maintenance

/matts-second-brain:maintenance — the nightly self-consolidation job

A local cron runs maintenance/run.mjs, a tested harness that launches a headless claude -p constrained to the mcp__second-brain__* tools to: scan recent transcripts, recall + persist genuinely durable knowledge (deduping, not duplicating), and reindex for semantic search. It is local because the brain backend (e.g. Trilium) is local — a cloud /schedule routine could not reach it.

Resolve paths + config first

• Plugin dir: ${CLAUDE_PLUGIN_ROOT} if set, else the repo root containing this skill.
• Maintenance dir: <plugin>/maintenance.
• Read the plugin userConfig from ~/.claude.json (pluginConfigs[*matts-second-brain*].options), falling back to the manifest defaults:
  • scheduledMaintenance (default false), maintenanceSchedule (0 4 * * *), maintenanceModel (claude-sonnet-4-6), scanSinceDays (2), maintenanceTimeoutMin (10).

Subcommands

install

If scheduledMaintenance is false, tell the user to enable it (/config or the plugin config) first — or confirm they want to install anyway. Then run, from the maintenance dir:

BRAIN_MAINT_CRON="<maintenanceSchedule>" \
BRAIN_MAINT_MODEL="<maintenanceModel>" \
BRAIN_MAINT_SINCE_DAYS="<scanSinceDays>" \
BRAIN_MAINT_TIMEOUT_MS="$(( <maintenanceTimeoutMin> * 60000 ))" \
bash install.sh

Then show crontab -l | grep matts-second-brain-maintenance to confirm. Remind the user the cron runs as them, so the brain backend must be reachable from their account.

status

Show the installed entry (crontab -l | grep matts-second-brain-maintenance) and the last few structured runs (tail -n 5 maintenance/logs/maintenance.log). Each line is JSON with a status (ok / locked / unhealthy / failed / timeout) and, for completed runs, an agent-summary of what was persisted/updated/skipped.

run

node maintenance/run.mjs — a real run. Warn first: it scans recent transcripts and may persist a few notes. Afterwards show the new agent-summary log line.

dry-run

node maintenance/run.mjs --dry-run — health-check + print the prompt, no agent invoked. Safe.

uninstall

bash maintenance/install.sh --uninstall — removes the cron entry. Leaves logs in place.

test

cd maintenance && node --test — runs the harness test suite (lock/health/timeout/exit semantics). Dependency-free; no build step.