Cologne Skills: Playbooks de Segurança e Manutenção para Sanskrit Lexicon
sanskrit-lexicon/cologne-skills
Skills reutilizáveis de segurança e manutenção para o ecossistema Cologne
Ver no GitHubColeção de skills portáteis para Claude Code que encapsulam procedimentos testados em produção para segurança, manutenção e digitalização de repositórios da organização Sanskrit Lexicon (Cologne Digital Sanskrit Dictionaries). Inclui auditorias de segurança, triagem de alertas, normalização de markup, correção de textos e OCR de frentes de dicionários, com heurísticas e gotchas já resolvidos.
Instalação
npx convez-ai sanskrit-lexicon/cologne-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auditoria de segurança org-wide (GitHub Actions, secrets, SAST)
- Busca e correção de vulnerabilidades XSS/SQL injection em PHP
- Triagem e remediação de alertas CodeQL e Semgrep
- Normalização de markup e correção de textos em PRs validadas
- OCR de frentes de dicionários com tradução EN/RU
- Manutenção de journals de estado (_ai_state.md) e handoff entre sessões
Veredito

Repositório especializado e bem documentado para times de segurança/manutenção que trabalham com o ecossistema Cologne. Valor real em playbooks contextualizados (escaping correto, heurísticas de falsos positivos, gates de segurança) em vez de templates genéricos; requer familiaridade com o domínio Sanskrit Lexicon e GitHub Actions para máximo proveito.
README
A tradução para o português deste README ainda está na fila.
cologne-skills
Created: 16-06-2026 · Last updated: 11-07-2026
Portable Claude Code skills for security, maintenance, and digitization work across the Sanskrit Lexicon (Cologne Digital Sanskrit Dictionaries) GitHub org.
These are the shareable cut of the personal /cologne-* command family — version-controlled so they work on any machine and can be shared with collaborators, instead of living only in a personal ~/.claude/commands/. Each skill encodes a battle-tested playbook (escaping decision tables, false-positive heuristics, the gotchas) rather than a generic template.
Skills
Codex + Claude Code portable skills
Reusable SKILL.md folders live in skills/. They are written as portable runbooks:
Codex can invoke them as skills, while Claude Code can read the same SKILL.md files and follow
the instructions manually.
| Skill | What it does |
|---|---|
ai-state-journal-maintainer | Maintain tracked .ai_state.md session journals and handoff state. |
cdsl-newsletter-publisher | Prepare paired CDSL newsletter blog posts and email drafts. |
claude-codex-skill-porter | Audit/update skills so they work in both Codex and Claude Code. |
cologne-markup-batch | Prepare safe Cologne dictionary markup-normalization batches. |
cologne-question-research | Research Cologne editorial questions before source edits. |
cologne-text-correction-pr | Prepare validated Cologne text-correction PRs with registry/read-only gates. |
commentary-strategies-editorial-pipeline | Sync CommentaryStrategies publication state across manuscripts, hubs, and journals. |
dependabot-pr-triage | Triage dependency update PRs and recommend merge/hold/close. |
publication-readiness-auditor | Audit scholarly article readiness gates and metadata consistency. |
sundara-lexical-layer-qa | QA Sundara lexical, etymology, and cross-text annotation layers. |
Security & maintenance
| Command | What it does |
|---|---|
/cologne-php-xss-sweep <repo|all> | Find + fix reflected-XSS / SQL-injection / injection in a repo's PHP web-frontend, with context-correct escaping (HTML body/attr → htmlspecialchars; JS-in-<script> → json_encode(JSON_HEX_*); URL-in-JS-in-attribute → urlencode; JSONP → whitelist; SQL literal → prepared/''; PCRE → preg_quote). PR-only. |
/cologne-security-audit-all [nondict|dict|all] | Org-wide audit: GitHub Actions (pwn-request / script-injection / token scope), committed secrets (incl. the Firebase-web-key vs Cloud-key trap), and SAST coverage. Read-mostly; PRs only for hardening. |
/cologne-alert-triage <repo> | Triage a repo's CodeQL + Semgrep alerts: fix the genuinely-exploitable ones (PR), dismiss false-positives/won't-fixes with written justifications. Handles the Semgrep echoed-request taint-mode quirk and "php → Semgrep, never CodeQL". |
Digitization
| Command | What it does |
|---|---|
/cologne-preface-ocr <CODE|all> | OCR a dictionary's front matter (title pages, prefaces, abbreviation lists, addenda) from the csldoc scans into faithful Markdown, then add English + Russian translations, consolidated single-file editions, and a README index. Bakes in the vision-OCR playbook: crop scans to native-resolution column-bands ≤1900 px before reading (a downsampled full page yields fluent-but-fabricated text), trust the toctree page order, keep Sanskrit/Devanāgarī verbatim, Cyrillic names in Russian, omit digitizer stamps, never commit temp crops. First run produced PWG/prefaces/ (27 pages × 3 languages). |
Example invocation
Real skill, real trigger — skills/cologne-question-research/SKILL.md
fires on an issue like this (its actual description: frontmatter):
"Use when an issue is labeled question, asks for scholarly judgment, compares readings across dictionaries or scans, lacks an exact correction, depends on the meaning of a local marker, abbreviation, tag, name, or editorial policy, or requires evidence before deciding whether a CDSL markup or text change is warranted."
