Day-to-day

Nothing to run by hand

Both housekeeping scripts run automatically via hooks, so you never update anything manually:

When	Hook	Does
session start	sync-session.sh	git pull --ff-only both repos (data and app, so hook/script fixes propagate too), then claude-sync.sh — config edited on another machine arrives and applies itself.
session end	log-session.sh	append the log line, refresh chats.index.json, then git add -A + commit + push the whole data repo — memory/, templates/, host config and all. Nothing needs a manual sync.

git add -A is safe because the data repo's .gitignore blocks secrets and transcripts (*.credentials*, *.jsonl, .claude.json), so those can never be staged.

Symlinked scopes (CLAUDE.md, agents/, hooks/, and the per-file links in commands/) go live on the pull alone — claude-sync.sh only has real work when settings.json changed or a new command/agent file appeared (it relinks commands/ from the app + data sources). You can still run either by hand (both idempotent); the hooks just mean you don't have to:

bin/claude-sync.sh         # re-apply config (auto-runs at SessionStart)
bin/claude-index-chats.sh  # refresh this host's chats.index.json (auto-runs at SessionEnd)

Escape hatches (env, in ~/.config/scrubjay/config or inline): SCRUBJAY_NOSYNC=1 (skip the start-of-session pull+sync), SCRUBJAY_SYNC_NOPULL=1 (sync without pulling). The full toggle list is in the Reference cheatsheet.

Find a past chat

Every session is logged by the SessionEnd hook to scrubjay-data/logs/<host>.log (one line: time | host | cwd | "topic" | session=id | harness=… | model=… | turns=… | size=…) and pushed, so all machines' histories are searchable from any clone of the data repo. The topic is the first real prompt, or — if you publish with /sjlog — a one-sentence essence the model writes; the trailing key=value fields are additive, so older lines simply have fewer:

git -C ~/.scrubjay/scrubjay-data pull
grep -i refactor ~/.scrubjay/scrubjay-data/logs/*.log

The full transcript (the .jsonl) lives in scrubjay-chats / on the NAS under <host>/<slug>/<session>.jsonl.

For recall by topic (rather than an exact word you remember typing) from inside a live session, use the sjmcp MCP server.

Troubleshooting

Hooks only activate on the next session. claude-sync.sh (and onboard.sh, which runs it) install the hooks — symlinking ~/.claude/hooks and registering them in ~/.claude/settings.json. But Claude reads its hooks at session start, so a claude instance that was already running when you onboarded won't see them: its SessionStart already fired, with no hooks to load. Quit and reopen claude after onboarding — only sessions started afterwards run SessionEnd (the transcript/subagents/plans relay + log/index push). This is the usual reason a freshly-onboarded machine "ships nothing".

• SessionEnd triggers on both /exit and /clear (the latter then starts a fresh session). Either way, a session that began before the hooks were active won't ship.
• "SessionEnd hook … Hook cancelled": Claude cancels hooks that haven't returned by the time the session process exits. log-session.sh detaches its network work (git push + relay) so shutdown can't interrupt it — just ensure the machine has pulled the app repo (git -C <app-clone> pull; SessionStart does this automatically next time).
• Expecting memory/ on the NAS? Memory isn't relayed — it rides the scrubjay-data git sync, not the session relay (see Transcripts: relay + NAS).