh": self.config.root_ssh_dir,
            "root-env": self.config.root_env_dir,
            "root-lark-cli": self.config.root_lark_cli_dir,
            "backup-info": self.config.metadata_dir,
        }
        if self.config.backup_npm_enabled:
            dirs["root-npm"] = self.config.root_npm_dir
        if self.config.extra_dirs:
            dirs.update(self.config.extra_dirs)
        return dirs

    def _restore_root_dirs(self) -> dict[str, Path]:
        dirs = {
            "root-config": self.config.root_config_dir,
            "root-codex": self.config.root_codex_dir,
            "root-claude": self.config.root_claude_dir,
            "root-agents": self.config.root_agents_dir,
            "root-ssh": self.config.root_ssh_dir,
            "root-env": self.config.root_env_dir,
            "root-lark-cli": self.config.root_lark_cli_dir,
            "backup-info": self.config.metadata_dir,
        }
        if self.config.restore_npm_enabled:
            dirs["root-npm"] = self.config.root_npm_dir
        if self.config.extra_dirs:
            dirs.update(self.config.extra_dirs)
        return dirs

    def _get_last_backup_time(self) -> Optional[dt.datetime]:
        """Get the timestamp of the last successful backup from metadata.

        Checks metadata_dir for last-backup-metadata.json, then falls back to
        downloading latest-backup.json and the corresponding .meta.json from
        remote storage.
        """
        metadata_path = self.config.metadata_dir / "last-backup-metadata.json"
        try:
            if metadata_path.exists():
                with open(metadata_path, "r") as f:
                    metadata = json.load(f)
                    last_backup_str = metadata.get("last_backup_time")
                    if last_backup_str:
                        return dt.datetime.fromisoformat(last_backup_str)
        except Exception:
            pass

        return self._get_last_backup_time_from_remote()

    def _get_remote_backup_chain_info(self) -> Optional[dict]:
        """Get the complete backup chain info from remote storage.

        Downloads latest-backup.json and the latest .meta.json to get:
        - last_backup_time: timestamp of last backup
        - parent_meta_path: path to parent backup's .meta.json
        - chain_id: the chain this backup belongs to
        - volumes: list of volume filenames (for split backups)
        - backup_type: "full" or "incremental"

        This is the source of truth for backup chain continuity.
        """
        try:
            latest_json_path = "latest-backup.json"
            latest_json_local = self._download_backup(latest_json_path, self.config.work_dir)
            with open(latest_json_local, "r") as f:
                latest_json = json.load(f)
            meta_path = latest_json.get("latest", "")
            if not meta_path:
                return None

            meta_local = self._download_backup(meta_path, self.config.work_dir)
            with open(meta_local, "r") as f:
                meta = json.load(f)

            meta_filename = Path(meta_path).name
            base_archive_name = meta_filename.replace(".meta.json", "")

            return {
                "last_backup_time": meta.get("last_backup_time") or meta.get("created_at_utc"),
                "parent_meta_path": meta.get("parent"),  # The parent field from the .meta.json file
                "chain_id": meta.get("chain_id"),
                "volumes": meta.get("volumes", [base_archive_name]),
                "backup_type": meta.get("backup_type", "incremental"),
                "is_split": meta.get("is_split", False),
            }
        except Exception:
            return None

    def _get_last_backup_time_from_remote(self) -> Optional[dt.datetime]:
        """Download latest-backup.json and corresponding .meta.json from remote to get last backup time.

        This is used when local metadata is not available (e.g., first run on a new machine).
        """
        try:
            latest_json_path = "latest-backup.json"
            print(f"[_get_last_backup_time_from_remote] Downloading {latest_json_path}...")
            latest_json_local = self._download_backup(latest_json_path, self.config.work_dir)
            print(f"[_get_last_backup_time_from_remote] Downloaded to {latest_json_local}")
            with open(latest_json_local, "r") as f:
                latest_json = json.load(f)
            meta_path = latest_json.get("latest", "")
            print(f"[_get_last_backup_time_from_remote] latest path: {meta_path}")
            if not meta_path:
                return None

            print(f"[_get_last_backup_time_from_remote] Downloading {meta_path}...")
            cached_meta = self._get_cached_meta(meta_path)
            if cached_meta is not None:
                meta = cached_meta
                print(f"[_get_last_backup_time_from_remote] Using cached metadata for {meta_path}")
            else:
                meta_local = self._download_backup(meta_path, self.config.work_dir)
                print(f"[_get_last_backup_time_from_remote] Downloaded to {meta_local}")
                with open(meta_local, "r") as f:
                    meta = json.load(f)
                self._set_cached_meta(meta_path, meta)
            last_backup_str = meta.get("last_backup_time")
            print(f"[_get_last_backup_time_from_remote] last_backup_time: {last_backup_str}")
            if last_backup_str:
                return dt.datetime.fromisoformat(last_backup_str)
        except Exception as e:
            print(f"[_get_last_backup_time_from_remote] Error: {type(e).__name__}: {e}")
            pass
        return None

    def _save_backup_metadata(
        self,
        archive_path: Path,
        file_count: int,
        is_full_backup: bool = True,
        parent_meta_path: Optional[str] = None,
        chain_id: Optional[str] = None
    ) -> None:
        """Save metadata about this backup for future incremental backups.

        This creates the per-backup .meta.json file content.
        The metadata is saved locally and also uploaded to remote storage.

        Args:
            archive_path: Path to the archive file (may be split into multiple volumes)
            file_count: Number of files in the backup
            is_full_backup: Whether this is a full backup
            parent_meta_path: Path to parent backup's .meta.json file (for incremental)
            chain_id: Chain ID for this backup (generated for full, inherited for incremental)
        """
        if chain_id is None:
            chain_id = str(uuid.uuid4())[:8]
        prev_metadata = self._get_backup_metadata()

        if is_full_backup:
            backup_type = "full"
            chain_id = str(uuid.uuid4())[:8]  # Full backup generates new chain_id
        else:
            backup_type = "incremental"
            # For incremental, use provided chain_id or inherit from remote
            if not chain_id:
                chain_id = prev_metadata.get("chain_id", str(uuid.uuid4())[:8])

        checksum = self._calculate_checksum(archive_path)
        archive_size = archive_path.stat().st_size

        if is_full_backup:
            incremental_count = 0
            last_full_backup_time = utc_now().isoformat(timespec="seconds")
        else:
            incremental_count = (prev_metadata.get("incremental_count", 0) + 1) if prev_metadata else 1
            last_full_backup_time = prev_metadata.get("last_full_backup_time") if prev_metadata else None

        metadata = {
            "version": METADATA_VERSION,
            "backup_type": backup_type,
            "chain_id": chain_id,
            "parent": parent_meta_path,
            "volumes": [archive_path.name],
            "checksum": f"sha256:{checksum}",
            "created_at_utc": utc_now().isoformat(timespec="seconds"),
            "last_backup_time": utc_now().isoformat(timespec="seconds"),
            "file_count": file_count,
            "archive_size": archive_size,
            "is_latest": True,
            "created_by": "openclaw-backup",
            "incremental_count": incremental_count,
            "last_full_backup_time": last_full_backup_time,
        }

        if prev_metadata:
            prev_metadata["is_latest"] = False
            prev_path = self.config.metadata_dir / "last-backup-metadata.json"
            with open(prev_path, "w") as f:
                json.dump(prev_metadata, f, indent=2)

        metadata_path = self.config.metadata_dir / "last-backup-metadata.json"
        with open(metadata_path, "w") as f:
            json.dump(metadata, f, indent=2)

    def _get_backup_metadata(self) -> Optional[dict]:
        """Get the current backup metadata from local files.

        Checks metadata_dir for last-backup-metadata.json.
        """
        metadata_path = self.config.metadata_dir / "last-backup-metadata.json"
        try:
            if metadata_path.exists():
                with open(metadata_path, "r") as f:
                    return json.load(f)
        except Exception:
            pass
        return None

    def _get_skipped_count_file(self) -> Path:
        """Get path to the skipped incremental count file."""
        return self.config.metadata_dir / ".skipped-incremental-count"

    def _load_skipped_count(self) -> int:
        """Load the count of consecutively skipped incremental backups."""
        skipped_file = self._get_skipped_count_file()
        try:
            if skipped_file.exists():
                return int(skipped_file.read_text().strip())
        except (ValueError, OSError):
            pass
        return 0

    def _save_skipped_count(self, count: int) -> None:
        """Save the count of consecutively skipped incremental backups."""
        skipped_file = self._get_skipped_count_file()
        try:
            self.config.metadata_dir.mkdir(parents=True, exist_ok=True)
            skipped_file.write_text(str(count))
        except OSError:
            pass

    def _clear_skipped_count(self) -> None:
        """Clear the skipped incremental count (when a backup is performed)."""
        skipped_file = self._get_skipped_count_file()
        try:
            skipped_file.unlink(missing_ok=True)
        except OSError:
            pass

    def _update_volumes_in_metadata(self, volumes: list[str], checksum: Optional[str] = None, archive_size: Optional[int] = None) -> None:
        """Update volumes field in metadata after split/encryption.

        This is necessary because _save_backup_metadata is called BEFORE split/encryption,
        so it records the original archive name. After splitting and/or encrypting, we need
        to update with the actual volume names and checksums.

        Args:
            volumes: List of split volume filenames
            checksum: New checksum for the encrypted archive (optional)
            archive_size: New archive size in bytes (optional)
        """
        try:
            metadata_path = self.config.metadata_dir / "last-backup-metadata.json"
            if metadata_path.exists():
                with open(metadata_path, "r") as f:
                    metadata = json.load(f)
                metadata["volumes"] = volumes
                if checksum:
                    metadata["checksum"] = checksum
                if archive_size is not None:
                    metadata["archive_size"] = archive_size
                if any(v.endswith(".enc") for v in volumes):
                    metadata["encrypted"] = True
                    metadata["encryption_algorithm"] = "AES-256-CBC"
                if any(".part-" in v for v in volumes):
                    metadata["is_split"] = True
                with open(metadata_path, "w") as f:
                    json.dump(metadata, f, indent=2)
        except Exception as e:
            print(f"Warning: Could not update metadata: {e}")

    def _should_do_full_backup(self, use_cached: bool = False) -> bool:
        """Determine if a full backup should be performed instead of incremental.

