research

Plugin research from marketplace dansolombrino-skills in dansolombrino/skills. Commands read from the repo's manifest — today.

Skills (12)

• cineca-login

  Renew the CINECA Leonardo SSH certificate non-interactively via an OIDC password grant with programmatic TOTP, distribute it to the rig hub, and optionally run on a systemd timer so the 12-hour certificate never expires unnoticed. Use when the certificate is about to expire, when sweep-supervisor fi

• environment-sync

  Provision and verify exact, lock-keyed uv-managed Python wave environments across configured Research 2.0 rigs. Use when installing or repairing a supported project environment, building the environment for a wave or for the staged lock before the pre-dispatch smoke, diagnosing Python/package drift,

• experiment-design

  Compile a requested scientific step into a reproducible Research 2.0 experiment - NNN naming, hashed-v2 run_id election, checkpoints, schema-v3 status telemetry, offline metric logging, and smoke criteria. Use when creating a new experiment or sub-experiment, writing training, finetuning, or evaluat

• experiments-tracking

  Keep a Research 2.0 EXPERIMENTS.md true, derive active-run, lane, and wave ETAs from provenance-valid status plus artifacts, and prune finished wave worktrees. Use when runs are generated, launched, checked, reported, or discussed, when the user asks about experiment status or ETA, or when session-s

• research-journal

  Keep a Research 2.0 JOURNAL.md as the append-only chronological story of what was done, why, and what was learned. Use after meaningful scientific or engineering events, when a commit is blocked by the journal guard, and whenever the user asks to journal, log, or write up the work; never as factual

• research-project-init

  Scaffold a new or empty Research 2.0 repository - the standard experiment taxonomy, exact uv environment, rig configuration, status and run_id helpers, journal guard, host notes, and the project's private GitHub repository. Use when starting, initializing, bootstrapping, or scaffolding a fresh resea

• rig-board

  Read and maintain the fleet-wide GPU lane board shared by every Research 2.0 project on the hub, so any chat knows which rig GPUs are held, by which project and wave, since when, and with what ETA, plus the read-only Slurm queue of a cluster such as Leonardo. Use when the user asks whether the GPUs

• rig-sync

  Deploy exact tested Git revisions as per-wave worktrees and move artifacts across configured Research 2.0 GPU rigs, including the continuous hub-ward offload that keeps a disk-constrained rig usable. Use when validating declared rig paths, preparing a rig clone, copying the project .env or shell cac

• run-identity

  Own the hashed-v2 run_id identity of a Research 2.0 experiment - elect its identity params, look up a run hash or prefix (old hashes resolve through RUN_ID_HISTORY.json) or find the runs matching param values, and migrate an experiment to the next run_id version when a sweep adds or drops an identit

• sweep-dispatch

  Generate and launch experiment runs/sweeps across the GPU rigs (rig-4090, rig-3090-ti, rig-3080-ti, behemoth — also "4090", "3080 ti", "pro 6000", "bw", "blackwell", "server-pro-6000-bw") and the CINECA Leonardo Slurm cluster ("leonardo", "cineca", "slurm", "sbatch", "HPC") as one pull queue over a

• sweep-supervisor

  Supervise a dispatched Research 2.0 wave from the hub with no chat needed - a hub service feeds each GPU lane from one pull queue, relaunches lanes after a rig fault, offloads and collects, keeps the fleet board current, reports to a watching chat (and any `notify_command`), and wakes a fresh unatte

• visualizations

  Create or update Research 2.0 plots as self-contained interactive HTML from experiment evaluation outputs - titles that state the selected params by full name, meaningful axis and metric text, a PARAM_GLOSSARY.json with hover explanations, run selection inside the file, one file per figure under the

Subagents (2)

• plot-checker

  Render research plots, run the plot check, fix layout errors until clean, read the screenshots, and return the verdict with links. Use for any plot render or re-render in a research project, so the screenshots stay out of the main chat.

• wave-diagnoser

  Find out why a run, lane or wave failed or stalled, read-only, and return the cause with evidence and a recommended action. Use when a wave watch reports a failure or warning, or the user asks why something went wrong in a wave. Request the agent profile declared in the registry's [supervisor.agent]

Hooks (2)

• SessionStart

  python3 "${CLAUDE_PLUGIN_ROOT}/hooks/orchestration_note.py"

• SessionStart

  python3 "${CLAUDE_PLUGIN_ROOT}/skills/sweep-supervisor/scripts/wave_reminder.py"

Similar plugins