SLURM resurrection (slurm-resurrect)

An optional extra (plugin slurm-resurrect, in extras/slurm-resurrect/) for one setting only: development on a compute node of a computing cluster that uses the SLURM scheduler, with Claude Code or Codex running in tmux inside a batch job on that node. It is of no use on a laptop, a workstation, or a cluster login node.

A batch job ends at its time limit, and every agent session running in it ends with it. This plugin rebuilds your tmux session in a new job and resumes every Claude Code session that was running in it, with --resume, and every Codex session it can identify, with codex resume. Windows, panes, layout and working directories are restored. Nothing in the other open-science plugins depends on it. How to start such a job and connect to it is in Working in tmux.

claude plugin install slurm-resurrect@open-science

For Codex: codex plugin add slurm-resurrect@open-science (see Codex).

The full reference is extras/slurm-resurrect/README.md; how it works is in extras/slurm-resurrect/reference/mechanism.md.

Requirements

• tmux. Claude Code or Codex must run inside tmux, and the tmux server must run inside the SLURM batch job.
• A SLURM batch job. Account, partition and time limit are read from the running job each time you register.
• bash, jq, flock, setsid, sbatch, squeue, scancel.
• A state directory on a filesystem that the compute nodes share. Default: ${XDG_STATE_HOME:-$HOME/.local/state}/slurm-resurrect; set RR_STATE_DIR to use another one. A tmpfs or a path under /tmp is not shared.

Turn it on

Only you can register a session. From any Claude Code pane in the tmux session, type:

/slurm-resurrect:resurrect register

The plugin's UserPromptSubmit hook runs the command before the model sees the prompt. The first register only shows a warning about the settings below and registers nothing; run it again to register. The successor job is queued as soon as a session is registered. Codex has no such command: run bash <plugin dir>/scripts/rr_registry.sh register in a plain terminal pane of the session.

Agents cannot register: the skill cannot be invoked by the model, and the script refuses register, reset, set and set-notify when it runs under Claude Code (other than through the prompt hook) or Codex, and logs the attempt. This stops an agent from registering by accident; it is not a security boundary.

Options of register:

option	default	what
--permission-mode MODE	none: each pane's own mode	Claude Code only: the widest permission mode a resumed pane may get (any mode claude --permission-mode accepts, for example acceptEdits), and the mode for a pane whose own mode cannot be read
--remote-control on\|off	on	Claude Code only: whether the resumed sessions can be read and driven from any device logged in to your Claude account
session ...	the current session	register named tmux sessions instead

A resumed session runs with no one watching. Each Claude pane resumes in the permission mode its process was started with (--permission-mode, or --dangerously-skip-permissions, which counts as bypassPermissions), or in MODE if that is narrower; a pane started without a mode flag resumes without one, so Claude Code's settings decide as before. A pane that ran in bypassPermissions therefore resumes in it unless you register with a narrower --permission-mode; in that mode it runs every command, including edits and deletions, without asking. A session whose registration record is missing resumes in at most manual, with Remote Control off. Both settings are recorded per session and applied at every hop.

Folder trust. A resumed Claude session can open in a folder Claude Code has not trusted yet (the pane's directory or the config directory changed), and then asks whether to trust it. The plugin answers "Yes, I trust this folder" only with set auto_trust true, which setup asks about. Accepting lets that folder's .claude/settings.json hooks and MCP servers run without your review, so a folder with code you did not write (a cloned repository, for example) could run commands as you. With auto_trust false, the default, the plugin presses nothing and notifies you each time. A Codex folder-trust question is always left to you.

Other safeguards. tmux servers are started or contacted only on a socket whose directory is yours alone (owned by you, no group or other permissions, not a symlink); otherwise the session is rebuilt on a new private socket and the notice gives the attach command. Values typed into a pane's shell are quoted. Notes and resume prompts typed into a pane lose their control characters and cannot start with /, ! or # (except the context plugin's own /open-science-context:continue-context prompt), and a note is always delivered behind a fixed [slurm-resurrect] prefix. If no Claude or Codex pane is alive after a hop, the lineage ends and you are notified.