        Args:
            use_cached: If True, return cached result from previous call in same backup cycle.
                       If False, always compute fresh (but will cache if this is first call).
        """
        cached_attr = '_cached_should_full_backup'

        if use_cached and hasattr(self, cached_attr):
            return getattr(self, cached_attr)

        if not self.config.incremental:
            return True

        last_backup_time = self._get_last_backup_time()
        if last_backup_time is None:
            print("[DEBUG] No previous backup found, performing full backup")
            result = True
        else:
            time_since_last = utc_now() - last_backup_time
            if time_since_last.total_seconds() < self.config.incremental_interval_minutes * 60:
                print(f"[DEBUG] Incremental interval not reached ({time_since_last.total_seconds():.0f}s < {self.config.incremental_interval_minutes * 60}s), skipping backup")
                result = False
            else:
                result = self._should_force_full_backup()

        setattr(self, cached_attr, result)
        return result
    
    def _should_force_full_backup(self) -> bool:
        """Check if full backup should be forced based on interval/count rules."""
        metadata = self._get_backup_metadata()

        last_full_backup_time_str = None
        incremental_count = 0
        if metadata:
            last_full_backup_time_str = metadata.get("last_full_backup_time")
            incremental_count = metadata.get("incremental_count", 0)

        if last_full_backup_time_str:
            last_full_backup_time = dt.datetime.fromisoformat(last_full_backup_time_str)
            hours_since_full = (utc_now() - last_full_backup_time).total_seconds() / 3600
            if hours_since_full >= self.config.full_backup_interval_hours:
                print(f"[DEBUG] Full backup interval reached ({hours_since_full:.1f}h >= {self.config.full_backup_interval_hours}h), forcing full backup")
                return True

        if incremental_count >= self.config.max_incremental_backups:
            print(f"[DEBUG] Max incremental backups reached ({incremental_count}/{self.config.max_incremental_backups}), forcing full backup")
            return True

        return False

    def _get_changed_files(self, directory: Path, since: dt.datetime) -> list[Path]:
        """Get list of files that have been modified since the given timestamp."""
        changed_files = []
        if not directory.exists():
            return changed_files

        for item in directory.rglob("*"):
            if item.is_file():
                try:
                    mtime = dt.datetime.fromtimestamp(item.stat().st_mtime, tz=dt.UTC)
                    if mtime > since:
                        changed_files.append(item)
                except (OSError, ValueError):
                    # Skip files we can't stat
                    continue
        return changed_files

    def create_archive(self, timestamp: Optional[str] = None) -> Path | None:
        if not self.config.backup_source_dir.exists():
            raise FileNotFoundError(f"backup source directory does not exist: {self.config.backup_source_dir}")

        stamp = timestamp or utc_now().strftime("%Y%m%d-%H%M%S")
        self.config.work_dir.mkdir(parents=True, exist_ok=True)

        # Clean up stale checkpoints from previous interrupted uploads before starting
        self._cleanup_stale_checkpoints()

        # Get remote backup chain info FIRST (source of truth)
        # This ensures continuity even if local metadata is lost
        remote_chain_info = self._get_remote_backup_chain_info()

        # CRITICAL: Call _should_do_full_backup FIRST with caching
        # This must happen BEFORE _apply_dynamic_strategy to ensure consistent decisions
        do_full_backup = self._should_do_full_backup(use_cached=False)
        print(f"[BACKUP] Full backup decision: {do_full_backup} (cached=False call)")

        # Apply dynamic backup strategy (pass already-computed decision)
        strategy = self._apply_dynamic_strategy(force_full_backup=do_full_backup)

        # Check if backup should be skipped
        if strategy.get("skip_backup", False):
            print(f"[Dynamic Strategy] Backup skipped: {strategy.get('reason', 'Unknown')}")
            return None

        # Apply dynamic settings
        dynamic_compression = strategy.get("compression_level", self.config.compression_level)
        dynamic_split = strategy.get("split_size", self.config.split_size)

        # Temporarily override config for this backup
        original_compression = self.config.compression_level
        original_split = self.config.split_size
        self.config.compression_level = dynamic_compression
        self.config.split_size = dynamic_split

        try:
            # Use cached result from earlier call
            last_backup_time = self._get_last_backup_time()

            # Get parent_meta_path and chain_id from remote chain info
            parent_meta_path = None
            chain_id = None
            if remote_chain_info:
                chain_id = remote_chain_info.get("chain_id")
                if do_full_backup:
                    parent_meta_path = None  # Full backup starts new chain
                    print(f"[BACKUP] Creating new chain (full backup), chain_id={chain_id}")
                else:
                    # For incremental backup: find the latest backup in the chain
                    # Use chain_id to find all backups, the latest one is the parent
                    chain = self._find_all_backups_by_chain_id(chain_id, repo_id=self.config.dataset_repo)
                    if chain and len(chain) > 0:
                        parent_meta_path = chain[-1]  # Last one is the latest
                        print(f"[BACKUP] Remote chain: parent={parent_meta_path}, chain_id={chain_id}")
                    else:
                        # Fallback: use remote_chain_info's parent_meta_path
                        # This handles cases where we can't enumerate all backups
                        parent_meta_path = remote_chain_info.get("parent_meta_path")
                        if parent_meta_path:
                            print(f"[BACKUP] Remote chain (fallback): parent={parent_meta_path}, chain_id={chain_id}")
                        else:
                            print(f"[BACKUP] Warning: No parent found, starting new chain, chain_id={chain_id}")

            archive = self.config.work_dir / f"openclaw-backup-{stamp}.tar.gz"
            file_count = 0

            # Use configured compression level (1-9, where 9 is maximum compression)
            # For large files (700MB+), level 6 provides good balance
            compression_level = max(1, min(9, self.config.compression_level))
            
            with tarfile.open(archive, "w:gz", compresslevel=compression_level) as tar:
                if do_full_backup or last_backup_time is None:
                    print(f"[BACKUP] Creating FULL backup: {archive.name}")
                    print(f"[BACKUP] ============================================")
                    print(f"[BACKUP] Backup source: {self.config.backup_source_dir}")

                    # Backup main state directory
                    if self.config.backup_source_dir.exists():
                        state_files = [f for f in self.config.backup_source_dir.rglob("*") if f.is_file()]
                        tar.add(str(self.config.backup_source_dir), arcname="openclaw-state")
                        file_count += len(state_files)
                        print(f"[BACKUP]   ✓ openclaw-state: {len(state_files)} files")
                    else:
                        print(f"[BACKUP]   ✗ openclaw-state: directory not found!")

                    # Backup extra directories
                    for archive_root, local_dir in self._extra_root_dirs().items():
                        if local_dir.exists():
                            extra_files = [f for f in local_dir.rglob("*") if f.is_file()]
                            tar.add(str(local_dir), arcname=archive_root)
                            file_count += len(extra_files)
                            print(f"[BACKUP]   ✓ {archive_root}: {len(extra_files)} files")
                        else:
                            print(f"[BACKUP]   - {archive_root}: directory not found (skipped)")
                else:
                    print(f"[BACKUP] Creating INCREMENTAL backup: {archive.name}")
                    print(f"[BACKUP] ============================================")
                    print(f"[BACKUP] Last backup: {last_backup_time.isoformat()}")

                    # Add changed files from main state directory
                    changed_files = self._get_changed_files(self.config.backup_source_dir, last_backup_time)
                    for file_path in changed_files:
                        arcname = f"openclaw-state/{file_path.relative_to(self.config.backup_source_dir)}"
                        tar.add(str(file_path), arcname=arcname)
                    file_count += len(changed_files)
                    print(f"[BACKUP]   ✓ openclaw-state: {len(changed_files)} changed files")
                    if changed_files:
                        for fp in changed_files[:5]:
                            print(f"[BACKUP]      - {fp.name}")
                        if len(changed_files) > 5:
                            print(f"[BACKUP]      ... and {len(changed_files) - 5} more")

                    # Add changed files from extra directories
                    for archive_root, local_dir in self._extra_root_dirs().items():
                        if local_dir.exists():
                            changed_files = self._get_changed_files(local_dir, last_backup_time)
                            for file_path in changed_files:
                                arcname = f"{archive_root}/{file_path.relative_to(local_dir)}"
                                tar.add(str(file_path), arcname=arcname)
                            file_count += len(changed_files)
                            print(f"[BACKUP]   ✓ {archive_root}: {len(changed_files)} changed files")
                            if changed_files:
                                for fp in changed_files[:5]:
                                    print(f"[BACKUP]      - {fp.name}")
                                if len(changed_files) > 5:
                                    print(f"[BACKUP]      ... and {len(changed_files) - 5} more")
                        else:
                            print(f"[BACKUP]   - {archive_root}: directory not found (skipped)")

                # Always add extra files (they're typically small config files)
                for archive_name, local_file in self.config.extra_files.items():
                    if local_file.exists() and local_file.is_file():
                        tar.add(str(local_file), arcname=archive_name)
                        file_count += 1

            # Save metadata for next backup
            is_full_backup = do_full_backup or last_backup_time is None
            # Determine parent_meta_path and chain_id from remote chain info (source of truth)
            # For incremental backups, parent is the previous backup's .meta.json
            # For full backups, parent is None and new chain_id is generated
            if not is_full_backup and remote_chain_info:
                # Use parent_meta_path and chain_id from remote - already set above
                pass  # parent_meta_path and chain_id already set from remote_chain_info
            elif is_full_backup:
                parent_meta_path = None  # Full backup starts new chain
                chain_id = None  # Will be generated in _save_backup_metadata
            self._save_backup_metadata(archive, file_count, is_full_backup, parent_meta_path, chain_id)

            # Check archive size and warn if too large
            archive_size_mb = archive.stat().st_size / (1024 * 1024)
            if archive_size_mb > self.config.size_warning_mb:
                print(f"⚠️  WARNING: Backup size ({archive_size_mb:.1f}MB) exceeds warning threshold ({self.config.size_warning_mb}MB)")
                print(f"   Consider: 1) Excluding large files 2) Using split archives 3) Reducing backup frequency")

            if is_full_backup:
                print(f"Full backup complete: {file_count} files, {archive_size_mb:.1f}MB")
            else:
                print(f"Incremental backup complete: {file_count} files, {archive_size_mb:.1f}MB")

            # Encrypt archive if enabled (before split)
            if self.config.encryption_enabled and self.config.encryption_password:
                result = self._encrypt_archive(archive)
                if isinstance(result, Path):
                    enc_checksum = self._calculate_checksum(result)
                    enc_size = result.stat().st_size
                    enc_volumes = [result.name]
                    self._update_volumes_in_metadata(enc_volumes, f"sha256:{enc_checksum}", enc_size)
            else:
                result = archive

            # Split archive if configured and file is large
            # Note: checksum must be calculated BEFORE split for non-encrypted backups
            if self.config.split_size and archive_size_mb > 500:
                # For non-encrypted backups, calculate checksum before split
                if not (self.config.encryption_enabled and self.config.encryption_password):
                    pre_split_checksum = self._calculate_checksum(result)
                    pre_split_size = result.stat().st_size
                result = self._split_archive(result)
                # Update metadata with actual volumes (split files)
                # This must happen AFTER split since _save_backup_metadata is called BEFORE split
                if isinstance(result, list):
                    is_encrypted = self.config.encryption_enabled and self.config.encryption_password
                    if is_encrypted:
                        # For encrypted + split backups, checksum was already stored when encryption
                        # happened (before split). After split, we only update volumes list.
                        # Checksum verification will happen after merge+decrypt in restore flow.
                        self._update_volumes_in_metadata([v.name for v in result])
                    else:
                        # For non-encrypted split backups, use checksum calculated before split
                        total_size = sum(v.stat().st_size for v in result)
                        self._update_volumes_in_metadata(
                            [v.name for v in result],
                            checksum=f"sha256:{pre_split_checksum}",
                            archive_size=total_size
                        )
            
            return result
            
        finally:
            # Restore original config
            self.config.compression_level = original_compression
            self.config.split_size = original_split

    def _encrypt_archive(self, archive: Path) -> Path:
        """Encrypt archive using AES-256-CBC with password.

