Skip to content
Back to skills

Svsi

BSecurity

Pre-revisión semántica de informes de internacionalización antes de subir al Drive compartido. Detecta inconsistencias entre cuestionario, diagnóstico y plan. Propone correcciones focales (surgical_edit) en sesión o genera handoff para tmux (regen_section). Trigger: "revisa el informe de X", "pre-review X", "checkea X antes de Drive", "/svsi-review X".

  • 11 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added June 6, 2026
ai-agentspythonbash

Works with

  • claude code
  • cli

Security analysis

B75/100
  • criticalAccesses sensitive system or user directories

Pro shows the line behind each finding and how to fix it

Scanned June 6, 2026

npx -y skills add senda-labs/DQIII8 --skill svsi --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Svsi?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Svsi
[![Security: B — Skills Directory](https://www.skillsdirectory.com/api/skills/senda-labs-svsi/badge)](https://www.skillsdirectory.com/skills/senda-labs-svsi)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: svsi
description: >
  Pre-revisión semántica de informes de internacionalización antes de subir al Drive compartido.
  Detecta inconsistencias entre cuestionario, diagnóstico y plan. Propone correcciones focales
  (surgical_edit) en sesión o genera handoff para tmux (regen_section).
  Trigger: "revisa el informe de X", "pre-review X", "checkea X antes de Drive", "/svsi-review X".
command: /svsi-review
allowed-tools: [Bash, Read, Edit, Grep]
user-invocable: true
---

# /svsi-review — Pre-revisión Semántica de Informes

Proyecto: `/root/dqiii8/my-projects/intl-reports/`
Plan canónico: `docs/SVSI_PLAN.md` (v1.3)
Engine: `tools/review/engine.py` → `review_slug(slug)`
Reporter: `tools/review/reporter.py` → `write_report(report)`

## Cuándo usar

- Antes de subir un informe al Drive de ANOVA
- Cuando un consultor pida verificar un informe específico
- Para detectar inconsistencias que `qa_pre_render` existente no captura
- Trigger explícito: "revisa {slug}", "/svsi-review {slug}", "pre-review {slug}"

## Cuándo NO usar

- Para verificar la generación misma (eso lo hace `qa_pre_render` dentro del pipeline)
- Para regenerar secciones (eso requiere tmux externo con `core.cli`)
- Para batch overnight (usar `scripts/pre_drive_review.py` desde tmux externo)

## Flujo en sesión (Interactive Mode)

**Paso 1 — Ejecutar engine:**
```bash
cd /root/dqiii8/my-projects/intl-reports
python3 -c "
from tools.review.engine import review_slug
from tools.review.reporter import write_report
report = review_slug('{slug}')
write_report(report)
print(f'Verdict: {report.summary[\"verdict\"]}  C={report.summary[\"critical\"]}  H={report.summary[\"high\"]}  issues={len(report.issues)}')
"
```

**Paso 2 — Leer report estructurado:**
```bash
cat companies/{slug}/review/review_report.json
```

**Paso 3 — Para cada issue, según `remediation`:**

| remediation | Acción en sesión |
|-------------|-----------------|
| `surgical_edit` | Proponer Edit usando `edit_hint.path` + `edit_hint.new_template`. Pedir confirmación. Aplicar. |
| `regen_section` | Mostrar `handoff_command`. Escribir `data/svsi_handoff_{slug}.txt`. NO ejecutar. |
| `human_review` | Mostrar al usuario para decisión manual del consultor ANOVA. |
| `dismiss_only` | Mostrar como informativo. Preguntar si registrar como FP. |

**Paso 4 — Re-render DOCX** (solo si se aplicaron surgical_edits):
```bash
python3 scripts/render_batch.py --slug {slug} --doc-type plan --force
```

**Paso 5 — Re-validar** (máximo 3 iteraciones del loop):
```bash
python3 -c "
from tools.review.engine import review_slug
report = review_slug('{slug}')
print(f'Nuevo verdict: {report.summary[\"verdict\"]}')
for i in report.issues:
    print(f'  {i.rule_id} {i.severity}: {i.title}')
"
```

**Paso 6 — Reportar al usuario:**
- N issues resueltos quirúrgicamente (surgical_edits aplicados)
- M issues en handoff a tmux (comando exacto)
- K issues para consultor ANOVA
- Estado final: PASS/WARN/FAIL

## Restricciones absolutas

- NUNCA ejecutar `core.cli run/resume` — CLAUDECODE=1 bloquea el orchestrator
- NUNCA editar `tools/agent_writer.py`, `tools/block_writer.py`, `core/`
- NUNCA declarar "informe listo" sin verificar que el DOCX tiene > 500 KB
- Las correcciones se proponen y confirman antes de aplicar — no edición silenciosa
- Máximo 3 iteraciones del loop detect → edit → re-render → re-validate

## Comandos de referencia

```bash
# Engine + reporter (una empresa)
python3 -c "from tools.review.engine import review_slug; from tools.review.reporter import write_report; r=review_slug('{slug}'); write_report(r)"

# Ver report MD generado
cat companies/{slug}/review/review_report.md

# Re-render DOCX tras surgical_edit
python3 scripts/render_batch.py --slug {slug} --doc-type plan --force

# Batch shadow desde tmux (NO en Claude Code)
python3 scripts/pre_drive_review.py --slug-list data/tanda3_slugs.txt

# Stats acumuladas
python3 scripts/pre_drive_review.py --stats

# Marcar FP
python3 scripts/pre_drive_review.py --mark-fp {issue_id} --note "motivo"
```

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.