HPC pack

Skills for running and inspecting work on an HPC cluster from Clio Coder. The pack joins the cluster's own tools into one workflow: the scheduler, the software environment, the parallel I/O stack, the I/O profiler and a durable record of what ran. It is content only. Installing it starts nothing and declares no MCP server.

clio-coder library install skill:slurm-jobs --project
clio-coder library install plugin:hpc --project

The pack requires the slurm-jobs skill package, which submits through the clio-kit Slurm MCP server. Run /library reload in a running session after installing.

Skill	What it does	Drives
`slurm-jobs` (required package)	Submits a job or job array, polls each task to a terminal state, reports each task's exit and output	`mcp_slurm__*` from `clio-kit mcp-server slurm`
`hpc-sweep`	Plans a parameter sweep as a job array with a dry run, submits it on the operator's word, checks each run with a declared verify check, and writes a results table that keeps failed and unverified runs	`assets/scripts/sweep.py`, `mcp_slurm__*`, `verify`, `clio-coder claims`
`hpc-modules`	Finds and loads Lmod modules and Spack installs, keeping each load in the shell that uses it	`module`, `$LMOD_CMD`, `spack` over `bash`
`hpc-parallel-io`	Correctness rules and tuning for parallel HDF5 and MPI-IO, checked by readback and Darshan counters	`h5pcc`, `h5py`, `mpirun`, `lfs`
`hpc-darshan`	Summarizes a Darshan log and names the counters behind each I/O finding	`darshan-parser` through `assets/scripts/darshan_summary.py`
`hpc-chronolog`	Records run events in ChronoLog and reads them back: chrono-mcp on released 3.x, or chronolog-mcp with a completeness verdict on unreleased 4.0 builds	`chronolog-mcp`, and `chronolog` on 4.0

Setup the operator owns

Slurm. Declare the clio-kit Slurm server in the user file <Clio config directory>/mcp.yaml with the id slurm, as the Slurm guide shows. A user declaration with that id is what makes polling run without approval prompts, so the pack does not declare the server itself. Give the entry an absolute cwd, such as your scratch directory. Jobs run from the server's working directory and write logs/slurm_output/ there. Without cwd, a user entry runs in Clio's config directory, and the job logs land in it. Submitting and cancelling ask at default, where the safety judge reviews them against the operator's messages, and a headless run denies what the judge does not allow.
ChronoLog. Declare ChronoLog's MCP server in the same user file, with its connection settings in env and per-tool toolActionClasses; hpc-chronolog shows both entries. The released line is ChronoLog 3.x (latest 3.4.0), whose server, chrono-mcp, ships in the ChronoLog source tree at plugins/chrono-mcp. ChronoLog 4.0 is unreleased development: it exists only as a source build, with its own chronolog launcher and chronolog-mcp.
Environment. Clio passes an MCP server only PATH, HOME, USER, LOGNAME, LANG, LC_ALL, the temp and XDG_* directories and a few platform variables, plus the env its declaration names. A variable a server needs (CHRONO_HOST for 3.x, CHRONOLOG_MCP_IDENTITY for 4.0, CLIO_KIT_CACHE_DIR) goes in that env, never as a credential value.

Offline behavior

The pack makes no network calls of its own. Its skills drive tools installed on the cluster. On an air-gapped login node they work with these conditions:

clio-kit mcp-server slurm runs each server from its own lock file in a uv environment under $XDG_CACHE_HOME/clio-kit or ~/.cache/clio-kit (set CLIO_KIT_CACHE_DIR in the declaration's env to move it). The first launch downloads that environment and fails without network. Every later launch from the warm cache works offline. Launch it once while online, or copy a warm cache from a connected node.
spack install needs network or a configured mirror. The skills run it only when the user asks. Lookups (spack find, spack location) are local.
Lmod, Darshan, parallel HDF5, MPI-IO and ChronoLog are local or cluster-network tools, so they need no internet access once the site has installed them.
The pack declares no Darshan, Spack or ChronoLog MCP server (see below), so nothing else needs a first online launch.

Not yet in the pack

clio-kit also ships Darshan, Spack, Lmod and ChronoLog MCP servers. The pack does not declare them yet. The Darshan and Spack servers wait on a live check against a real darshan-util and Spack install. The Lmod server runs module as a program, but module is a shell function, so its loads cannot reach the caller. clio-kit is retiring its ChronoLog server in favor of ChronoLog's own chrono-mcp (not yet in a clio-kit release). The skills drive the command-line tools directly instead.

Checks

assets/scripts/darshan_summary.py is the pack's declared check script (stdlib Python 3.10+). It reads darshan-parser's text output and never writes to the log. It was written against the darshan-util 3.5.0 output format.

assets/scripts/sweep.py (Python 3.10+) plans and tabulates sweeps. It reads the sweep file as YAML when PyYAML is importable and as JSON otherwise, writes only under the sweep's sweeps/<name>/ directory in the project, and never submits. Its per-run verdicts come from the claims Clio records when verify(check="<id>/<run>") judges a run, which needs a .clio-coder/verifiers.yaml check that declares runs; the hpc-sweep skill shows the entry. Task states come from sacct, read-only.

Licenses and terms

The pack itself is Apache-2.0. It bundles no third-party code or data, and it queries no external data service, so no data terms of use apply. The tools it drives keep their own licenses: clio-kit (BSD-3-Clause), Lmod (MIT), Spack (Apache-2.0 or MIT), Darshan (its University of Chicago copyright notice), ChronoLog (BSD-2-Clause), and the site's MPI and HDF5 builds.