Its actual workflow (from the same file) separates evidence-gathering from
editing: state the research question in one sentence → gather primary local
evidence (source record, generated XML, scans, neighboring/parallel dictionary
entries) → classify the outcome as one of answer-only /
ready-for-correction / ready-for-markup-batch / needs-human-review → if
ready-for-correction, hand off to $cologne-text-correction-pr rather than
editing source in the same pass. Invoke it directly against a real issue with:
/cologne-question-research PWG 123
Agents
Read-only worker agents in .claude/agents/ that the skills fan out (and that you can invoke directly). Copy them to ~/.claude/agents/ the same way as the commands.
| Agent | Role |
|---|---|
adversarial-verifier | Verify one finding → CONFIRMED / PLAUSIBLE / REFUTED with the trigger. The check that stops both shipped bugs and buried real issues. |
cologne-security-reviewer | Hunt injection in a repo's web code with the escaping playbook; returns ranked candidates (no fixes). |
fact-check-against-source | Verify every claim in a doc/report against ground-truth code/data; flags inaccuracies with quotes. |
cologne-data-integrity-auditor | Audit data repos for corpus-signal leakage, overlay-wipe seeders, lossy normalization, generated-vs-canonical drift. |
All are read-only (no Edit/Write; no gh pr create/merge, no git push) — the skill or main loop does the PR-ing, so a fanned-out agent can never mutate a repo.
Install
As Codex skills: copy or symlink the desired folders from skills/ into
~/.codex/skills/.
As Claude Code runbooks: open the relevant skills/<name>/SKILL.md and follow it directly.
Each skill includes a compatibility rule for Codex and Claude Code.
As personal commands (available in every Claude Code session, any directory):
# copy them
cp .claude/commands/cologne-*.md ~/.claude/commands/
# …or symlink, so they stay in sync with this repo
ln -s "$PWD/.claude/commands/"cologne-*.md ~/.claude/commands/
As project commands (available only when working inside a specific repo): copy the .md files into that repo's .claude/commands/.
Invoke with the slash name, e.g. /cologne-php-xss-sweep csl-santam.
Prerequisites
- GitHub CLI (
gh) authenticated withreposcope (andread:projectfor any project-board work). phpon PATH for thephp -llint step (Windows: e.g. XAMPP'sC:\xampp\php\php.exe— alias it tophpor adjust the lint command).python3for the helper scans (enumeration, workflow-risk, secrets).- For
/cologne-preface-ocr:curl(scan download) and Pillow (pip install pillow) for native-resolution image cropping. OCR is done by the model's own vision — no Tesseract/OCR engine needed.
Conventions baked in
- PR-only: branch off
origin/<default>, never push the default branch; one tight PR per repo. - Resolve the default branch per repo (
gh api repos/{owner}/{name} --jq .default_branch) — it is not alwaysmaster(e.g.mw-devdefaults tomain). - Fork gotcha: pass
--repo <owner>/<name>togh pr create, or a PR may target a parent fork. - Check
origin/<default>+ open PRs first — an external actor may run the same sweep; close duplicates rather than rebasing. - Windows/UTF-8: Python scripts set
sys.stdout.reconfigure(encoding='utf-8')and passencoding='utf-8'tosubprocess.run.
Provenance
Distilled from a security + maintenance pass across the Sanskrit Lexicon org (reflected-XSS sweep, SQLi/ReDoS hardening, CodeQL/Semgrep triage, GitHub Actions audit). The companion mechanical-batch skills (cologne-codeql-all, cologne-dependabot-automerge-all, the issue/tooling runbooks) and their .deploy_*.py helpers are not yet included here — they can be folded in once their machine-specific paths are parameterized.
License
MIT — see LICENSE.
Dr. Mārcis Gasūns