phoenix

██████╗ ██╗  ██╗ ██████╗ ███████╗███╗   ██╗██╗██╗  ██╗
██╔══██╗██║  ██║██╔═══██╗██╔════╝████╗  ██║██║╚██╗██╔╝
██████╔╝███████║██║   ██║█████╗  ██╔██╗ ██║██║ ╚███╔╝
██╔═══╝ ██╔══██║██║   ██║██╔══╝  ██║╚██╗██║██║ ██╔██╗
██║     ██║  ██║╚██████╔╝███████╗██║ ╚████║██║██╔╝ ██╗
╚═╝     ╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝  ╚═══╝╚═╝╚═╝  ╚═╝

        ◎ Reboot without losing a single session ◎
                   CLAUDE AGENT SYSTEM
              ⚠ BETA — macOS + Ghostty only ⚠

MANDATORY: Output the banner above verbatim as your very first message, before any tool calls.

⚠ BETA — macOS + Ghostty only

Newly shipped and validated against a real 15-session setup, but not yet battle-tested across machines. It reopens sessions in Ghostty only (the reopen step uses ghostty -e), and the auto-reboot needs macOS. If anything misbehaves: your sessions are armed independently of the reboot, so the safe path is /phoenix snapshot → /phoenix arm → reboot manually. Cancel a pending restore anytime with /phoenix disarm. Feedback: https://github.com/Kasempiternal/Claude-Agent-System/issues

Phoenix makes rebooting cost-free when you have many claude sessions open. It captures the live session set, arms a one-shot login agent, and (optionally) reboots. On the next login, every session reopens as a tab in one Ghostty window, cd'd to its original directory and resumed with claude --resume <id>.

How it works (one moving part: a deterministic CLI)

All behavior lives in a single dependency-free Node script. The skill's only job is to run the right subcommand and print its stdout to the user verbatim.

• Detection is from the OS, not a registry: ps finds the claude CLI process trees and lsof recovers each one's working directory plus the conversation transcript (~/.claude/projects/<enc>/<uuid>.jsonl) it holds open — the transcript basename is the session id. This correctly separates multiple sessions sharing one directory (each tree holds its own transcript open). The Electron desktop app and daemon "spare" processes are filtered out.
• Restore reopens every session as a tab in ONE Ghostty window. The first session opens the window (open -na Ghostty --args -e zsh -lc "…cd <dir> && exec <abs claude> --resume <id>"); each remaining session is added as a tab via keystroke automation (osascript → activate Ghostty → Cmd+T → paste the command → Return), because Ghostty has no CLI to open a tab on macOS (+new-window is unsupported there; new_tab exists only as a keybind). The claude path is absolute (resolved + stored at snapshot time) because Ghostty's zsh -lc is a non-interactive login shell that does not source ~/.zshrc, so claude would not be on PATH otherwise.
• State lives at a stable path ~/.cas/phoenix/ (snapshot.json, ARMED flag, a copy of the script, restore.log). The login agent is ~/Library/LaunchAgents/com.cas.phoenix.restore.plist.

Subcommand routing

Treat the first word of $ARGUMENTS (default to status if empty, and confirm intent before any destructive action):

User typed	Run
/phoenix restart	node "${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js" restart
/phoenix snapshot	node "${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js" snapshot
/phoenix restore	node "${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js" restore
/phoenix arm	node "${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js" arm
/phoenix disarm	node "${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js" disarm
/phoenix status	node "${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js" status

Useful flags: restart --dry-run (snapshot + arm, print the reboot command instead of rebooting); restore --dry-run (print the exact launch commands without opening windows).

What each does

• restart — the full auto path: snapshot → arm → 5-second abort countdown → reboot via osascript … System Events restart. After login the one-shot agent reopens everything, then removes itself. This reboots the machine and ends the current session — confirm with the user first, and tell them their sessions (including this one) will reappear after they log back in.
• snapshot — capture the current sessions to ~/.cas/phoenix/snapshot.json. Non-destructive.
• arm / disarm — install / remove the one-shot login restore agent. arm does not reboot.
• restore — reopen everything from the snapshot now (also what the login agent runs). Avoid running this while the captured sessions are still live — it would open second windows on them.
• status — armed? snapshot age + captured sessions, agent installed?, recent restore log.

Recommended manual flow (if the user prefers to control the reboot): /phoenix snapshot → /phoenix arm → reboot manually whenever ready → sessions auto-reopen on next login.

Honest limitations (tell the user when relevant)

• Ghostty only. Restore opens one Ghostty window with one tab per session. If a session was running in a different terminal, it still reopens — just in Ghostty. Requires Ghostty.app.
• Tabs need Accessibility permission. Adding tabs uses keystroke automation (Cmd+T + paste), which requires the controlling process to be allowed under System Settings → Privacy & Security → Accessibility. Run manually (/phoenix restore) from a session that already has it and tabs build reliably (proven). At login (the auto-restore agent), keystroke permission is less certain — if tabs don't appear after a reboot, the first session's window still opens; re-run /phoenix restore from it, or grant Accessibility to the relevant process. Focus briefly moves to Ghostty while tabs are created, so don't type elsewhere during restore.
• The reboot may need permission. The first System Events restart can prompt for Automation/ Accessibility consent, and macOS may pause if an app has unsaved work. If the auto-reboot doesn't fire, sessions are already armed — just reboot manually and they reopen on next login.
• Multiple sessions in one directory are separated by which transcript each process holds open. A session that is completely idle with no held transcript falls back to "resume most recent in that directory" (claude -c), which can be ambiguous if several share the dir.
• One-shot by design. The agent fires once on the next login and deletes itself. Re-arm (or run /phoenix restart) for the next reboot. Cancel a pending restore anytime with /phoenix disarm.
• Snapshots are point-in-time. Sessions started after a snapshot won't be restored — re-run snapshot (or use restart, which snapshots automatically) right before rebooting.

Reference

• Engine: ${CLAUDE_PLUGIN_ROOT}/skills/phoenix/scripts/phoenix.js
• State: ~/.cas/phoenix/ (snapshot.json, snapshot.md, ARMED, phoenix.js copy, restore.log)
• Login agent: ~/Library/LaunchAgents/com.cas.phoenix.restore.plist
• Feedback: https://github.com/Kasempiternal/Claude-Agent-System/issues