Rclone Drive Backup

When to use

Load this skill when the user asks to:

• Upload, sync, or snapshot local directories to Google Drive via rclone
• Refresh md-plain mirrors, tree copies, or snapshots/ tarballs on Drive
• Check % complete of an in-flight rclone upload
• Debug why a Drive destination is stale or missing

Procedure

1. Pre-flight: remote health + what is already there

rclone lsd gdrive: --max-depth 2          # remote reachable?
rclone about gdrive:                        # quota headroom?
rclone lsl gdrive:SWAL/snapshots/ | tail    # date of last snapshot upload
rclone lsd gdrive:SWAL/tree/apps/           # date of last tree sync
rclone lsd gdrive:SWAL/md-plain/            # 'directory not found' = mirror never ran

Compare remote dates against local scope before choosing full vs incremental: a mirror dir that does not exist means that pipeline stage never completed, not that it is up to date.

2. Verify the filter file BEFORE uploading

rclone lsf is NON-recursive by default — a bare lsf shows only top-level entries and hides what the filter really passes. Always verify with:

cd <same-root-the-sync-runs-from>
rclone lsf -R --files-only <source-root> --filter-from <filter-file> | wc -l
rclone lsf -R --files-only <source-root> --filter-from <filter-file> | grep -v '<allowed-extension>$' | head

Run from the SAME root the sync uses, or **/ patterns match differently and the check is worthless. The second command must print nothing: any non-allowed extension in output means the filter leaks.

3. Launch long uploads in background and confirm liveness immediately

Launch with terminal(background=true, notify=true), then poll the session at once: background launches can die silently, and a missing process only shows up if you check. Capture the session_id — it is the only handle for later polls.

4. Monitor via pgrep + remote listing, never via piped output

Do NOT pipe the long command through tail: tail buffers until the pipe closes, so process polls show zero progress for hours. Monitor instead with:

pgrep -af "rclone copy" | grep -v hermes-snap   # in-flight transfer = current repo
rclone lsd gdrive:SWAL/md-plain/ | wc -l         # completed units

5. Report % as completed-units / total-units + current unit + ETA

ls -d <local-source>/*/ | wc -l    # total units

Heavy repos stall the % for a long time while the process is healthy — report the in-flight repo name and ETA alongside the %, not the % alone. Answer repeated "que % falta?" checks with fresh remote listings, never from memory of the last check.

6. Completion evidence (all required before claiming done)

• Background process exited (poll shows completed, or pgrep empty).
• Fresh (today) timestamps in EVERY destination stage (rclone lsl on snapshots, rclone lsd on mirrors).
• Report: files/bytes transferred per stage + remote paths.

Standing gates

• Gate multi-GB rclone syncs on unmetered network. On mobile/metered data, run snapshot (tarball) uploads only, or ask first — bulk small-file syncs are throttled to KiB/s and burn the data budget. When the user announces wifi is back, update the network-regime memory entry the same turn the upload is re-enabled.
• Never commit secrets: rclone.conf contains OAuth tokens — redact token/client_secret in any output you show.