        Args:
            archive: Path to the archive to encrypt

        Returns:
            Path to the encrypted archive (.tar.gz.enc)
        """
        if not self.config.encryption_password:
            raise ValueError("Encryption password not set")

        encrypted_path = archive.with_suffix(archive.suffix + ".enc")

        print(f"Encrypting archive: {archive.name} -> {encrypted_path.name}")

        try:
            import subprocess
            import tempfile
            import os as _os

            with tempfile.NamedTemporaryFile(mode="w", delete=False) as pw_file:
                _os.chmod(pw_file.file.fileno(), 0o600)
                pw_file.write(self.config.encryption_password)
                pw_file.flush()
                pw_file_path = pw_file.name

            try:
                result = subprocess.run(
                    [
                        "openssl", "enc", "-aes-256-cbc", "-salt", "-pbkdf2",
                        "-in", str(archive),
                        "-out", str(encrypted_path),
                        "-pass", f"file:{pw_file_path}",
                    ],
                    capture_output=True,
                    text=True,
                    check=True,
                )
            finally:
                _os.unlink(pw_file_path)

            archive.unlink()
            print(f"Encrypted: {encrypted_path.name} ({encrypted_path.stat().st_size / (1024*1024):.1f}MB)")
            return encrypted_path
        except subprocess.CalledProcessError as e:
            raise RuntimeError(f"Encryption failed: {e.stderr}")

    def _decrypt_archive(self, encrypted_path: Path, target_path: Path) -> Path:
        """Decrypt archive using AES-256-CBC with password.

        Args:
            encrypted_path: Path to the encrypted archive (.tar.gz.enc)
            target_path: Path where decrypted archive should be written

        Returns:
            Path to the decrypted archive
        """
        if not self.config.encryption_password:
            raise ValueError("Decryption password not set")

        print(f"Decrypting archive: {encrypted_path.name} -> {target_path.name}")

        try:
            import subprocess
            import tempfile
            import os as _os

            with tempfile.NamedTemporaryFile(mode="w", delete=False) as pw_file:
                _os.chmod(pw_file.file.fileno(), 0o600)
                pw_file.write(self.config.encryption_password)
                pw_file.flush()
                pw_file_path = pw_file.name

            try:
                result = subprocess.run(
                    [
                        "openssl", "enc", "-d", "-aes-256-cbc", "-pbkdf2",
                        "-in", str(encrypted_path),
                        "-out", str(target_path),
                        "-pass", f"file:{pw_file_path}",
                    ],
                    capture_output=True,
                    text=True,
                    check=True,
                )
            finally:
                _os.unlink(pw_file_path)

            print(f"Decrypted: {target_path.name} ({target_path.stat().st_size / (1024*1024):.1f}MB)")
            return target_path
        except subprocess.CalledProcessError as e:
            raise RuntimeError(f"Decryption failed: {e.stderr}")

    def _split_archive(self, archive: Path) -> list[Path]:
        """Split large archive into smaller volumes."""
        import subprocess
        
        print(f"Splitting archive into {self.config.split_size} volumes...")
        
        # Parse split size (e.g., "500M" -> 500)
        split_mb = 500
        if self.config.split_size:
            size_str = self.config.split_size.upper()
            if size_str.endswith('M'):
                split_mb = int(size_str[:-1])
            elif size_str.endswith('G'):
                split_mb = int(size_str[:-1]) * 1024
        
        split_bytes = split_mb * 1024 * 1024
        
        # Use split command to create volumes
        base_name = archive.name
        volumes_pattern = self.config.work_dir / f"{base_name}.part-"
        
        try:
            result = subprocess.run(
                ["split", "-b", str(split_bytes), str(archive), str(volumes_pattern)],
                capture_output=True,
                text=True,
                check=True
            )
            
            # Remove original archive after splitting
            archive.unlink()
            
            # Return list of volume files
            volumes = sorted(self.config.work_dir.glob(f"{base_name}.part-*"))
            print(f"Created {len(volumes)} volume(s)")
            for vol in volumes:
                vol_size_mb = vol.stat().st_size / (1024 * 1024)
                print(f"  - {vol.name}: {vol_size_mb:.1f}MB")
            
            return volumes
            
        except subprocess.CalledProcessError as e:
            print(f"Warning: Failed to split archive: {e}")
            print(f"Continuing with unsplit archive")
            return [archive]
        except FileNotFoundError:
            print(f"Warning: 'split' command not found. Install coreutils.")
            print(f"Continuing with unsplit archive")
            return [archive]

    def _upload_file_with_retry(self, path_or_fileobj, path_in_repo: str, commit_message: str, max_retries: int = 3, repo_id: Optional[str] = None) -> None:
        """Upload file with retry logic for rate limiting.

        Args:
            path_or_fileobj: File path or file object to upload
            path_in_repo: Path in the repository
            commit_message: Commit message
            max_retries: Maximum number of retries
            repo_id: Optional repository ID to upload to. If not specified, uses dataset_repo
        """
        upload_repo = repo_id or self.config.dataset_repo

        for attempt in range(max_retries):
            try:
                self.api.upload_file(
                    repo_id=upload_repo,
                    repo_type=self.config.repo_type,
                    path_or_fileobj=path_or_fileobj,
                    path_in_repo=path_in_repo,
                    commit_message=commit_message,
                )
                return
            except Exception as e:
                error_str = str(e)
                # Check for rate limit error (429)
                if "429" in error_str or "Too Many Requests" in error_str:
                    if attempt < max_retries - 1:
                        # Extract retry wait time from error message
                        match = re.search(r'Retry after (\d+) seconds', error_str)
                        if match:
                            wait_time = int(match.group(1))
                            print(f"Rate limit hit (429), waiting {wait_time} seconds (server requested) before retry {attempt + 1}/{max_retries}...")
                        else:
                            # Exponential backoff: 60, 120, 240 seconds
                            wait_time = 60 * (2 ** attempt)
                            print(f"Rate limit hit (429), waiting {wait_time} seconds (exponential backoff) before retry {attempt + 1}/{max_retries}...")
                        time.sleep(wait_time)
                        continue
                # Re-raise if not rate limit or max retries reached
                raise

    def upload_backup(self, archive: Path | list[Path]) -> str:
        if isinstance(archive, list):
            volumes = archive
            is_split = True
        else:
            volumes = [archive]
            is_split = False

        # Get current backup metadata - first from local, then from remote if not found
        backup_meta = self._get_backup_metadata()
        if not backup_meta:
            # Try to get metadata from remote (in case local cache was lost)
            remote_chain_info = self._get_remote_backup_chain_info()
            if remote_chain_info:
                # Build minimal metadata from remote chain info
                backup_meta = {
                    "chain_id": remote_chain_info.get("chain_id"),
                    "parent": remote_chain_info.get("parent_meta_path"),
                    "backup_type": remote_chain_info.get("backup_type"),
                }
                print(f"[UPLOAD] Using remote chain info: chain_id={backup_meta.get('chain_id')}")
            else:
                raise RuntimeError("Backup metadata not found. Cannot upload.")

        # Generate meta.json name early (needed for checkpoint)
        first_volume_name = volumes[0].name
        base_name = first_volume_name
        if ".part-" in base_name:
            base_name = base_name.rsplit(".part-", 1)[0]
        meta_json_name = f"{base_name}.meta.json"
        if self.config.path_prefix:
            meta_json_remote = f"{self.config.path_prefix}/{meta_json_name}"
        else:
            meta_json_remote = meta_json_name

        # Create checkpoint file for recovery - upload BEFORE any volume to establish a recovery point
        checkpoint_json = {
            "status": "uploading",
            "backup_meta_path": meta_json_remote,
            "volumes": [vol.name for vol in volumes],
            "volumes_uploaded": [],
            "created_at_utc": utc_now().isoformat(timespec="seconds"),
        }
        checkpoint_name = f"{base_name}.checkpoint.json"
        if self.config.path_prefix:
            checkpoint_remote = f"{self.config.path_prefix}/{checkpoint_name}"
        else:
            checkpoint_remote = checkpoint_name

        print(f"Creating checkpoint file: {checkpoint_name}")
        self._upload_file_with_retry(
            path_or_fileobj=(json.dumps(checkpoint_json, ensure_ascii=False, indent=2) + "\n").encode("utf-8"),
            path_in_repo=checkpoint_remote,
            commit_message=f"backup: {checkpoint_name} (checkpoint)",
        )

        # Upload all volumes, updating checkpoint after each
        remote_volumes = []
        for vol in volumes:
            if self.config.path_prefix:
                remote_path = f"{self.config.path_prefix}/{vol.name}"
            else:
                remote_path = vol.name

            print(f"Uploading {vol.name}...")
            self._upload_file_with_retry(
                path_or_fileobj=str(vol),
                path_in_repo=remote_path,
                commit_message=f"backup: {vol.name}",
            )
            remote_volumes.append(remote_path)

            # Update checkpoint after each volume upload (for recovery)
            checkpoint_json["volumes_uploaded"].append(vol.name)
            print(f"Updating checkpoint: {vol.name} uploaded")
            self._upload_file_with_retry(
                path_or_fileobj=(json.dumps(checkpoint_json, ensure_ascii=False, indent=2) + "\n").encode("utf-8"),
                path_in_repo=checkpoint_remote,
                commit_message=f"backup: {checkpoint_name} (checkpoint update)",
            )

        # Update backup_meta with remote paths for volumes
        # Note: volumes in .meta.json should be local names (without path_prefix)
        # because restore code will prepend path_prefix when downloading
        backup_meta["volumes"] = [vol.name for vol in volumes]  # Local names only
        backup_meta["is_split"] = is_split

        # Generate and upload per-backup .meta.json
        print(f"Uploading {meta_json_name}...")
        self._upload_file_with_retry(
            path_or_fileobj=(json.dumps(backup_meta, ensure_ascii=False, indent=2) + "\n").encode("utf-8"),
            path_in_repo=meta_json_remote,
            commit_message=f"backup: {meta_json_name}",
        )

        # Update latest-backup.json to point to the new .meta.json
        latest_backup_json = {
            "dataset": self.config.dataset_repo,
            "latest": meta_json_remote,  # Full path with path_prefix
            "created_at_utc": utc_now().isoformat(timespec="seconds"),
        }

        # latest-backup.json should always be uploaded to dataset root (same level as backups/)
        latest_json_path = "latest-backup.json"

        print("Updating latest-backup.json...")
        self._upload_file_with_retry(
            path_or_fileobj=(json.dumps(latest_backup_json, ensure_ascii=False, indent=2) + "\n").encode("utf-8"),
            path_in_repo=latest_json_path,
            commit_message="backup: update latest-backup.json",
        )

        # Delete checkpoint file now that backup is complete
        print(f"Deleting checkpoint file: {checkpoint_name}")
        self._delete_checkpoint_with_retry(checkpoint_remote)

    def _delete_checkpoint_with_retry(self, checkpoint_remote: str, max_retries: int = 3) -> None:
        """Delete checkpoint file with retry logic.

