RonanCodes/snapshot-sessions

/ro:snapshot-sessions

A pre-restart / pre-shutdown safety sweep across every live Claude Code session on the machine. It is a fire-alarm, not a dashboard and not a resume system: its one irreducible value is the cross-session view that no single session can produce (each session only sees its own cwd, and /ro:close-session only sweeps the repo it runs in).

Why this exists

You have N Claude sessions open across repos and worktrees in iTerm tabs. You need to restart the machine. Three things are true:

1. Transcripts and working trees survive a restart. Conversations are journalled to ~/.claude/projects/<slug>/<session-id>.jsonl; claude --resume <id> replays one verbatim. Uncommitted files are on disk and survive reboot. So you lose almost nothing on a clean restart.
2. What you DO lose is the cross-session map (which tab was in which folder/worktree, doing what) and any background work (workflows, ScheduleWakeup chains, dev servers).
3. Tab names lie. A tab labelled "NUTMEG" might be a shell, and a week-idle tab is invisible to any timestamp-based scan.

This skill solves #2 by keying off live processes, so it catches every open session regardless of how long it has been idle.

What it does

Run the bundled script:

python3 "$(dirname "$0")/scripts/snapshot.py"   # or: python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/Dev/ronan-skills}/skills/snapshot-sessions/scripts/snapshot.py"

It:

1. pgrep -fl 'claude --dangerously-skip-permissions' for live session pids.
2. Per pid, lsof -p <pid> -a -d cwd -Fn for the working directory (Claude does NOT hold the transcript open continuously, so key off cwd, then map to the newest transcript in ~/.claude/projects/<slug>/; slug = every non-alphanumeric char replaced with -).
3. Per session, gather: branch, meaningful uncommitted count (ignores agent/tooling state: .claude/RESUME.md, .swarm/, .nightshift/, .ralph/, .claude/worktrees/, the snapshot file itself), unpushed commits (@{u}..HEAD), open PR for the branch (gh pr list --head), and a one-line "what" (prefers a .claude/RESUME.md / .swarm/run-state.md summary line, falls back to the last real user instruction).
4. List orphaned worktrees: worktrees of any live repo with no session attached. Clean + merged ones are reap candidates.
5. Write ~/.claude/session-snapshot.md danger-first: rows with uncommitted/unpushed/open-PR work at the top, clean rows next, orphaned worktrees, then a paste-ready resume block.

After writing, open it so the user sees it:

open ~/.claude/session-snapshot.md

snapshot-show

"snapshot-show" / "show my saved sessions" is the read-back half. After a restart the live processes are gone, but the file persists, so just:

cat ~/.claude/session-snapshot.md   # or: open ~/.claude/session-snapshot.md

No re-scan needed (and impossible anyway, the sessions are closed). Capture needs live processes; show just reads the saved file.

The restart flow

1. Run /ro:snapshot-sessions while the sessions are still alive (it needs the processes).
2. Deal with the ⚠️ danger rows first: anything uncommitted/unpushed/open-PR is work you are about to walk away from. For a session with real WIP, run /ro:close-session in that tab rather than just relying on --resume.
3. Restart.
4. Read ~/.claude/session-snapshot.md, recreate tabs, run the resume command per row.

Scope discipline (do not grow this)

• It is a fire-alarm, not a daily dashboard. Its value is inversely proportional to your session discipline: if you close sessions promptly (/ro:close-session), you rarely need it. It earns its keep digging out of session sprawl and as a crash safety net.
• Resume is the safety net, not the goal. The disciplined path is /ro:close-session per tab (durable capture: issues, ADRs, wiki, memory), then start fresh and light (Pocock's <100k-token sessions). This skill does not replace that.
• Do not auto-resume or auto-recreate tabs. It hands the user the commands; the human runs them.

Related

• /ro:close-session — the per-session durability sweep; the disciplined front door this skill is the fallback for.
• /ro:handoff — compacts one session into a rich handoff doc; its summary line is what this skill's "what" column reads when present.
• /ro:worktree — reap mode audits the orphaned worktrees this skill surfaces before removing them.

Provenance

Created 2026-06-14 after a multi-repo night-shift run left 10 sessions open across repos and worktrees with a forced restart pending. Built from a proven script written live that session. Named "snapshot" deliberately: the file is read AFTER the sessions close (so "open-sessions" would be wrong), and "checkpoint" collides with /ro:way-of-working's commit-cadence vocabulary. See ronan-skills#77.