Skill: pr-agent-judge (SPEC-124)

SkillAI & models

Use when adding pr-agent as an external judge in the Code Review Court.

Available today. Use it from your connected AI after setup.

No other account needed.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Skill: pr-agent-judge (SPEC-124) skill

What this skill tells your AI

The instructions your AI receives, as published by gonzalezpazmonica/pm-workspace in .claude/skills/pr-agent-judge/SKILL.md and read by ahel’s review.

Wrapper sobre qodo-ai/pr-agent (10.9k ⭐, 60.1% F1) como juez externo opt-in del Code Review Court.

Rol dentro del Court

Court actual: 4 jueces (correctness, security, architecture, cognitive). Con opt-in: 5 jueces → añade diversidad (jueces externos con benchmark público).

Se activa solo si COURT_INCLUDE_PR_AGENT=true en pm-config.md.

Invocación

# Modo local (sin CI)
bash scripts/pr-agent-run.sh --pr-number 593 --mode review --output court-format

Salida JSON compatible con formato Court:

{
  "judge": "pr-agent",
  "version": "qodo-ai/pr-agent@0.27",
  "verdict": "approve | request_changes | comment",
  "findings": [
    {
      "severity": "medium",
      "category": "correctness",
      "file": "src/foo.ts",
      "line": 42,
      "message": "..."
    }
  ],
  "summary": "..."
}

Modos

review — code review completo
describe — descripción automática del PR
improve — sugerencias de mejora

Requirements

pr-agent instalado: pip install pr-agent o Docker
GITHUB_TOKEN env var (para acceder al PR)
Anthropic API key o compatible (PR_AGENT_MODEL en pm-config.md). Por defecto resuelto via tier: mid → ~/.savia/preferences.yaml model_mid.

Si pr-agent no está instalado, el wrapper falla gracefully reportando {"status":"SKIPPED","reason":"pr-agent not installed"} sin bloquear al Court.

Feature flags en pm-config.md

COURT_INCLUDE_PR_AGENT   = false        # default opt-in
PR_AGENT_VERSION         = "0.27"       # pin
PR_AGENT_MODEL           = "mid"        # tier, resuelto via preferences.yaml
PR_AGENT_MAX_LINES       = 1000         # evita token blowout

Relación con agentes internos

Complementa correctness-judge, security-judge, architecture-judge, cognitive-judge.
NO reemplaza ninguno — opt-in para diversity, no para rebaja.
Orchestrator (court-orchestrator) agrega los 5 verdicts con policy ya definida.