        Args:
            checkpoint_remote: Path to checkpoint file in repo
            max_retries: Maximum deletion attempts
        """
        for attempt in range(max_retries):
            try:
                self.api.delete_file(
                    path_in_repo=checkpoint_remote,
                    repo_id=self.config.dataset_repo,
                    repo_type=self.config.repo_type,
                )
                print(f"Checkpoint file deleted successfully")
                return
            except Exception as e:
                error_str = str(e)
                if "404" in error_str or "Entry Not Found" in error_str:
                    print(f"Checkpoint file already deleted or never existed")
                    return
                if attempt < max_retries - 1:
                    wait_time = 2 ** attempt
                    print(f"Warning: Could not delete checkpoint file (attempt {attempt + 1}/{max_retries}): {e}")
                    print(f"Retrying in {wait_time}s...")
                    time.sleep(wait_time)
                else:
                    print(f"Warning: Could not delete checkpoint file after {max_retries} attempts: {e}")
                    print(f"The checkpoint will be cleaned up during the next prune operation")
                    raise

    def _cleanup_stale_checkpoints(self) -> None:
        """Clean up stale checkpoint files from previous interrupted uploads.

        A checkpoint is considered stale if:
        - It is older than 24 hours
        - It has no corresponding backup archive uploaded
        """
        try:
            files = self.api.list_repo_files(
                repo_id=self.config.dataset_repo,
                repo_type=self.config.repo_type,
            )

            cutoff_time = utc_now() - dt.timedelta(hours=24)

            cleaned = 0
            for path in files:
                if not isinstance(path, str):
                    continue
                if not path.endswith(".checkpoint.json"):
                    continue

                try:
                    with tempfile.TemporaryDirectory() as tmp_dir:
                        checkpoint_local = self._download_backup(path, Path(tmp_dir))
                        with open(checkpoint_local, "r") as f:
                            checkpoint_data = json.load(f)
                        created_at = checkpoint_data.get("created_at_utc", "")
                        if created_at:
                            checkpoint_time = dt.datetime.fromisoformat(created_at)
                            if checkpoint_time < cutoff_time:
                                self.api.delete_file(
                                    path_in_repo=path,
                                    repo_id=self.config.dataset_repo,
                                    repo_type=self.config.repo_type,
                                )
                                print(f"[CLEANUP] Deleted stale checkpoint: {path}")
                                cleaned += 1
                except Exception:
                    match = re.search(r"(\d{8}-\d{6})", path)
                    if match:
                        try:
                            file_time = dt.datetime.strptime(match.group(1), "%Y%m%d-%H%M%S")
                            if file_time < cutoff_time:
                                self.api.delete_file(
                                    path_in_repo=path,
                                    repo_id=self.config.dataset_repo,
                                    repo_type=self.config.repo_type,
                                )
                                print(f"[CLEANUP] Deleted stale checkpoint (by filename): {path}")
                                cleaned += 1
                        except (ValueError, Exception):
                            pass

            if cleaned > 0:
                print(f"[CLEANUP] Cleaned up {cleaned} stale checkpoint(s)")
        except Exception as e:
            print(f"[CLEANUP] Warning: Could not clean up stale checkpoints: {e}")

    def _verify_upload(self, files_to_verify: list[str], repo_id: Optional[str] = None) -> None:
        """Verify that the uploaded files exist in the repository.

        Uses lightweight API checks for large files and selective download
        verification for small metadata files only.

        Args:
            files_to_verify: List of file paths to verify
            repo_id: Optional repository ID to verify in. If not specified, uses dataset_repo
        """
        files_to_verify = [f for f in files_to_verify if f]
        if not files_to_verify:
            raise RuntimeError("Upload verification failed: no files to verify")

        verify_repo = repo_id or self.config.dataset_repo

        time.sleep(2)

        try:
            repo_files = self.api.list_repo_files(
                repo_id=verify_repo,
                repo_type=self.config.repo_type,
            )
            repo_files_set = set(repo_files)

            for path in files_to_verify:
                if path not in repo_files_set:
                    raise RuntimeError(f"Upload verification failed: {path} not found in repository")

                # For small metadata files (< 1MB), do actual download verification
                # For large archive files, presence in list_repo_files is sufficient
                if path.endswith(".meta.json") or path.endswith("latest-backup.json"):
                    try:
                        with tempfile.TemporaryDirectory() as tmp_dir:
                            self.api.hf_hub_download(
                                repo_id=verify_repo,
                                filename=path,
                                local_dir=tmp_dir,
                                repo_type=self.config.repo_type,
                                token=self.token,
                            )
                    except Exception:
                        raise RuntimeError(f"Upload verification failed: {path} exists but cannot be downloaded")

                print(f"  Verified: {path}")

        except Exception as e:
            raise RuntimeError(f"Upload verification failed: {e}")

    def _list_backup_archive_paths(self) -> list[str]:
        if not hasattr(self.api, "list_repo_files"):
            return []

        files = self.api.list_repo_files(
            repo_id=self.config.dataset_repo,
            repo_type=self.config.repo_type,
        )
        path_prefix = self.config.path_prefix.strip("/")
        path_prefix_slash = f"{path_prefix}/" if path_prefix else ""
        
        # Group by backup timestamp, collect all paths for each backup
        backups_by_time: dict[str, list[str]] = {}

        for path in files:
            if not isinstance(path, str):
                continue

            relative_path = path
            if path_prefix_slash:
                if not path.startswith(path_prefix_slash):
                    continue
                relative_path = path[len(path_prefix_slash) :]

            if "/" in relative_path:
                continue

            match = self._ARCHIVE_NAME_RE.fullmatch(relative_path)
            if match is None:
                continue

            timestamp = match.group(1)
            if timestamp not in backups_by_time:
                backups_by_time[timestamp] = []
            backups_by_time[timestamp].append(path)

        # Sort backups by timestamp descending
        sorted_timestamps = sorted(backups_by_time.keys(), reverse=True)
        
        # Expand back to full list, keeping all volumes for each backup
        result = []
        for timestamp in sorted_timestamps:
            result.extend(backups_by_time[timestamp])

        return result

    def _find_meta_file_for_timestamp(self, timestamp: str, files: list, path_prefix_slash: str) -> str | None:
        """Find the meta.json file path for a backup timestamp.

        Args:
            timestamp: The backup timestamp (e.g., "20260425-013501")
            files: List of files in the repository
            path_prefix_slash: Path prefix with trailing slash

        Returns:
            The meta file path if found, None otherwise
        """
        for path in files:
            if not isinstance(path, str):
                continue
            if timestamp in path and path.endswith(".meta.json"):
                return path
        return None

    def prune_old_backups(self) -> list[str]:
        print("[PRUNE] Starting old backup cleanup...")
        print(f"[PRUNE] Retention policy: keep_count={self.config.keep_count}, min_full_backups=12")

        if not hasattr(self.api, "delete_file"):
            print("[PRUNE] API does not support delete_file, skipping cleanup")
            return []

        keep_count = self.config.keep_count if self.config.keep_count >= 1 else 48
        min_full_backups = 12

        if not hasattr(self.api, "list_repo_files"):
            print("[PRUNE] API does not support list_repo_files, skipping cleanup")
            return []

        print("[PRUNE] Fetching file list from remote repository...")
        files = self.api.list_repo_files(
            repo_id=self.config.dataset_repo,
            repo_type=self.config.repo_type,
        )
        print(f"[PRUNE] Found {len(files)} total files in repository")

        path_prefix = self.config.path_prefix.strip("/")
        path_prefix_slash = f"{path_prefix}/" if path_prefix else ""

        backups_by_time: dict[str, list[str]] = {}
        skipped_non_backup = 0

        for path in files:
            if not isinstance(path, str):
                continue

            relative_path = path
            if path_prefix_slash:
                if not path.startswith(path_prefix_slash):
                    continue
                relative_path = path[len(path_prefix_slash) :]

            if "/" in relative_path:
                continue

            match = self._ARCHIVE_NAME_RE.fullmatch(relative_path)
            if match is None:
                skipped_non_backup += 1
                continue

            timestamp = match.group(1)
            if timestamp not in backups_by_time:
                backups_by_time[timestamp] = []
            backups_by_time[timestamp].append(path)

        print(f"[PRUNE] Scanned files: {len(files)} total, {skipped_non_backup} non-backup files skipped")
        print(f"[PRUNE] Identified {len(backups_by_time)} backup timestamp(s)")

        if not backups_by_time:
            print("[PRUNE] No backups found, nothing to prune")
            return []

        for ts, paths in sorted(backups_by_time.items(), reverse=True):
            print(f"[PRUNE]   Backup {ts}: {len(paths)} volume(s)")

        sorted_timestamps = sorted(backups_by_time.keys(), reverse=True)

        full_backup_timestamps = set()
        for ts in sorted_timestamps:
            meta_file = self._find_meta_file_for_timestamp(ts, files, path_prefix_slash)
            if meta_file:
                try:
                    cached_meta = self._get_cached_meta(meta_file)
                    if cached_meta is not None:
                        meta = cached_meta
                    else:
                        local_meta = self._download_backup(meta_file, self.config.work_dir)
                        with open(local_meta, "r") as f:
                            meta = json.load(f)
                        self._set_cached_meta(meta_file, meta)
                    if meta.get("backup_type") == "full":
                        full_backup_timestamps.add(ts)
                except Exception:
                    pass

        print(f"[PRUNE] Identified {len(full_backup_timestamps)} full backup(s), ensuring minimum 12 are kept")

        if len(full_backup_timestamps) < min_full_backups:
            print(f"[PRUNE] Not enough full backups ({len(full_backup_timestamps)}) to meet minimum {min_full_backups}, skipping cleanup")
            return []

        kept_full_backups_list = sorted(full_backup_timestamps, reverse=True)[:min_full_backups]
        kept_full_backups = set(kept_full_backups_list)
        oldest_kept_full = kept_full_backups_list[-1] if kept_full_backups_list else None
        print(f"[PRUNE] Will keep {len(kept_full_backups)} newest full backups: {kept_full_backups_list[:3]}...")
        if oldest_kept_full:
            print(f"[PRUNE] Oldest kept full backup: {oldest_kept_full}")

        old_full_timestamps = [ts for ts in sorted_timestamps if ts in full_backup_timestamps and ts not in kept_full_backups]

        incremental_timestamps = [ts for ts in sorted_timestamps if ts not in full_backup_timestamps]
        timestamps_to_delete_candidates = list(reversed(old_full_timestamps))

        for ts in reversed(incremental_timestamps):
            if oldest_kept_full and ts >= oldest_kept_full:
                continue
            timestamps_to_delete_candidates.append(ts)

        timestamps_to_delete = timestamps_to_delete_candidates[:keep_count]

        if not timestamps_to_delete:
            print("[PRUNE] No backups to delete after applying retention policy")
            return []

        print(f"[PRUNE] Will delete {len(timestamps_to_delete)} backup(s): {len([t for t in timestamps_to_delete if t in full_backup_timestamps])} full, {len([t for t in timestamps_to_delete if t not in full_backup_timestamps])} incremental")
        print(f"[PRUNE] Oldest backup to delete: {timestamps_to_delete[-1]}")

