Usage¶

sinteractive [LAUNCH OPTIONS] [SBATCH ARGS...]     # launch a session
sinteractive <COMMAND> [ARGS...]

A bare sinteractive launches; everything else is a subcommand (sinteractive --help, sinteractive <command> --help, man sinteractive).

Commands¶

Sessions¶

Watching¶

Driving a session from outside¶

A person at a prompt attaches; a script or an agent reaches in with these.

Claude Code¶

claude install writes the last four into your Claude Code settings; you never type them yourself.

Generated output¶

status, list, cancel, queue, monitor, quota, doctor and launch --detach take --json. TARGET is a JOBID or a session NAME; inside a session it defaults to the current one.

The pre-grouping spellings — ensure, peek, send, events, refresh, snapshot, agent-context, hook, statusline, mcp, install-claude, completions, man and schema as top-level commands — still work but are hidden from --help.

Launch options¶

All other arguments are passed directly to sbatch, in any order, so you can use any sbatch option (--gres=gpu:1, --qos=long, --account=...). session ensure takes the same options after the name.

Examples¶

# Default: 1-day session, 2 CPUs, 8G memory
sinteractive

# Named, 8 hours, 4 CPUs, 16G
sinteractive -n rna-seq -t 8h -j 4 -m 16G

# Run on a specific node
sinteractive --node compute01

# GPU session; unknown flags go to sbatch
sinteractive --partition=gpu --gres=gpu:1 --mem=16G

# Longer session on the normal partition (Bodhi: up to 3 days)
sinteractive --time=1-12:00:00 --partition=normal

# Launch without attaching, then come back to it
sinteractive --detach -n build
sinteractive attach build

# What is running, and how long is left?
sinteractive list
sinteractive status build

# The queue, refreshed every 5 s (q to exit)
sinteractive queue --watch

# The build session's node, nvitop-style, from the login node
sinteractive monitor build

# Read the last 40 lines of a session's screen
sinteractive session peek build -n 40

Inside a session¶

Ctrl+b is the only chord (tmux muscle memory); press it, then one key. Everything else zellij binds by default is cleared, so shell and editor keys pass through untouched. Ctrl+b h shows the same legend in the status bar.

The status bar¶

● sint 31761255 · rusttest · c3cpu-a2-u3-4 · 22m left · 3R launched · ^b h help

• The dot spins while the session is starting, and turns yellow, then red, as the walltime runs down (SINTERACTIVE_WARN_YELLOW / _RED: an hour and ten minutes by default). SINTERACTIVE_GRACE seconds before the limit the session ends itself, so teardown runs cleanly instead of under Slurm's SIGKILL.
• 3R launched counts the running (R) and pending (PD) jobs started from this session — not every job you own, which the session cannot do anything about. Slurm records the node a job was submitted from and the session id of the process that submitted it, so a job is a candidate when it was submitted on this node and that process is either gone or names this session. A running job then has the last word itself: Slurm gives a job the environment of the shell that submitted it, and a session exports SINTERACTIVE_JOB_ID into every shell it runs, so the job's own processes say which session started it, however long ago that shell exited. A job that a previous session on this node started is dropped on its first sample; a pending job has no processes yet and stays on Slurm's word. Absent when nothing has been launched, which is most of the time.
• ▣ N jobs monitorable ^b m appears when there is a host the monitor panel could show and the panel is closed. The panel shows the same jobs the count counts — the ones launched from this session.
• ⚠ N notices ^b n appears when the session has something to say — a quota overage (red, shimmering), a walltime trimmed before a maintenance window, a hint to run claude install while Claude Code is running without the hooks. Absent when there is nothing to say; sinteractive status prints the same text from the login node.

Segments drop from the right as the terminal narrows; the job id is the last to go.

The monitor panel¶

Ctrl+b m opens a six-row panel between the shell and the bar: a strip of every running job launched from this session (this one first), then bars for CPU and memory against the selected job's cgroup limits, and a row per GPU. A job with no GPU spends the spare row on where its load has been, a sparkline the width of the bars above it. Jobs on other nodes are sampled over ssh every ten seconds; one on this node is sampled here.

In a pane 122 columns or wider the resources go two to a row — cpu beside mem, then the GPUs in pairs — so a four-GPU node fits with a row to spare. Narrower than that the rows stack as below, and a third and fourth GPU fall off the bottom.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
147845 mywork · 246422 sint-mods   1/2      ←→ job · t top · esc shell · x close
cpu ███████░░░░░░░░░░░░░  34% of 8 · load 3.2
mem ███████░░░░░░░░░░░░░  37% 12.0G / 32G
gpu0 █████████████████░░░  87% 31/40G 61°C 240W ████░░░░

Ctrl+b m focuses the panel rather than toggling it, so the panel is a pane you step into and out of:

Ctrl+b chords keep working while the panel holds the focus — zellij reads the prefix before the focused pane sees a key — so Ctrl+b m steps back out to the shell, Ctrl+b d detaches, and so on. From the shell, Ctrl+b , and Ctrl+b . step through jobs without taking the focus at all.

Mouse, copy and paste¶

Mouse mode is on by default: scroll with the wheel, click to focus a pane, drag borders to resize, and select text to copy it — the selection lands in your local system clipboard over SSH via OSC 52. Hold Shift to select with the terminal instead. --no-mouse or SINTERACTIVE_MOUSE=off turns mouse mode off for a session.