Commands

Type them as /slurm-resurrect:resurrect <command>, or run bash <plugin dir>/scripts/rr_registry.sh <command> in a plain terminal pane of the session.

command	what
register [options] [session ...]	opt the tmux session in (see above)
status	print the lineage settings (including the queued successor job) and the tmux sessions registered in this job
timeleft	the time left in the current job
remove [session]	take one tmux session out
stop	end the lineage and cancel the queued successor job
reset [N]	a new hop budget (default cap 10 hops)
set KEY VALUE	change a setting; set alone lists the keys
set-notify CMD	run CMD with the message in $RR_MSG at each notice
note "<text>"	(agents may run this) save a message delivered to this pane after the hop

Settings for set: queue_mode, early_lead_seconds, handoff_timeout_seconds, handoff_grace_seconds, snapshot_interval_seconds, pause_threshold_seconds, winddown_threshold_seconds, launch_cmd, sbatch_extra, account, partition, default_time_limit, nodes, ntasks, cpus_per_task, context_window_suffix, core_scripts_dir, auto_trust (true or false; see "Folder trust" above).

To resume with a command other than claude (for example a wrapper that sets the config directory), run set launch_cmd <command>; for Codex, set codex_launch_cmd <command>.

To send resurrection notices through opsci notify (useful with the Slack back end), type /slurm-resurrect:resurrect set-notify <path to opsci> notify "$RR_MSG", with the full path from command -v opsci, since the batch job's PATH may differ.

Queueing

queue_mode	how the successor waits
afterany (default)	--dependency=afterany:<job>: it starts after the current job ends. On some clusters a job waiting on a dependency gains no age priority, so the gap can be long
early	--begin=<end of current job - early_lead_seconds> (default 4 hours), no dependency. If it starts before the current job ends, it asks the current job for a final snapshot, cancels it, then rebuilds

Before the limit

At winddown_threshold_seconds before the limit (default 30 minutes) each active Claude Code or Codex pane gets one message: the job is near its limit and the session will be resumed; if the work in hand will not finish, stop cleanly and leave a note for after the hop. At pause_threshold_seconds (default 120 seconds) the final snapshot is taken.

With context management

If open-science-context is installed, the two plugins coordinate. From the wind-down message until the hop, session jumps are inhibited in the registered panes. Where Claude Code runs the context plugin's mod (the default on a recent Claude Code), the mod does the rest: a resumed session keeps its session id, and with it its registered task and its queued SLURM wakers, and a session that was waiting on background tasks, which ended with the old job, is woken with /open-science-context:continue-context. Without the mod, this plugin does it by typing into the panes: after the hop, a jump that the time limit interrupted is finished, and a pane that was waiting after a wait jump is woken with /open-science-context:continue-context. Without open-science-context, the plugin works the same way and skips these steps. Details: extras/slurm-resurrect/reference/jump-hook.md.

Turn it off

• remove [session] takes one tmux session out.
• stop ends the lineage and cancels the queued successor.
• Disabling or uninstalling the plugin does not cancel a successor that is already queued: queued jobs run a copy of the scripts kept in the state directory. Run stop first.

Known limits

• A session started with --plugin-dir loses that flag on resume unless launch_cmd includes it.
• If a resumed session shows the workspace trust dialog and auto_trust is true, the plugin selects "Yes, I trust this folder"; if it cannot select it, it presses nothing and notifies you. With auto_trust false it presses nothing and notifies you.
• Registration is checked by process ancestry, not enforced by the operating system.
• Further limits of the tests are listed in the plugin's README.md.

Codex

After adding the marketplace, install with codex plugin add slurm-resurrect@open-science. Mixed Claude Code and Codex panes are supported. Codex panes resume with codex resume, preserving their original launch options and model; Claude permission settings are not applied to them. Unknown launch options or a missing thread identity leave a plain shell and a notice.

Register from your own plain terminal pane using the rr_registry.sh commands in the extra's README. Codex agents cannot register a session. The core context plugin's trusted hooks provide thread identity; wind-down notices and wake-ups use codex queue. Codex folder-trust prompts always need your response, whatever auto_trust says. An interrupted Codex jump is not re-driven after a hop: the thread is resumed as it was. Fake-scheduler tests and live Codex resume checks cover these adapters; a real SLURM hop with Codex has not been tested.