Add scnet-hpc skill

[lql341]

What

Adds scnet-hpc, a skill for operating SCNet HPC clusters through profile-based SSH and Slurm workflows.

Scope

• Profile-specific SCNet connection, partition, memory, module, and accelerator guidance.
• SSH setup, Slurm job generation, cluster discovery, profile refresh, and compute-node probes.
• Hygon DCU/DTK development and evidence-based compatibility checks.
• Linux and macOS native scripts; Windows through WSL2.

Structure

• SKILL.md contains routing, safety boundaries, and operational invariants.
• references/ contains focused setup, environment, troubleshooting, accelerator, and compatibility guidance.
• scripts/ contains deterministic shell/Python helpers.
• clusters/ contains public profile templates and packaged SCNet profiles.

Verification

• Agent Skill validator passes.
• bash -n passes for all shell scripts.
• python3 -m py_compile passes for the compute probe.
• git diff --check passes.

The canonical source is maintained at https://github.com/lql341/scnet-hpc. The submission intentionally excludes the repository installer, CI files, tests, and DSH-specific bundle files.

[98zc5g5jyw-arch]

Babysit review — dangling script reference

Extensive, well-organized HPC skill. Blocking issue: SKILL.md references ./scripts/install.sh but that file is NOT in this PR (only _common.sh, probe-cluster.sh, new-job.sh, refresh-cluster.sh, setup-ssh.sh, etc.) — the README.zh-CN also references it. Either include install.sh or fix the references. Private-key handling with BEGIN PRIVATE KEY validation looks safe.

[lql341]

Thanks for catching this. You’re right—the installer was unintentionally omitted from the PR. I’ll add `scripts/install.sh` and verify the references in both `SKILL.md` and `README.zh-CN` in the next update.

[lql341]

Correction to my previous reply: the standalone repository installer is intentionally excluded from this upstream submission. I removed the stale `install.sh` references from `references/quickstart-en.md` instead and renumbered the remaining sections. The submitted skill is now self-contained.

[98zc5g5jyw-arch]

Babysit follow-up — fix confirmed ✅

Verified the latest diff: the dangling installer reference is gone and LICENSE.txt (MIT) is now included. Thanks for the fast turnaround.

One last nit (non-blocking): SKILL.md frontmatter still lacks a license: key — e.g. add license: Complete terms in LICENSE.txt for consistency with repo convention and marketplace tooling that reads the frontmatter.

— Hermes Agent (cron babysitter)

[lql341]

The requested license metadata has been added to the SKILL.md frontmatter:

license: Complete terms in LICENSE.txt

Commit: f213217

[98zc5g5jyw-arch]

Reviewed by Hermes Agent. Frontmatter compliant (name/desc 413 chars/license with LICENSE.txt shipped). Scripts are safe: r() and ssh_cmd() use parameterized ssh with BatchMode=yes + ConnectTimeout, no eval, no shell=True; rm -rf guarded with quotes and EXIT trap; cluster .conf files contain hostnames only, no credentials. Solid HPC skill. LGTM.