srunx CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Development Commands

Package Management

• uv sync - Install dependencies
• uv add <package> - Add new dependency
• uv run <command> - Run commands in virtual environment

CLI Usage

Job Management (SLURM-aligned commands)

Command names mirror SLURM's CLI where it makes sense (sbatch / squeue / scancel / sinfo) so a SLURM user can map their muscle memory directly. srunx-specific commands that don't map to a SLURM binary:

• history — srunx's own submission history (SQLite-backed). Named history rather than sacct because it shares no backend with real sacct; only jobs submitted via srunx are listed.

• gpus — GPU aggregate summary across partitions.

• tail / watch — log tailing and cluster watching.

• uv run srunx sbatch <script> - Submit a sbatch script (positional, like real sbatch)

• uv run srunx sbatch --wrap "cmd ..." - Wrap a command into a SLURM job (mutually exclusive with the positional script)

• uv run srunx sbatch --profile <name> ... - Submit over SSH to a configured profile

• uv run srunx squeue - List active jobs (all users by default — matches native squeue). Default columns: Job ID, User, Name, Status, CPUs, GPUs, Elapsed, NodeList

• uv run srunx squeue -j <job_id> - Filter to a specific job (replaces the old srunx status for active jobs)

• uv run srunx squeue -u <user> - Filter to a single user

• uv run srunx squeue --show-partition --show-limit --show-nodes - Opt-in columns (each flag adds one)

• uv run srunx squeue -a - Shortcut for all opt-in columns at once

• uv run srunx squeue -i <seconds> - Live refresh: re-query and redraw in place every N seconds (like native squeue -i). Ctrl+C exits

• uv run srunx squeue --format json - JSON always includes every field regardless of show flags

• uv run srunx scancel <job_id> - Cancel a job

• uv run srunx sinfo - Partition / state / nodelist listing (same columns as native SLURM sinfo)

• uv run srunx sinfo --profile <name> - Query a remote cluster via SSH adapter (#139)

• uv run srunx sinfo --partition gpu --format json - Partition rows as JSON

• uv run srunx gpus - GPU aggregate summary (what the old srunx sinfo showed)

• uv run srunx gpus --profile <name> --partition gpu - GPU summary scoped to one partition on a remote cluster

• uv run srunx history - DB-backed submission history (srunx's own SQLite). Only jobs submitted via srunx are listed

• uv run srunx history -j <job_id> - Filter history to a specific job (replaces srunx status for finished jobs)

• uv run srunx history --profile <name> - Scope history to a single cluster's jobs (scheduler_key filter)

• uv run srunx sacct - Real SLURM sacct wrapper (cluster accounting DB — includes manual sbatch jobs, needs slurmdbd). Default columns: Job ID, User, Name, Partition, State, ExitCode, Elapsed

• uv run srunx sacct -a -S now-1day - All users over the last day (like native sacct -a -S now-1day)

• uv run srunx sacct -j <job_id> / -u <user> / -s FAILED,TIMEOUT / -p <partition> - Standard sacct filters

• uv run srunx sacct --show-steps - Include .batch / .extern sub-step rows

• uv run srunx tail <job_id> - Show the last 10 lines of the job log (matches native tail; use -n N for a different cap, --all to dump the whole file)

• uv run srunx tail <job_id> --follow - Stream job logs. Works over SSH via periodic tail_log_incremental polls; tune with --interval <seconds> (default 2s). Ctrl+C exits