        # Limit batch size to avoid rate limiting (max 50 files per cleanup)
        max_files_per_cleanup = 50
        total_files_to_delete = 0
        for ts in timestamps_to_delete:
            total_files_to_delete += len(backups_by_time[ts])
            # Add 1 for meta.json (approximate)
            total_files_to_delete += 1

        if total_files_to_delete > max_files_per_cleanup:
            # Limit to oldest backups first, but ensure at least one backup is cleaned
            limited_timestamps = []
            limited_count = 0
            for ts in reversed(timestamps_to_delete):  # Oldest first
                ts_files = len(backups_by_time[ts]) + 1
                if limited_count + ts_files <= max_files_per_cleanup or not limited_timestamps:
                    limited_timestamps.append(ts)
                    limited_count += ts_files
                else:
                    break
            timestamps_to_delete = list(reversed(limited_timestamps))  # Back to oldest-last order
            print(f"[PRUNE] Limiting cleanup to {len(timestamps_to_delete)} backup(s) ({limited_count} files) to avoid rate limiting")

        to_delete = []
        seen_paths = set()
        for timestamp in timestamps_to_delete:
            archive_paths = backups_by_time[timestamp]
            for archive_path in archive_paths:
                if archive_path not in seen_paths:
                    to_delete.append(archive_path)
                    seen_paths.add(archive_path)
            print(f"[PRUNE] Marking for deletion: backup {timestamp} ({len(archive_paths)} volume(s))")

            meta_added = False
            for archive_path in archive_paths:
                archive_name = Path(archive_path).name
                if ".part-" in archive_name:
                    base_name = archive_name.rsplit(".part-", 1)[0]
                else:
                    base_name = archive_name
                if base_name.endswith(".tar.gz"):
                    meta_name = base_name[:-7] + ".meta.json"
                else:
                    meta_name = base_name + ".meta.json"

                if path_prefix_slash:
                    meta_path = f"{path_prefix_slash}{meta_name}"
                else:
                    meta_path = meta_name

                if meta_path in files and meta_path not in seen_paths:
                    to_delete.append(meta_path)
                    seen_paths.add(meta_path)
                    if not meta_added:
                        print(f"[PRUNE]   + meta.json: {meta_name}")
                        meta_added = True

                # Also add checkpoint file if exists
                checkpoint_name = base_name + ".checkpoint.json"
                if path_prefix_slash:
                    checkpoint_path = f"{path_prefix_slash}{checkpoint_name}"
                else:
                    checkpoint_path = checkpoint_name
                if checkpoint_path in files and checkpoint_path not in seen_paths:
                    to_delete.append(checkpoint_path)
                    seen_paths.add(checkpoint_path)
                    print(f"[PRUNE]   + checkpoint.json: {checkpoint_name}")

        # Find orphaned checkpoint files (checkpoint without corresponding backup)
        orphaned_checkpoints = []
        for path in files:
            if not isinstance(path, str):
                continue
            if not path.endswith(".checkpoint.json"):
                continue
            # Extract base name from checkpoint
            checkpoint_name = Path(path).name
            base_name = checkpoint_name[:-len(".checkpoint.json")]
            # Check if corresponding backup exists
            has_backup = False
            for ts, archive_paths in backups_by_time.items():
                for ap in archive_paths:
                    ap_name = Path(ap).name
                    if ".part-" in ap_name:
                        ap_base = ap_name.rsplit(".part-", 1)[0]
                    else:
                        ap_base = ap_name
                    if ap_base == base_name:
                        has_backup = True
                        break
                if has_backup:
                    break
            if not has_backup and path not in seen_paths:
                orphaned_checkpoints.append(path)
                seen_paths.add(path)

        if orphaned_checkpoints:
            print(f"[PRUNE] Found {len(orphaned_checkpoints)} orphaned checkpoint(s)")
            for cp in orphaned_checkpoints:
                to_delete.append(cp)
                print(f"[PRUNE]   + orphaned checkpoint: {Path(cp).name}")

        print(f"[PRUNE] Total files to delete: {len(to_delete)}")

        prune_repo = self.config.dataset_repo
        deleted_count = 0
        skipped_count = 0
        rate_limit_hits = 0
        import time
        for path in to_delete:
            print(f"[PRUNE] Deleting ({deleted_count + skipped_count + 1}/{len(to_delete)}): {path}")
            retry_count = 0
            max_retries = 3
            while retry_count < max_retries:
                try:
                    self.api.delete_file(
                        repo_id=prune_repo,
                        repo_type=self.config.repo_type,
                        path_in_repo=path,
                        commit_message=f"backup: prune {Path(path).name}",
                    )
                    deleted_count += 1
                    break
                except Exception as e:
                    error_str = str(e)
                    if "404" in error_str or "Entry Not Found" in error_str:
                        print(f"[PRUNE]   ⚠ File already gone or never existed, skipping: {path}")
                        skipped_count += 1
                        break
                    elif "429" in error_str or "Too Many Requests" in error_str:
                        retry_count += 1
                        rate_limit_hits += 1
                        if retry_count >= max_retries:
                            print(f"[PRUNE]   ✗ Rate limit exceeded after {max_retries} retries, stopping cleanup: {path}")
                            print(f"[PRUNE] Cleanup interrupted due to rate limiting. Deleted {deleted_count} file(s), skipped {skipped_count} (not found), rate limit hits: {rate_limit_hits}")
                            return to_delete[:deleted_count + skipped_count]
                        wait_time = min(2 ** retry_count * 5, 60)
                        print(f"[PRUNE]   ⚠ Rate limit hit ({retry_count}/{max_retries}), waiting {wait_time}s before retry...")
                        time.sleep(wait_time)
                        continue
                    else:
                        print(f"[PRUNE]   ✗ Delete failed: {e}")
                        raise

        print(f"[PRUNE] Cleanup complete, deleted {deleted_count} file(s), skipped {skipped_count} (not found), rate limit hits: {rate_limit_hits}")
        return to_delete

    def _manage_local_backups(self, archive: Path | list[Path]) -> None:
        """Manage local backup files based on retention policy."""
        if not self.config.keep_local_backup:
            # Clean up temporary files in work_dir
            if isinstance(archive, list):
                for vol in archive:
                    if vol.exists():
                        vol.unlink()
                        print(f"Cleaned up temporary volume: {vol.name}")
            else:
                if archive.exists():
                    archive.unlink()
                    print(f"Cleaned up temporary archive: {archive.name}")
            return
        
        # Keep local backups - move to local_backup_dir if specified
        target_dir = self.config.local_backup_dir or self.config.work_dir
        target_dir.mkdir(parents=True, exist_ok=True)
        
        # Move/copy archives to target directory
        archives_to_keep = []
        if isinstance(archive, list):
            for vol in archive:
                target_path = target_dir / vol.name
                if target_dir != self.config.work_dir:
                    shutil.copy2(vol, target_path)
                    vol.unlink()
                archives_to_keep.append(target_path)
        else:
            target_path = target_dir / archive.name
            if target_dir != self.config.work_dir:
                shutil.copy2(archive, target_path)
                archive.unlink()
            archives_to_keep.append(target_path)
        
        # Prune old local backups
        self._prune_local_backups(target_dir)
        
        print(f"Local backup kept: {target_dir}")
    
    def _prune_local_backups(self, directory: Path) -> None:
        """Remove old local backups keeping only keep_local_count most recent."""
        # Find all backup archives in directory
        archives = sorted(
            directory.glob("openclaw-backup-*.tar.gz*"),
            key=lambda p: p.stat().st_mtime,
            reverse=True
        )
        
        # Group by base name (for split archives)
        base_names = {}
        for arch in archives:
            # Extract base name (remove .part-xx if present)
            base_name = arch.name
            if ".part-" in base_name:
                base_name = base_name.rsplit(".part-", 1)[0]
            
            if base_name not in base_names:
                base_names[base_name] = []
            base_names[base_name].append(arch)
        
        # Keep only keep_local_count most recent backup sets
        keep_count = self.config.keep_local_count
        sorted_bases = sorted(base_names.keys(), reverse=True)
        
        for base_name in sorted_bases[keep_count:]:
            for arch in base_names[base_name]:
                arch.unlink()
                print(f"Pruned old local backup: {arch.name}")

    def _cleanup_stale_lock(self, lock_path: Path) -> bool:
        """Clean up stale lock file from previous crashed process.

