claude-skills

Guilhem Lavaux's personal Claude Code plugin marketplace.

Install

/plugin marketplace add <this-repo>
/plugin install forgejo
/plugin install cv
/plugin install euclid
/plugin install report

Plugins

forgejo

Manage Forgejo repos via the fj CLI (forgejo-cli). Skills:

• forgejo-pr — create/update pull requests from current branch
• forgejo-issue — create, search, triage, update, close issues
• forgejo-release — manage releases/tags, upload/download assets
• forgejo-actions — inspect task status, dispatch workflows, manage variables/secrets

Requires fj CLI installed and authenticated.

cv

Academic CV tooling. Skills:

• cv-data-store — build and maintain a machine-checked YAML store of CV facts at ~/.config/academic-cv: positions, publications, grants and compute allocations, supervision, talks, meeting organisation, commissions of trust, teaching, outreach, collaborations, prizes and narrative rubrics, person-scoped so several PIs' CVs can live side by side. Extracts from a LaTeX dossier, an ERC Part B1 PDF or an ADS query; merges on merge keys rather than ids so a re-ingest inserts nothing; records a contradiction between two documents as a disputed: block instead of overwriting it; and validates the whole store with a Python verifier covering syntax, references, partial dates, amounts and units, controlled vocabulary and aggregate consistency (check-cv.sh --rules lists all 92 rules). Commands: /cv-store-init, /cv-store-ingest, /cv-store-check.

• cv-render-cnrs — turn the store back into LaTeX for a CNRS/CNU dossier: one .tex fragment per rubric (35 of them, named by rubric and window, never by section number) matching the dossier's own macros, longtable style and French typography, plus a drift report classifying every difference as content, style or anchor, and a page-budget check measured from the compiled .toc rather than estimated. Never edits the dossier — the diff command is granted no Edit or Write tool. Commands: /cv-render, /cv-render-diff.

• cv-render-website — turn the store back into the 5 fixed-name RST fragments a personal website's CV page consumes (short.rst, grants.rst, responsabilities.rst, activities.rst, outreach.rst), matching the site's own grid-table/bullet-list RST style with an .en-preferred/.fr-fallback policy where every fallback is a named finding. Curated fragments (grants, responsabilities, outreach) render an honest placeholder until a website_highlight selection list is written in selections.yaml — no ingest, no guessed curation. Plus a --diff mode classifying drift as content/style/anchor, with --only-en-gaps narrowing to records genuinely missing English text. Never edits the live site — the diff command is granted no Edit or Write tool. Commands: /cv-render-website, /cv-render-website-diff.

• cv-activity-report — update French academic evaluation dossiers (CNU/HCERES, évaluation à vague / à mi-vague) written in LaTeX: CV / "Rapport d'activité" / "Productions scientifiques" sections, enforce page limits, the 10/5-semester window and the 10/5 publication cap, cross-check publications against arXiv/ADS, require each curated production to state its portée, its impact and the candidate's personal contribution plus a direct open-archive link to the full text itself (HAL preferred, arXiv accepted, per the CNRS open-science policy), audit the "Rapport d'activité" against the six expected content rubrics (activité de recherche, collaborations, contrats, place dans l'unité, mobilités, distinctions), and leave \textcolor{red}{BOT: ...} notes in the .tex source for anything unresolved

report

Compute-time allocation activity reports for the French national centres (GENCI/DARI: CINES Adastra, IDRIS Jean Zay, TGCC Irene), following the official CINES template Demande d'Attribution de Ressources Informatiques — Rapport d'activité. Skills:

• compute-report-init — start the report as a compilable French LaTeX document. Two steps: a YAML holding every fact the form asks for, then the .tex. The intermediate file is what makes the numbers checkable — hours are parsed (2,5 millions, 2.5M, 2 500 000) or refused by name, each machine line carries the unit the form itself prints for that partition, and the consumed/allocated ratio is checked against thresholds before any LaTeX exists. --from-store prefills from the cv-data-store store at ~/.config/academic-cv but never assigns a store record to a partition: {name: Irene, site: TGCC} could be SKL, Rome or V100, and those differ in unit. Every unanswered field is rendered as \textcolor{red}{BOT: ...}, visible in the compiled PDF rather than silently blank. Command: /compute-report-init.

• compute-report-review — audit an existing report and mark what is missing. Read-only script, no write path: it reports each finding with a line number and the exact BOT string to place there, and the insertion is done separately. Catches missing sections, unreplaced template placeholders, a missing dossier number, GPU hours written on a core-hour line, allocation and consumption blocks that list different machines, an unexplained under- or over-consumption, empty bibliography or conferences sections, and sub-projects over the 2-page limit — measured from the \label{sousprojet:...} page numbers in the compiled .aux, never estimated from source length. Command: /compute-report-review.

The machine list, units and thresholds live in plugins/report/assets/machines.yaml; a centre adding a partition is an edit to that file, not to the code.

euclid

Euclid Consortium effort tracking. Skills:

• euclid-effort-tracking — fill the Euclid Consortium semester effort declaration (the Euclid Tracking Tool / dynacase record) from mail, calendar, ADS and git evidence, weighted per the EC effort-tracking rulebook, emitting a paste-ready table, justification memo, and HTML worksheet. Commands: /euclid-tracker-init, /euclid-tracker-report.