For keyboard copying, Ctrl+b [ enters scroll mode; e opens the whole scrollback in $EDITOR, which is the easiest way to search or copy a long stretch of output.

Reconnecting after a disconnect¶

If your SSH connection drops or you detach (Ctrl+b d), the session keeps running on the compute node and your work is safe. From the login node:

# List your running sessions
sinteractive list
#   JOBID       NAME                  NODE            PARTITION     ELAPSED     TIMELIMIT   CWD
#   12345       rna-seq               compute01       cpu           01:23:45    1-00:00:00  ~/projects/rna-seq

# Reattach
sinteractive attach 12345
sinteractive attach rna-seq

If you have only one session running, a bare sinteractive attach goes straight to it — no need to look up the job id first. With several running, it lists them with ready-to-run commands to pick from.

attach reconnects through Slurm (srun --overlap), which needs no SSH access to the node. attach --ssh uses ssh -X instead, which is the way to get X11 forwarding.

Cancelling the job¶

Exiting the last shell (type exit or Ctrl+d in every pane) ends the Slurm job. You can also cancel it from the login node, by name or job id:

sinteractive cancel myproj
sinteractive cancel 12345
scancel 12345              # equivalent, job id only

Pressing Ctrl+c while a launch is still waiting in the queue cancels the pending job too.

Waiting for a job to start¶

When the cluster is busy your job may sit in the queue. While it does, sinteractive shows why it is waiting (Slurm's pend reason — free resources, higher-priority jobs ahead of you) and, when Slurm can estimate one, the expected start time:

 ⠹ waiting for free resources — est. start 14:32 (2m elapsed)

sinteractive queue shows the same for every job you have, plus the last day's history with a memory right-sizing hint; --all adds everyone's jobs in the partitions you can see.

What a session is for¶

A session is a durable place to work from — editing, git, scheduler queries, and keeping long-lived state across SSH drops. It is not a compute target: the default interactive partition is the smallest on the cluster, and anything heavy you run in the session competes with the shell you are typing in.

Run work in an allocation sized for it instead:

# One-off job
srun -p rna -c 8 --mem 32G -t 1:00:00 -J make-test --comment=make-test -- make test

# Sustained work: hold one allocation and reuse it
salloc --no-shell -p rna -c 32 --mem 96G -t 4:00:00 -J cargo-ci --comment=cargo-ci
srun --overlap --jobid=ID -- cargo build --release
scancel ID

Name every job in both fields — -J NAME and --comment=NAME, the same short descriptive value — so the queue says what is running and why:

squeue --me -o "%.10i %.20j %.20P %.10M %k"   # %j is the name, %k the comment

SLURM_* is stripped from a session, so srun and salloc run from inside one create their own allocations rather than steps of the session's job.

Environment variables¶

Set personal defaults in your ~/.bashrc; explicit flags always win.

# Example: a bigger default allocation, cache on a filesystem with room
export SINTERACTIVE_MEM=16G
export SINTERACTIVE_CPUS=4
export SINTERACTIVE_CACHE=/projects/$USER/.cache/sinteractive

Configuring for Alpine (CU Boulder)¶

The defaults match Bodhi, but everything scheduler-specific is overridable. On Alpine three things differ:

• CPU partition + QOS — the general-purpose CPU queue is acpu and an explicit --qos is mandatory. acpu/cpu-normal are the names that took effect with Alpine's 2026-08-05 rename of amilan/normal; both name sets are accepted.
• /home is 2 GB — put the cache on /projects.
• name clash on PATH — Alpine already provides an older, screen-based sinteractive in /usr/local/bin, which is ahead of ~/.local/bin on PATH. An alias forces your copy to win.

Add this to your ~/.bashrc:

# Use the ~/.local/bin copy instead of Alpine's older /usr/local/bin one
alias sinteractive="$HOME/.local/bin/sinteractive"

export SINTERACTIVE_PARTITION=acpu         # CPU queue (was 'amilan' pre-2026-08-05)
export SINTERACTIVE_QOS=cpu-normal         # 1-day max walltime; QOS is required on Alpine
export SINTERACTIVE_CACHE=/projects/$USER/.cache/sinteractive

Then sinteractive launches a 1-day CPU session. For a longer run (up to 7 days), override the QOS: sinteractive --time=2d --qos=cpu-long. The default account (amc-general for most users) is applied automatically; pass --account=<name> if you need a different allocation. Alpine has no quota daemons, so sinteractive quota reports "unavailable" there — curc-quota is the tool.

Configuring for Bodhi¶

Nothing to set: the built-in defaults are Bodhi's (interactive partition, no QOS, the quota daemons and /cluster/scripts/quota_current.txt). Longer sessions go to the normal partition (sinteractive -t 1-12:00:00 -p normal), GPU work to gpu (-p gpu --gres=gpu:1).

Over-quota sessions carry a red QUOTA over by … notice, checked every ten minutes against the cluster's quota daemons. The check is cached per user, not per session, so six open sessions do not mean six times the polling. After deleting something, don't wait out the interval:

sinteractive quota --check      # re-checks now, updates every open session
# OVER QUOTA: 30.2T of 30T used (100.7%), over by 204.8G
# Quota OK: 24.1T of 30T used (80.3%)

Tab completion¶

make install installs completions for bash, zsh and fish (generated by the binary itself: sinteractive gen completions bash|zsh|fish). Start a new shell after installing to pick them up. zsh needs ~/.local/share/zsh/site-functions on $fpath.