        Returns True if lock was cleaned up or no lock existed.
        Returns False if lock exists and is held by a live process.
        """
        if not lock_path.exists():
            return True
        try:
            with open(lock_path, "r") as f:
                lines = f.readlines()
            if not lines:
                lock_path.unlink(missing_ok=True)
                return True
            try:
                pid = int(lines[0].strip())
            except (ValueError, IndexError):
                lock_path.unlink(missing_ok=True)
                return True
            try:
                os.kill(pid, 0)
                return False
            except ProcessLookupError:
                print(f"[LOCK] Stale lock from dead process {pid}, cleaning up")
                lock_path.unlink(missing_ok=True)
                return True
            except OSError:
                return False
        except Exception:
            try:
                lock_path.unlink(missing_ok=True)
            except OSError:
                pass
            return True

    def _acquire_lock(self, lock_path: Path) -> Optional[int]:
        """Acquire exclusive backup lock with stale lock cleanup and timeout."""
        cleanup_result = self._cleanup_stale_lock(lock_path)
        if not cleanup_result:
            print("[LOCK] Lock file exists and is held by a live process, waiting...")
        lock_fd = os.open(str(lock_path), os.O_CREAT | os.O_RDWR)
        start_time = time.time()
        timeout = 300
        while True:
            try:
                fcntl.flock(lock_fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
                lock_uuid = str(uuid.uuid4())
                os.write(lock_fd, f"{os.getpid()}\n{utc_now().isoformat()}\n{lock_uuid}\n".encode())
                os.fsync(lock_fd)
                self._lock_uuid = lock_uuid
                return lock_fd
            except Exception as e:
                if time.time() - start_time > timeout:
                    os.close(lock_fd)
                    raise TimeoutError(f"Could not acquire backup lock after {timeout}s") from e
                time.sleep(1)

    def backup(self) -> Path | list[Path] | None:
        lock_path = self.config.work_dir / "openclaw-backup.lock"
        lock_path.parent.mkdir(parents=True, exist_ok=True)
        lock_fd = None
        result = False
        try:
            lock_fd = self._acquire_lock(lock_path)
            return self._backup_inner()
        except TimeoutError as e:
            print(f"[LOCK] {e}, backup skipped")
            return None
        finally:
            if lock_fd is not None:
                try:
                    fcntl.flock(lock_fd, fcntl.LOCK_UN)
                except (OSError, IOError):
                    pass
                os.close(lock_fd)
                try:
                    lock_path.unlink(missing_ok=True)
                except OSError:
                    pass

    def _backup_inner(self) -> Path | list[Path] | None:
        self.ensure_dataset_repo()

        if not self._verify_existing_backup():
            print("backup aborted: cannot verify existing backup is restorable", file=sys.stderr)
            print("  Please fix the restore issue before creating new backups.", file=sys.stderr)
            return None

        archive = self.create_archive()

        if archive is None:
            print("backup skipped by dynamic strategy")
            return None

        remote_path = self.upload_backup(archive)
        print(f"backup uploaded: {self.config.dataset_repo}/{remote_path}")

        self._manage_local_backups(archive)

        return archive

    def _verify_existing_backup(self) -> bool:
        """Verify that the existing latest backup metadata is valid.

        Note: This method only validates metadata, not the actual backup files.
        Full integrity verification is deferred to the restore process.

        Returns True if:
        - No existing backup (first backup)
        - Existing backup metadata is valid

        Returns False if:
        - Existing backup metadata is invalid or corrupted
        """
        print("Checking if existing backup metadata is valid...")
        print(f"  Checking backup target: {self.config.dataset_repo}")

        verify_repo = self.config.dataset_repo

        # Try to get metadata
        try:
            # latest-backup.json is ALWAYS at the dataset root (no path_prefix), not in backups/
            meta_filename = "latest-backup.json"
            print(f"  Downloading {meta_filename}...")
            meta_path = self._download_backup(meta_filename, self.config.work_dir)
            print(f"  Downloaded to {meta_path}")
            with open(meta_path, "r") as f:
                latest_meta = json.load(f)
            print(f"  Found backup metadata: {latest_meta.get('latest', 'unknown')}")

            # Now download the actual .meta.json file to get volumes and is_split
            actual_meta_path = latest_meta.get("latest", "")
            if not actual_meta_path:
                print("  No 'latest' entry found in latest-backup.json")
                return False

            print(f"  Downloading actual meta file: {actual_meta_path}...")
            cached_meta = self._get_cached_meta(actual_meta_path)
            if cached_meta is not None:
                backup_meta = cached_meta
                print(f"  Using cached metadata for {actual_meta_path}")
            else:
                actual_meta_local = self._download_backup(actual_meta_path, self.config.work_dir, repo_id=verify_repo)
                with open(actual_meta_local, "r") as f:
                    backup_meta = json.load(f)
                self._set_cached_meta(actual_meta_path, backup_meta)

            is_split = backup_meta.get("is_split", False)
            volumes = backup_meta.get("volumes", [])
            print(f"  backup_meta keys: {list(backup_meta.keys())}")
            print(f"  is_split: {is_split}")
            print(f"  volumes: {volumes}")

            # Validate that volumes are specified
            if not volumes:
                print("  Warning: No volumes found in backup metadata")
                return False

            # Verify that backup files exist in the repository using API (no download)
            print("  Verifying backup files exist in repository...")
            try:
                repo_files = self.api.list_repo_files(
                    repo_id=verify_repo,
                    repo_type=self.config.repo_type,
                )
                repo_files_set = set(repo_files)

                for vol_path in volumes:
                    # Add path_prefix if not already present
                    check_path = vol_path
                    if self.config.path_prefix and not vol_path.startswith(self.config.path_prefix):
                        check_path = f"{self.config.path_prefix}/{vol_path}"

                    if check_path not in repo_files_set:
                        print(f"  ✗ Backup file not found: {check_path}")
                        return False
                    print(f"  ✓ Backup file exists: {check_path}")

                print(f"  ✓ All {len(volumes)} backup file(s) verified to exist")
            except Exception as e:
                print(f"  Warning: Could not verify backup files exist: {e}")
                # Continue anyway, don't fail the backup due to API issues

            print("  Note: Full integrity verification will be performed during restore")
            return True

        except Exception as e:
            # No metadata means no existing backup (first backup)
            print(f"  No existing backup found (first backup): {type(e).__name__}: {e}")
            return True

    @staticmethod
    def replace_dir_contents(source_dir: Path, target_dir: Path) -> None:
        target_dir.mkdir(parents=True, exist_ok=True)

        for existing in target_dir.iterdir():
            try:
                if existing.is_dir() and not existing.is_symlink():
                    shutil.rmtree(existing)
                else:
                    existing.unlink()
            except (OSError, PermissionError):
                pass

        for item in source_dir.iterdir():
            destination = target_dir / item.name
            if item.is_symlink():
                try:
                    if destination.exists() or destination.is_symlink():
                        destination.unlink()
                except OSError:
                    pass
                try:
                    destination.symlink_to(os.readlink(item))
                except OSError:
                    pass
            elif item.is_dir():
                if destination.exists():
                    shutil.rmtree(destination)
                shutil.copytree(item, destination, symlinks=True)
            else:
                try:
                    shutil.copy2(item, destination)
                except FileExistsError:
                    try:
                        destination.unlink()
                        shutil.copy2(item, destination)
                    except Exception:
                        pass

    def _download_backup(self, remote_path: str, local_dir: Path, repo_id: Optional[str] = None) -> Path:
        """Download a specific backup from the dataset.

        First attempts to use hf_hub_download (with cache bypass), then falls back
        to direct HTTP download via urllib. Always ensures latest version is obtained.

        Uses in-memory caching for .meta.json files to avoid repeated downloads during
        the same session.

        Args:
            remote_path: Path to the file in the repository
            local_dir: Local directory to download to
            repo_id: Optional repository ID to download from. If not specified,
                    uses dataset_repo as default
        """
        # Determine which repo to download from
        # Default to dataset_repo (caller should explicitly pass repo_id for correct dataset)
        download_repo = repo_id or self.config.dataset_repo

        # Check in-memory cache first (for .meta.json files to avoid repeated downloads)
        if remote_path in self._meta_file_cache:
            cached_path = self._meta_file_cache[remote_path]
            if cached_path.exists():
                return cached_path

        local_dir = Path(local_dir)
        local_dir.mkdir(parents=True, exist_ok=True)
        local_file_path = local_dir / remote_path

        if local_file_path.exists():
            local_file_path.unlink()

        if hf_hub_download is not None:
            try:
                print(f"[_download_backup] Trying hf_hub_download for {remote_path} from {download_repo}...")
                downloaded_path = hf_hub_download(
                    repo_id=download_repo,
                    repo_type=self.config.repo_type,
                    filename=remote_path,
                    local_dir=str(local_dir),
                    force_download=True,
                    token=self.token,
                )
                print(f"[_download_backup] hf_hub_download succeeded: {downloaded_path}")
                result_path = Path(downloaded_path)
                if remote_path.endswith(".meta.json"):
                    self._meta_file_cache[remote_path] = result_path
                return result_path
            except Exception as e:
                print(f"[_download_backup] hf_hub_download failed: {type(e).__name__}: {e}")
                print(f"[_download_backup] Falling back to direct HTTP download...")

        print(f"[_download_backup] Using direct HTTP for {remote_path} from {download_repo}...")
        endpoint = f"https://huggingface.co/datasets/{download_repo}/resolve/main/{remote_path}"
        headers = {}
        if self.token:
            headers["Authorization"] = f"Bearer {self.token}"

        req = urllib.request.Request(endpoint, headers=headers)
        with urllib.request.urlopen(req, timeout=60) as response:
            local_file_path.write_bytes(response.read())

        print(f"[_download_backup] Direct HTTP download succeeded: {local_file_path}")
        if remote_path.endswith(".meta.json"):
            self._meta_file_cache[remote_path] = local_file_path
        return local_file_path

    def _clear_hf_cache(self, remote_path: str) -> None:
        """Clear Hugging Face Hub cache for a specific file to force re-download."""
        try:
            # Get the default cache directory
            cache_dir = Path.home() / ".cache" / "huggingface" / "hub"
            if not cache_dir.exists():
                return

            # Construct the repo cache directory name
            repo_name = self.config.dataset_repo.replace("/", "--")
            repo_cache_dir = cache_dir / f"datasets--{repo_name}"

            if not repo_cache_dir.exists():
                return

            # Find and remove cached files matching the remote_path
            # The cache structure is: blobs/, refs/, snapshots/
            # We need to find files in blobs/ that match our filename
            filename = Path(remote_path).name
            removed = False

            # Search for cached files
            for blob_file in repo_cache_dir.rglob("*"):
                if blob_file.is_file() and blob_file.name == filename:
                    blob_file.unlink()
                    print(f"    Cleared cache: {blob_file}")
                    removed = True

            if not removed:
                print(f"    No cache found for: {filename}")

        except Exception as e:
            print(f"    Warning: Could not clear cache: {e}")

    def _get_cached_meta(self, meta_path: str) -> Optional[dict]:
        """Get cached metadata JSON, or None if not cached."""
        return self._meta_cache.get(meta_path)

    def _set_cached_meta(self, meta_path: str, meta: dict) -> None:
        """Cache metadata JSON."""
        self._meta_cache[meta_path] = meta

    def should_restore(self) -> bool:
        """Determine if restore should be performed based on configuration.

        Restore logic:
        - Container restart/rebuild always triggers restore
        - This is by design to ensure data consistency

        Users can configure OPENCLAW_RESTORE_DATASET_REPO via HF Space variables
        to control restore behavior (defaults to OPENCLAW_BACKUP_DATASET_REPO).

        Returns:
            True (restore is always performed on container start)
        """
        print("Restore will be performed (container restart/rebuild detected)")
        return True

    def _format_size(self, size_bytes: int) -> str:
        """Format byte size to human readable string."""
        for unit in ["B", "KB", "MB", "GB", "TB"]:
            if size_bytes < 1024:
                return f"{size_bytes:.1f} {unit}"
            size_bytes /= 1024
        return f"{size_bytes:.1f} PB"

    def _merge_directory(self, source: Path, target: Path) -> None:
        """Merge source directory into target directory.

