remote-compute-ssh

What it does

Submits and monitors jobs on the user’s SSH/SLURM hosts. It handles orchestration tasks (partitions, env activation, job scripts, file transfer, recovery) while the science runs under separate skills. Each submission prompts the user for approval before using allocation, and the skill records host/provider learnings for future sessions, aiming to land the first submission efficiently and document insights for subsequent runs.

How it works

All host.compute.* calls run through the repl tool (control-plane kernel). Job submission opens an approval modal and the SSH connection from the orchestrator's process. The workflow is:

• Prepare inputs in a python cell (write ./in.dat, pickle inputs).
• Run create → submit_job in a repl cell; it returns immediately while the daemon harvests results.
• Call wait_for_notification to wait for compute_done from the daemon poller.
• Read harvested files from hpc/<jobId>/ using python tool; publish features with save_artifacts(payload['featured_files']).
• If needed, re-enter a repl cell and use c.attach_job(job_id).result() to obtain the fuller result dict, then c.close().
• Use c.download() to fetch any additional files when necessary. End by returning the result verbatim.

A typical code example shows creating a host.compute, configuring inputs/outputs, submitting a job, and handling the results via the described tools and commands. The workflow emphasizes that inputs are staged as either src paths or artifacts, outputs are declared with glob patterns, and harvesting is capped by options like harvest and residency controls.

When to use it

Use when you need to orchestrate remote job submissions on SSH/SLURM hosts, including environment activation, scheduler directives, file staging, and result harvesting. Begin with compute_details({provider, mode:'read'}) and, once bound, use the suggested c = host.compute.create(provider) to discover host specifics and proceed.

What it can touch

The skill uses the repl tool for host.compute interactions and the python tool for in-workspace data handling. It performs file I/O in the workspace (writing in.dat, reading hpc/<jobId>/), and uses c.call_command(), c.submit_job(), c.attach_job(), c.close(), and c.download() to manage the lifecycle and data transfer. It describes using commands like source <path>/activate in the activation path when available, or discovering activation via c.call_command() results.

Caveats

The license declared is Apache-2.0. The skill depends on SSH access to the user's hosts and requires user approval for each submission (up to a maximum described as 10 approvals per batch). It records host/provider details across sessions via compute_details and uses a blocking wait for the compute_done notification, which may affect responsiveness if the daemon or network delays occur. It emphasizes not bypassing user consent and handling per-file confirmations for sensitive paths (dotfiles, etc.).