| name | remote-scientific-compute |
| description | Run reproducible scientific workloads through an existing SSH host, Slurm/HPC cluster, cloud GPU provider, or remote object storage. Use for remote execution, job submission and monitoring, checkpointed long runs, GPU allocation, data staging, result retrieval, queue diagnosis, or resuming work that can outlive the current Codex task. |
Remote Scientific Compute
Move an approved scientific workload to existing remote infrastructure while
protecting credentials, login nodes, private data, budgets, and reproducibility.
Reference usage
Read the backend checklists for the selected SSH, Slurm/HPC, cloud GPU, or object-storage route before staging, submitting, allocating, or writing remotely. Do not infer scheduler flags, cost controls, storage semantics, or cleanup procedures from a backend name.
Read the durable job runtime contract before constructing, approving, submitting, polling, cancelling, or collecting a local or Slurm job. A process receipt records execution state; it is not scientific acceptance.
Record the reference version and SHA-256 in the run when it controls a submission or transfer. The backend reference governs execution safety; it is not evidence for the scientific result.
Inspect without changing state
- Use only an existing host alias, provider profile, project, and account. Never
enroll a service, create IAM policy, or modify SSH configuration implicitly.
If the host is not already known, stop and show its host fingerprint for the
user to verify out of band; never auto-accept first contact or a changed key.
- Confirm the target type and remote working directory. Probe with minimal
read-only commands such as scheduler availability, queue/partition metadata,
disk quota, and accelerator inventory. Never read private keys or print tokens.
- Inventory files to transfer, their hashes, sensitivity, size, destination, and
exclusions. Prefer code/config plus immutable data references over bulk copies.
- Read references/backends.md only for the selected
SSH, Slurm, cloud GPU, or object storage backend.
One explicit approval packet
Before the first write, transfer, submission, allocation, or paid action, request
explicit approval for one concrete packet containing:
- target host/provider/account and remote path;
- code revision, command, environment or digest-pinned image;
- files and data leaving the local machine, including sensitivity;
- CPU/GPU type and count, memory, wall time, storage, and concurrency;
- estimated cost with a hard time/cost cap for paid compute;
- output/log/checkpoint locations and retrieval plan;
- cancellation, cleanup, and instance-termination commands.
- monitoring cadence, maximum monitoring duration, and terminal stop rule.
Approval covers reversible steps inside that packet only. Ask again if the target,
data boundary, cost cap, resource envelope, or scientific interpretation changes.
Execute and monitor
- Stage into a run-specific directory. Verify input and code hashes remotely.
Never persist credentials in commands, scripts, logs, artifacts, notebooks, or
environment snapshots.
- Smoke-test on the smallest input. Do not run long, GPU, or high-CPU work on an
SSH or cluster login node.
- For Slurm, submit with
sbatch; set account, partition, time, CPU, memory, GPU,
and %x-%j log paths explicitly. Use set -euo pipefail, change to the submit
directory, handle signals, and write resumable checkpoints.
- For a cloud GPU, use the already configured provider interface, pin the image
and region/type, enforce the approved cap, persist outputs outside ephemeral
storage, and terminate the instance after retrieval or failure.
- For object storage, preview transfers, keep destinations private, preserve
checksums, and use least-privilege existing credentials. Do not make a bucket
or object public for convenience.
- Record remote job/instance IDs immediately. Monitor without busy polling;
capture terminal status, exit code, resource accounting, logs, checkpoints,
and failure cause. Resume from a verified checkpoint instead of duplicating a
job. Cancel only on user request, a safety limit, or the approved stop rule.
- Retrieve outputs into
artifacts/<run-id>/, verify hashes, record costs and
retained remote resources, then run $science-provenance and $science-review.
Stop conditions
Stop for unknown ownership, host-key changes, missing authorization, secrets in a
planned command, unapproved private-data transfer, unavailable cancellation,
uncapped paid resources, a full filesystem, or a resource request unsuitable for
the target. Report the blocker without weakening the boundary.
Decision contract
Before execution, record the target, transfer boundary, argv or workflow, input hashes, runtime or container, resource and wall-time caps, cost cap, checkpoints, cancellation, outputs, and scientific acceptance criteria. Remote writes or allocation require explicit approval covering the exact job spec.
Outputs
Save the job spec, preflight and approval receipts, job ID and states, stdout and stderr hashes, failure class, checkpoint references, collected output hashes, and $science-provenance and $science-review handoffs.
Boundaries
Do not probe credentials, submit to unapproved accounts or partitions, stage sensitive data without approval, busy-poll, hide failed attempts, or treat process exit zero as scientific success.