        For incremental backups: newer files overwrite older ones.
        """
        target.mkdir(parents=True, exist_ok=True)

        for item in source.rglob("*"):
            if item.is_file():
                rel_path = item.relative_to(source)
                dest = target / rel_path
                dest.parent.mkdir(parents=True, exist_ok=True)
                try:
                    shutil.copy2(item, dest)
                except FileExistsError:
                    dest.unlink()
                    shutil.copy2(item, dest)
            elif item.is_dir():
                rel_path = item.relative_to(source)
                dest = target / rel_path
                dest.mkdir(parents=True, exist_ok=True)

    def restore(self) -> bool:
        """Restore from backup, automatically handling incremental backup chains.

        Flow according to documentation:
        1. Download latest-backup.json to get latest .meta.json path
        2. Parse backup chain: download all .meta.json files, trace back to full backup
        3. Download archives in order (from full to latest) and merge
        4. Restore files to target directory
        """
        lock_path = self.config.work_dir / "openclaw-backup.lock"
        lock_path.parent.mkdir(parents=True, exist_ok=True)

        lock_fd = None
        result = False
        restore_error_msg = None
        skip_flag = False
        try:
            lock_fd = self._acquire_lock(lock_path)
            if lock_fd is None:
                print("restore skipped: a backup or restore is already in progress")
                skip_flag = True
                return False

            os.write(lock_fd, f"{os.getpid()}\n{utc_now().isoformat()}\nrestore\n".encode())
            os.fsync(lock_fd)

            result = self._restore_inner()
        except TimeoutError as e:
            print(f"[LOCK] {e}, restore skipped")
            restore_error_msg = str(e)
            result = False
        except Exception as e:
            print(f"[RESTORE] Unexpected error during restore: {type(e).__name__}: {e}", file=sys.stderr)
            restore_error_msg = f"{type(e).__name__}: {e}"
            result = False
        finally:
            if lock_fd is not None:
                try:
                    fcntl.flock(lock_fd, fcntl.LOCK_UN)
                except (OSError, IOError):
                    pass
                try:
                    os.close(lock_fd)
                except OSError:
                    pass

            try:
                lock_path.unlink(missing_ok=True)
            except Exception:
                pass

            if not skip_flag:
                restore_completed_flag = Path("/tmp/openclaw-restore-completed")
                restore_completed_flag.touch()
                print(f"Created restore completion flag: {restore_completed_flag}", file=sys.stderr)

        return result

    def _restore_inner(self) -> bool:
        print("Starting restore process...")
        restore_repo = self._get_restore_repo()
        print(f"Restore source: {restore_repo}")
        print(f"Target: {self.config.backup_source_dir}")

        # Step 1: Determine starting point
        # Check if user specified a specific backup via environment variable
        specified_archive = os.getenv("OPENCLAW_RESTORE_ARCHIVE")
        if specified_archive:
            # User specified a backup archive name, add .meta.json suffix
            meta_path = f"{specified_archive}.meta.json"
            if self.config.path_prefix:
                meta_path = f"{self.config.path_prefix}/{meta_path}"
            print(f"Restoring specified backup: {specified_archive}")
        else:
            # Step 1: Download latest-backup.json from dataset root (always at root, regardless of path_prefix)
            latest_json_path = "latest-backup.json"
            try:
                latest_json_local = self._download_backup(latest_json_path, self.config.work_dir, repo_id=restore_repo)
                with open(latest_json_local, "r") as f:
                    latest_json = json.load(f)
                meta_path = latest_json.get("latest", "")
                print(f"Latest backup: {meta_path}")
            except Exception as e:
                print(f"Warning: latest-backup.json not found in dataset '{restore_repo}': {e}", file=sys.stderr)
                print(f"This appears to be a newly initialized dataset", file=sys.stderr)
                restore_skipped_flag = Path("/tmp/openclaw-restore-skipped-no-backup")
                restore_skipped_flag.touch()
                restore_completed_flag = Path("/tmp/openclaw-restore-completed")
                restore_completed_flag.touch()
                print(f"Created restore skipped flag: {restore_skipped_flag}", file=sys.stderr)
                return True

        # Step 2: Parse backup chain (download all .meta.json, trace back to full backup)
        print("\n=== Step 2: Parsing backup chain ===")
        chain_meta_paths = self._parse_backup_chain(meta_path, repo_id=restore_repo)

        if not chain_meta_paths:
            print("Error: Could not determine backup chain", file=sys.stderr)
            return False

        if len(chain_meta_paths) == 1:
            print(f"Single backup (full): {chain_meta_paths[0]}")
        else:
            print(f"Backup chain with {len(chain_meta_paths)} backup(s):")
            for i, mp in enumerate(chain_meta_paths, 1):
                backup_type = "full" if i == 1 else "incremental"
                print(f"  [{i}] {mp} ({backup_type})")

        # Step 3: Download archives in order and merge
        print("\n=== Step 3: Downloading and merging archives ===")
        extra_dirs = self._restore_root_dirs()
        extra_files = self.config.extra_files

        if extra_dirs:
            print("Extra directories to restore:")
            for archive_root, local_dir in extra_dirs.items():
                print(f"  - {archive_root} -> {local_dir}")

        if extra_files:
            print("Extra files to restore:")
            for archive_name, local_file in extra_files.items():
                print(f"  - {archive_name} -> {local_file}")

        restored_items = []

        with tempfile.TemporaryDirectory(prefix="openclaw-restore-") as tmp:
            tmp_dir = Path(tmp)

            # Download and merge archives
            merged_state = self._download_and_merge_chain(chain_meta_paths, tmp_dir, repo_id=restore_repo)

            if not merged_state.exists() or not merged_state.is_dir():
                print("Error: Merged state directory missing", file=sys.stderr)
                return False

            # Step 4: Restore files to target directory
            print("\n=== Step 4: Restoring files ===")

            # Restore main state
            print(f"Restoring state to: {self.config.backup_source_dir}")
            self.replace_dir_contents(merged_state / "openclaw-state", self.config.backup_source_dir)
            restored_items.append(f"state -> {self.config.backup_source_dir}")

            # Restore extra directories
            for archive_root, local_dir in extra_dirs.items():
                extracted_extra_dir = merged_state / archive_root
                if extracted_extra_dir.exists():
                    if not extracted_extra_dir.is_dir():
                        print(f"Warning: backup archive {archive_root} is not a directory, skipping")
                        continue
                    print(f"Restoring {archive_root} to: {local_dir}")
                    self.replace_dir_contents(extracted_extra_dir, local_dir)
                    restored_items.append(f"{archive_root} -> {local_dir}")

            # Restore extra files
            for archive_name, local_file in extra_files.items():
                extracted_file = merged_state / archive_name
                if extracted_file.exists() and extracted_file.is_file():
                    local_file.parent.mkdir(parents=True, exist_ok=True)
                    shutil.copy2(str(extracted_file), str(local_file))
                    print(f"Restored extra file: {archive_name} -> {local_file}")
                    restored_items.append(f"{archive_name} -> {local_file}")

        print("")
        restore_repo = self._get_restore_repo()
        print(f"✓ Restore completed successfully from {restore_repo}")
        print(f"  Chain: {len(chain_meta_paths)} backup(s) merged")
        print("  Restored items:")
        for item in restored_items:
            print(f"    - {item}")

        restored_metadata = merged_state / "backup-info" / "last-backup-metadata.json"
        if restored_metadata.exists():
            self.config.metadata_dir.mkdir(parents=True, exist_ok=True)
            target_metadata = self.config.metadata_dir / "last-backup-metadata.json"
            shutil.copy2(str(restored_metadata), str(target_metadata))
            print(f"  Copied metadata to: {target_metadata}")

        print("")
        print("Backup restore completed successfully")
        return True

    def _parse_backup_chain(self, latest_meta_path: str, repo_id: Optional[str] = None) -> list[str]:
        """Parse backup chain by following parent pointers in .meta.json files.

        Args:
            latest_meta_path: Path to the latest .meta.json file (with path_prefix)
            repo_id: Optional repository ID to download from. Uses dataset_repo if not specified.

        Returns:
            List of .meta.json paths in order from full backup to latest
        """
        download_repo = repo_id or self.config.dataset_repo
        chain = []
        current_path = latest_meta_path
        visited = set()
        max_chain_length = 100

        chain_id = None
        latest_time = None

        local_meta = self._download_backup(latest_meta_path, self.config.work_dir, repo_id=download_repo)
        with open(local_meta, "r") as f:
            latest_meta = json.load(f)
        latest_time_str = latest_meta.get("created_at_utc") or latest_meta.get("last_backup_time")
        if latest_time_str:
            try:
                latest_time = dt.datetime.fromisoformat(latest_time_str.replace("Z", "+00:00"))
            except (ValueError, AttributeError):
                latest_time = None

        while current_path and len(chain) < max_chain_length:
            if current_path in visited:
                print(f"Warning: Circular reference detected in backup chain at {current_path}")
                break
            visited.add(current_path)

            chain.append(current_path)

            # Download and parse this .meta.json (use cache if available)
            try:
                cached_meta = self._get_cached_meta(current_path)
                if cached_meta is not None:
                    meta = cached_meta
                else:
                    local_meta = self._download_backup(current_path, self.config.work_dir, repo_id=download_repo)
                    with open(local_meta, "r") as f:
                        meta = json.load(f)
                    self._set_cached_meta(current_path, meta)
            except Exception as e:
                print(f"Error: Could not download/parse {current_path}: {e}", file=sys.stderr)
                break

            # Track chain_id from first meta
            if chain_id is None:
                chain_id = meta.get("chain_id")

            # Get parent pointer
            parent = meta.get("parent")
            if not parent:
                # Full backup, end of chain
                if meta.get("backup_type") == "full":
                    # Normal full backup, end of chain
                    break
                else:
                    # Incremental backup with null parent - chain is broken (parent field lost)
                    # Need to find the full backup by chain_id
                    print(f"Warning: Incremental backup has null parent (chain may be broken). chain_id={chain_id}")
                    break

            # Build full path for parent (may need path_prefix)
            if self.config.path_prefix and not parent.startswith(self.config.path_prefix):
                parent = f"{self.config.path_prefix}/{parent}"

            current_path = parent

        if len(chain) >= max_chain_length:
            print(f"Warning: Chain length exceeded safety limit ({max_chain_length})")

        # If chain is broken (only incremental with null parent), try to find full backup
        if len(chain) == 1 and chain_id:
            first_meta_path = chain[0]
            try:
                local_meta = self._download_backup(first_meta_path, self.config.work_dir, repo_id=download_repo)
                with open(local_meta, "r") as f:
                    first_meta = json.load(f)
                if first_meta.get("backup_type") == "incremental" and not first_meta.get("parent"):
                    print(f"Attempting to rebuild chain using chain_id={chain_id}...")
                    all_backups = self._find_all_backups_by_chain_id(chain_id, repo_id=download_repo, latest_time=latest_time)
                    if all_backups:
                        print(f"Found {len(all_backups)} backup(s) in chain")
                        # all_backups is already sorted: [full, inc-1, inc-2, ...]
                        chain = all_backups
                        # Already in correct order (full first), no reverse needed
                    else:
                        print(f"Warning: Could not find backups for chain_id={chain_id}")
                        print(f"  This incremental backup may not restore correctly without its full backup")
            except Exception as e:
                print(f"Warning: Could not check first meta for chain rebuild: {e}")
        else:
            # Normal case: reverse the chain built by following parent pointers
            chain.reverse()

        return chain

    def _find_all_backups_by_chain_id(self, chain_id: str, repo_id: Optional[str] = None, latest_time: Optional[dt.datetime] = None) -> Optional[list[str]]:
        """Find all backups (full + incremental) for a given chain_id and sort by time.

