Snapshot

Snapshot Tmux Agents

Capture the current tmux session state to a snapshot file, so /tmux-reboot:restore can bring the agents back after tmux-resurrect restores the windows, panes, and working directories. The running claude and codex processes do not survive the reboot and must be resumed separately. Panes running a long-lived foreground command (dev servers and watchers like just dev, npm run dev, vite, cargo watch) or a read-only viewer (git log, git show, less FILE, man, htop, tig, lazygit) are also captured as command rows — these carry no session state, so restoring one just re-runs the command line (best-effort). Their restore_default field separates the two: false for dev servers, true for viewers.

Run:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.py --output .llm/resume-after-reboot-state.json

Stdout works too, for redirection: python3 ${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.py > FILE.

Then show the user the captured rows and confirm they look right before they reboot. The state file holds session ids personal to this machine, so keep it in the gitignored .llm/ directory.

The state file

JSON, schema resume-after-reboot/v1. One row per window:

{
	"schema": "resume-after-reboot/v1",
	"captured": "2026-07-29T20:22:47-04:00",
	"backend": "tmux",
	"session": "main",
	"rows": [
		{
			"slot": 1,
			"name": "webapp",
			"tool": "claude",
			"command": "claude --resume 3f2a1b0c-4d5e-4f60-8a71-b2c3d4e5f607",
			"cwd": "~/projects/webapp",
			"session_id": "3f2a1b0c-4d5e-4f60-8a71-b2c3d4e5f607",
			"note": "session mtime 2026-07-29 20:17"
		}
	]
}

• tool is claude, codex, command, or shell. shell rows are idle panes: command and session_id are null and nothing is restored into them.
• slot is a 1-based ordinal for ordering and display, not a tmux window index — restore pairs rows by cwd, because tmux-resurrect renumbers windows across a reboot.
• session_id is broken out of command so nothing has to re-parse the command string. It is null for command/shell rows and for the claude --continue / codex resume --last fallbacks.
• note records how the pairing was made, so a weak one is visible.

The format is tmux-shaped: a flat row list, since tmux-resurrect brings the windows and panes back on its own. The herdr-reboot plugin writes a nested resume-after-reboot/v2 document instead — workspaces, tabs, and splits included — so the two are not interchangeable.

Matching caveats

The script walks the process tree under each pane in the attached tmux session and matches the agent's working directory to recent transcripts on disk. Review these with the user when they affect the captured rows:

• Multiple agents in one working directory are assigned newest-transcript-first. The set of session ids is correct, but the window-to-session assignment is best-effort — that is what note records.
• An idle Codex session can have a stale transcript modification time. If the assignment looks wrong, note that the user may need the interactive codex resume picker after reboot.
• When no transcript matches, the snapshot records claude --continue or codex resume --last with a null session_id. Point out these fallback rows so the user knows to verify them after restoration.
• The script ignores mirrored tmux group sessions and snapshots only the attached session.
• command rows re-run their command line, not a saved session. Editors, REPLs, and shells with in-memory state (vim, psql, ssh) are intentionally not captured — only allowlisted long-running commands and read-only viewers are. Point these rows out as best-effort restores.
• git rows are judged by subcommand, not by program, so git push and git rebase are never captured. An alias is judged by what it expands to and restored as the alias you typed, so git la comes back as git la.
• A pager is captured only when it names a file. A bare less is draining a pipe whose writer dies with the reboot, and re-running it would hang the pane on stdin.
• tig and lazygit are captured as viewers, but both can stage, commit, and push from inside the TUI — a restored one is only as safe as what you then type at it.

Regenerate immediately before rebooting so the session ids are current. Keep refresh manual unless the user asks to automate it with cron or launchd.