Clean Up bd Auto-Backups | Gas City Español

Overview

bd (the beads CLI) maintains a Dolt-native backup of every workspace it touches. The backup writes to <root>/.beads/backup/ on a hardcoded remote named backup_export, throttled to one sync per fifteen minutes per bd process (gastownhall/beads: cmd/bd/backup_auto.go).

The backup remote is append-only. There is no retention, rotation, or garbage-collection logic in the bd auto-backup path today; the architectural redesign for snapshot collections is tracked at gastownhall/beads#2993. Over weeks of normal use the directory grows monotonically — and in busy multi-agent cities (where every orchestrator, agent, and order touches bd), the growth is fast enough to fill the disk in days.

gc doctor emits a bd-backup-size warning at 5 GB and an error at 15 GB to catch this before the next cascade. The canary checks the city workspace and managed rig scope roots, reporting the largest individual .beads/backup/ directory and the aggregate backup footprint.

Why it matters — the disk-full cascade

bd auto-backup's unbounded growth was the proximate root cause of a multi-hour outage on 2026-05-20/21 (qlandia incident, qlandia/gc-p831i):

.beads/backup/ hit 34 GB and filled the disk.
Dolt's mmap writes started returning ENOSPC.
Dolt entered a 95% CPU retry loop that did not recover when disk was freed.
Gas City's supervisor reconciler wedged on Dolt writes; named agents stuck stopped; recovery flags stuck.
Recovery required SIGTERM-to-dolt, supervisor restart, plist regeneration, and session-kill of every named-always agent — two days of disruption.

The gas-city-side amplifier (auto-recover-on-ENOSPC) was fixed in gastownhall/gascity#2347. This doc covers the still-live bd-side root cause.

Inspect

When gc doctor flags bd-backup-size, first inspect the directory:

du -sh ~/qlandia/.beads/backup       # total size
ls -lh ~/qlandia/.beads/backup       # file mix (.darc vs nbs_table_*)
cat ~/qlandia/.beads/backup/backup_state.json  # last sync watermark

You're looking for:

A current last_dolt_commit in backup_state.json (within minutes of now). If the backup is current, you can rotate it safely.
A mix of .darc archive files (one per conjoined chunk) and nbs_table_* files (raw uncompacted writes). Many nbs_table_* files indicate that conjoin has been failing — see "Atomicity / sync failure" below.

Cleanup options

Option A — disable bd auto-backup (least invasive)

If the only consumer of .beads/backup/ was the implicit auto-backup, turn it off and let mol-dog-backup (which writes to .dolt-backup/) handle backups instead:

bd config set backup.enabled false

After this, you can rm -rf .beads/backup/. Verify next gc doctor run that bd-backup-size reports no bd backup directory present.

You retain coverage from mol-dog-backup writing to .dolt-backup/, which is run by the dolt pack on a 6h cooldown and is operator-managed through gas-city.

Option B — rotate the backup (preserve coverage)

Use this when you want to keep auto-backup running but reclaim the accumulated bloat.

Risk: bd auto-backup is non-atomic (gastownhall/beads#4070). Mid-sync interruption can leave a broken backup. Schedule rotation at a quiet moment and verify the result.

# 1. Stop the supervisor so no agent is writing beads.
gc stop

# 2. Capture the existing backup as a safety net.
mv ~/qlandia/.beads/backup ~/qlandia/.beads/backup.old-$(date +%Y%m%d)

# 3. Trigger bd's auto-backup pipeline while the supervisor remains
#    stopped. Do not use `bd backup sync` here; that syncs the separate
#    explicit "default" backup remote, not the auto-backup `backup_export`
#    remote documented above.
bd -C ~/qlandia list --status open --limit 1 --json >/tmp/qlandia-bd-autobackup-smoke.json

# 4. Verify the new backup is sane.
du -sh ~/qlandia/.beads/backup        # should be ≲ source database size
cat ~/qlandia/.beads/backup/backup_state.json  # current timestamp

# 5. Restart the supervisor only after the first backup is known good.
gc start

# 6. After confirming a clean run for ~1 day, delete the safety net.
rm -rf ~/qlandia/.beads/backup.old-*

Option C — operator wipe (when the database has shrunk hard)

After running dolt gc --full on the source database (see dolt-bloat-recovery.md), the source's noms directory shrinks but the backup remote does not. The full backup chain — every chunk ever written — is still on disk in .beads/backup/.

If you've just run a source GC and have a verified safety copy of the data, Option B's wipe-and-resync is the only mechanism that propagates the GC into the backup.

Atomicity / sync failure

If you see every bd command print

Warning: auto-backup failed: sync to backup: sync backup backup_export:
Error 1105: error opening table file: table file not found:
/path/to/.beads/backup/<chunk-hash>

you've hit gastownhall/beads#4070: the backup manifest references chunks that were never delivered. Manual repair until upstream lands a fix:

# Identify missing chunks
python3 -c "
import os, sys
manifest = open(os.path.expanduser('~/qlandia/.beads/backup/manifest')).read()
parts = manifest.strip().split(':')
chunks = [parts[i] for i in range(5, len(parts), 2)]
existing = set()
for f in os.listdir(os.path.expanduser('~/qlandia/.beads/backup/')):
    if f.endswith('.darc'):
        existing.add(f[:-len('.darc')])
missing = [c for c in chunks if c not in existing]
for c in missing:
    print(c)
"
# If nbs_table_* files are present, treat the output as a candidate
# list and cross-check with bd backup status or Dolt tooling before
# copying. Raw table files can contain valid chunks that are not visible
# as standalone .darc files.
# Then copy each missing chunk from the live noms store:
#   cp ~/qlandia/.beads/dolt/<db>/.dolt/noms/<hash> ~/qlandia/.beads/backup/

Prevention

Track gastownhall/beads#2993. The architectural fix is snapshot-collection semantics with explicit retention. Until it lands, every bd-driven city accumulates this bloat.
Run gc doctor regularly. The bd-backup-size canary fires at 5 GB warn, 15 GB error — well before disk pressure breaks Dolt.
Prefer mol-dog-backup's .dolt-backup/ path for primary backup coverage. It's operator-managed through gas-city, runs on a 6h cooldown rather than per-command, and is the one targeted by retention work in gastownhall/gascity.
Watch gc doctor for dolt-noms-size alongside this check. Both surfaces grow together on busy cities; treating them as a single disk-pressure signal works better than reasoning about each in isolation.