        Args:
            chain_id: The chain ID to search for
            repo_id: Optional repository ID to search in. Uses dataset_repo if not specified.
            latest_time: Optional datetime to filter backups - only backups at or before this time are included

        Returns:
            List of .meta.json paths sorted by created_at_utc (earliest first), or None if no full backup found
        """
        download_repo = repo_id or self.config.dataset_repo

        try:
            from huggingface_hub import HfApi

            token = self.token
            api = HfApi(token=token)

            files = api.list_repo_files(repo_id=download_repo, repo_type=self.config.repo_type)
            meta_files = [f for f in files if f.endswith(".meta.json")]

            backups_with_time = []
            full_backup_path = None

            for meta_file in meta_files:
                try:
                    cached_meta = self._get_cached_meta(meta_file)
                    if cached_meta is not None:
                        meta = cached_meta
                    else:
                        local_meta = self._download_backup(meta_file, self.config.work_dir, repo_id=download_repo)
                        with open(local_meta, "r") as f:
                            meta = json.load(f)
                        self._set_cached_meta(meta_file, meta)

                    if meta.get("chain_id") == chain_id:
                        created_at = meta.get("created_at_utc") or meta.get("last_backup_time")
                        if created_at:
                            try:
                                if isinstance(created_at, str):
                                    created_at = dt.datetime.fromisoformat(created_at.replace("Z", "+00:00"))
                            except (ValueError, AttributeError):
                                created_at = None

                        if latest_time and created_at and created_at > latest_time:
                            continue

                        if meta.get("backup_type") == "full":
                            full_backup_path = (meta_file, created_at, meta)
                        else:
                            backups_with_time.append((meta_file, created_at, meta))
                except Exception:
                    continue

            if not full_backup_path:
                print(f"Warning: No full backup found for chain_id={chain_id}")
                return None

            backups_with_time.sort(key=lambda x: x[1] if x[1] else dt.datetime.min)

            chain = [full_backup_path[0]]
            chain.extend([b[0] for b in backups_with_time])

            return chain

        except Exception as e:
            print(f"Warning: Could not search for backups: {e}")
            return None

    def _download_and_merge_chain(self, chain_meta_paths: list[str], tmp_dir: Path, repo_id: Optional[str] = None) -> Path:
        """Download archives in chain order and merge them.

        Args:
            chain_meta_paths: List of .meta.json paths (from full to latest)
            tmp_dir: Temporary directory for downloads and extraction
            repo_id: Optional repository ID to download from. Uses dataset_repo if not specified.

        Returns:
            Path to merged state directory
        """
        # Default to dataset_repo if not specified
        download_repo = repo_id or self.config.dataset_repo
        merged_dir = tmp_dir / "merged-state"
        merged_dir.mkdir(parents=True, exist_ok=True)

        print(f"Merging {len(chain_meta_paths)} backup(s) in chain...")

        total_size = 0
        successful_backups = 0
        for i, meta_path in enumerate(chain_meta_paths):
            backup_type = "full" if i == 0 else "incremental"
            print(f"  [{i+1}/{len(chain_meta_paths)}] Processing: {meta_path} ({backup_type})")

            # Download and parse .meta.json (use cache if available)
            try:
                cached_meta = self._get_cached_meta(meta_path)
                if cached_meta is not None:
                    meta = cached_meta
                else:
                    local_meta = self._download_backup(meta_path, tmp_dir, repo_id=download_repo)
                    with open(local_meta, "r") as f:
                        meta = json.load(f)
                    self._set_cached_meta(meta_path, meta)
            except Exception as e:
                print(f"Error: Could not download/parse {meta_path}: {e}", file=sys.stderr)
                if i == 0:
                    raise RuntimeError(f"Failed to download metadata for full backup: {meta_path}")
                print(f"  Warning: Stopping backup chain at {meta_path}, using {successful_backups} successful backup(s)")
                break

            is_split = meta.get("is_split", False)
            is_encrypted = meta.get("encrypted", False)
            volumes = meta.get("volumes", [])
            expected_checksum = meta.get("checksum", "")

            # Download archive(s)
            archive_path = None
            try:
                if is_split and volumes:
                    # Download all split volumes and merge
                    print(f"      Downloading {len(volumes)} split volumes...")
                    local_volumes = []
                    for vol_path in volumes:
                        # Add path_prefix if not already present
                        vol_remote_path = vol_path
                        if self.config.path_prefix and not vol_path.startswith(self.config.path_prefix):
                            vol_remote_path = f"{self.config.path_prefix}/{vol_path}"
                        local_vol = self._download_backup(vol_remote_path, tmp_dir, repo_id=download_repo)
                        local_volumes.append(local_vol)

                    # Merge volumes
                    # For encrypted splits, merged file should have .enc extension for proper decryption
                    if is_encrypted:
                        merged_archive = tmp_dir / f"merged-{i}.tar.gz.enc"
                    else:
                        merged_archive = tmp_dir / f"merged-{i}.tar.gz"
                    with open(merged_archive, "wb") as outfile:
                        for vol in sorted(local_volumes):
                            with open(vol, "rb") as infile:
                                outfile.write(infile.read())
                    archive_path = merged_archive
                    print(f"      Merged {len(local_volumes)} volumes into {merged_archive.name}")
                else:
                    # Single file archive
                    if volumes:
                        # Add path_prefix if not already present
                        vol_remote_path = volumes[0]
                        if self.config.path_prefix and not volumes[0].startswith(self.config.path_prefix):
                            vol_remote_path = f"{self.config.path_prefix}/{volumes[0]}"
                        archive_path = self._download_backup(vol_remote_path, tmp_dir, repo_id=download_repo)
                    else:
                        raise RuntimeError(f"No volumes found in metadata for {meta_path}")

                # Verify checksum for all archives BEFORE decryption
                # For split backups, verify the merged archive checksum if available
                if expected_checksum:
                    try:
                        actual_checksum = self._calculate_checksum(archive_path)
                        if actual_checksum != expected_checksum.replace("sha256:", ""):
                            raise RuntimeError(
                                f"Checksum mismatch for {archive_path.name}: "
                                f"expected {expected_checksum}, got sha256:{actual_checksum}"
                            )
                        print(f"      ✓ Checksum verified: {archive_path.name}")
                    except Exception as e:
                        print(f"  Error: Integrity verification failed for {meta_path}: {e}", file=sys.stderr)
                        if i == 0:
                            raise RuntimeError(f"Full backup integrity check failed: {meta_path}")
                        print(f"  Warning: Stopping backup chain at {meta_path}, using {successful_backups} successful backup(s)")
                        break
                elif is_split:
                    print(f"      ⚠ Checksum not available for split backup, skipping verification (create new backups to enable)")

                if is_encrypted:
                    if not self.config.encryption_password:
                        raise RuntimeError(
                            f"Backup {meta_path} is encrypted but OPENCLAW_BACKUP_ENCRYPTION_PASSWORD is not set. "
                            f"Cannot restore encrypted backup without password."
                        )
                    decrypted_archive = tmp_dir / f"decrypted-{i}.tar.gz"
                    archive_path = self._decrypt_archive(archive_path, decrypted_archive)
            except Exception as e:
                print(f"  Error: Failed to download archive for {meta_path}: {e}", file=sys.stderr)
                if i == 0:
                    raise RuntimeError(f"Failed to download full backup: {meta_path}")
                print(f"  Warning: Stopping backup chain at {meta_path}, using {successful_backups} successful backup(s)")
                break

            # Checksum verification is now done BEFORE decryption (above)
            # This ensures all archive types (encrypted, split, both) are verified consistently
            # For split backups without expected_checksum, a warning was already printed above

            archive_size = archive_path.stat().st_size
            total_size += archive_size
            print(f"      Archive: {archive_path.name}, size: {self._format_size(archive_size)}")

            # Extract
            extract_dir = tmp_dir / f"extract-{i}"
            extract_dir.mkdir(parents=True, exist_ok=True)

            try:
                with tarfile.open(archive_path, "r:gz") as tar:
                    tar.extractall(extract_dir)
            except tarfile.ReadError as e:
                print(f"  Error: Failed to extract archive {archive_path}: {e}", file=sys.stderr)
                if i == 0:
                    raise RuntimeError(f"Failed to extract full backup: {meta_path}")
                print(f"  Warning: Stopping backup chain at {meta_path}, using {successful_backups} successful backup(s)")
                break

            # Count files
            file_count = sum(1 for _ in extract_dir.rglob("*") if _.is_file())
            print(f"      Files: {file_count}")

            # Merge into merged directory
            state_dir = extract_dir / "openclaw-state"
            if state_dir.exists():
                self._merge_directory(state_dir, merged_dir / "openclaw-state")

            # Merge extra directories
            for archive_root, _ in self._restore_root_dirs().items():
                extra_dir = extract_dir / archive_root
                if extra_dir.exists():
                    self._merge_directory(extra_dir, merged_dir / archive_root)

            successful_backups += 1

        if successful_backups == 0:
            raise RuntimeError("No backups could be restored successfully")

        print(f"  Total downloaded: {self._format_size(total_size)}")
        print(f"  Successfully processed {successful_backups}/{len(chain_meta_paths)} backup(s)")
        return merged_dir


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description="Backup/restore OpenClaw state with Hugging Face Dataset")
    parser.add_argument("action", choices=["backup", "restore"], help="action to perform")
    return parser


def main(argv: Optional[list[str]] = None) -> int:
    args = build_parser().parse_args(argv)

    try:
        config = BackupConfig.from_env()
    except ValueError as exc:
        print(f"skip {args.action}: {exc}", file=sys.stderr)
        return 0

    try:
        runner = OpenClawBackup(config=config)
        if args.action == "backup":
            runner.backup()
            return 0

        # Restore action - always perform restore on container start
        if args.action == "restore":
            # Perform restore (container restart/rebuild always triggers restore)
            success = runner.restore()
            return 0 if success else 1

        return 0
    except Exception as exc:  # pragma: no cover
        print(f"{args.action} failed: {exc}", file=sys.stderr)
        return 1


if __name__ == "__main__":
    raise SystemExit(main())