Submit a Slurm job through CLIO Coder

CLIO Coder can submit and follow a Slurm job for you through the clio-kit MCP server. Each submission or cancellation is confirmed first: the safety judge reviews it against your own messages, and you decide when it has no verdict. Use this when a task needs real compute and you want the job's outcome reported back into the session.

Install the server

uv tool install clio-kit
clio-kit mcp-server slurm

The second command lets clio-kit prepare its private runtime, which can take longer than the 15 seconds CLIO Coder allows a server to start on a first launch. Stop it with Ctrl-C once it is running.

Declare it

Run clio-coder paths to find your config directory, then add this as mcp.yaml there:

version: 1
servers:
  - id: slurm
    command: clio-kit
    args: [mcp-server, slurm]
    timeoutMs: 120000

Keep the id slurm. The slurm-jobs skill you install below expects it, and it is what switches on the confirmation for allocations.

clio-coder doctor --verbose
clio-coder mcp list --refresh

Until the first refresh, plain mcp list shows the declaration with no recorded tool catalog. --refresh starts each trusted server, lists its tools live, and records the catalog.

Plain doctor folds the Slurm checks into one row while none of them needs attention:

OK   slurm                  ready: clio-kit, mcp server, scheduler
INFO scheduler commands     sbatch is on PATH: scheduler commands run outside the shell sandbox with approval

doctor --verbose prints the slurm clio-kit, slurm mcp server and slurm scheduler rows separately, and a row with a warning always keeps its own line.

The server row names the mcp.yaml that declares it and its trust. The scheduler row means only that sbatch and squeue answered --version; it does not contact the controller. None of these rows fails the check.

The host in the picture below has no Slurm client, so its scheduler row reads INFO slurm scheduler sbatch not on PATH; squeue not on PATH. The two rows above are what a host with the clients prints.

Install the skill

clio-coder library install skill:slurm-jobs --project --dry-run
clio-coder library install skill:slurm-jobs --project

Submit one job

The skill checks that the server is declared, looks at the partition, and submits once. The submission is confirmed before the job reaches the scheduler. When the safety judge has no verdict, an approval card appears, so read the script, cores, memory, and time limit before you press Enter.

The skill then polls with a growing wait until the job reaches a terminal state, and fetches bounded output.

To keep a job's output beside its record rather than in scratch space, ask the agent to put it into the project store, or run clio-coder store put results <stdout path> --note "Slurm job <id>" yourself. The store lives on the host where CLIO Coder runs, here the login node, and records where each file came from.

Without the server

The agent can also run sbatch, squeue, scancel and the other Slurm, PBS, LSF and Flux clients through bash. They run outside the shell sandbox, because they reach the scheduler through a host daemon the sandbox blocks, and they get the same review as ssh.

At default the safety judge decides first and asks you when it has no verdict. Contained yolo runs one only when your own message names it, and unconfined yolo runs it.

A script that calls sbatch itself stays inside the sandbox and fails there, and so does a workspace file named like a client, such as ./sbatch. Submit from the command line the agent runs, not from a wrapper. With safety.sandbox: required at unconfined yolo, only a one-line sbatch, squeue, scancel, sacct or sinfo with listed options leaves the sandbox; srun, salloc and flux stay inside.

Cancel a job

Ask the agent to cancel the job by id. Cancelling is confirmed the same way, and the server refuses unless the job id is repeated exactly.