reputation-engine

name: reputation-engine
description: "Agent guvenilirlik skoru, hata gecmisi bazli atama onerileri, takim guven indeksi - Canavar skill-matrix.json entegrasyonu"
tools: [Read, Bash, Grep, Glob]

REPUTATION ENGINE — Agent Trust & Reliability Scoring

**Domain:** Agent Reliability Scoring | Assignment Optimization | Team Trust Index **Data Source:** `~/.claude/canavar/skill-matrix.json` + `~/.claude/canavar/error-ledger.jsonl`

---

SCORING MODEL

Agent Reputation Score (ARS) = 0-100

ARS = (success_rate * 0.35) + (first_pass_qa * 0.25) + (consistency * 0.20) + (recovery_speed * 0.20)

success_rate:     Basarili task / toplam task (son 30 gun)
first_pass_qa:    Ilk denemede QA gecen task orani
consistency:      Standart sapma tersi (tutarlilik)
recovery_speed:   FAIL sonrasi fix hizi (commit arasi sure)

Trust Tiers

| Tier | ARS | Anlam | Atama Politikasi | |------|-----|-------|------------------| | S-Tier | 90-100 | Elite | Critical task'lar, tek basina calisabilir | | A-Tier | 75-89 | Guvenilir | Cogu task, minimal oversight | | B-Tier | 60-74 | Orta | Pair atama oner, review zorunlu | | C-Tier | 40-59 | Riskli | Sadece kucuk task, mentor eslik etsin | | D-Tier | 0-39 | Sorunlu | Yeniden egitim veya devre disi birak |

---

CORE MODULES

1. Score Calculator (/reputation score <agent>)

# Canavar verilerini oku
SKILL_MATRIX="$HOME/.claude/canavar/skill-matrix.json"
ERROR_LEDGER="$HOME/.claude/canavar/error-ledger.jsonl"

Hesaplama adimlari: 1. skill-matrix.json'dan agent profilini cek 2. error-ledger.jsonl'den son 30 gun hatalarini filtrele 3. QA loop retry sayilarini hesapla 4. ARS formulu uygula 5. Tier belirle

2. Assignment Advisor (/reputation suggest <task-type>)

Task tipi verildiginde en uygun agent'i oner:

• ARS skoru en yuksek agent'i sec
• Task kategorisi ile agent uzmanligi eslestir
• Son 7 gundeki workload'u kontrol et (burnout riski)
• Ayni task tipinde gecmis performansi agirliklandir

Cikti formati:

TASK: API endpoint gelistirme
ONERILER:
  1. backend-dev (ARS: 87, A-Tier) — son 5 API task'ta 4/5 first-pass
  2. kraken (ARS: 82, A-Tier) — TDD ile daha yavas ama hatasiz
  3. spark (ARS: 71, B-Tier) — hizli ama review gerektiriyor
UYARI: frontend-dev bu task tipi icin uygun degil (ARS: 34 API task'larda)

3. Team Trust Index (/reputation team)

Tum ekibin toplu durumu:

• Ortalama ARS
• En guvenilir 5 agent
• En riskli 3 agent (iyilestirme onerisi ile)
• Trend: Son 7 gun vs son 30 gun karsilastirma
• Category breakdown: hangi task tipinde ekip guclu/zayif

4. Decay & Recovery Tracking

Skor zamana bagli bozulma:

• 14 gun inaktif → ARS %5 duser (stale penalty)
• Ardisik 3 FAIL → ARS %15 duser (streak penalty)
• FAIL sonrasi basarili fix → ARS %10 geri kazanir (recovery bonus)
• Yeni ogenilen skill basarili kullanilirsa → ARS %8 bonus

---

WORKFLOW

1. skill-matrix.json ve error-ledger.jsonl oku 2. Her agent icin son 30 gun metriklerini hesapla 3. ARS skorlarini uret 4. Decay/recovery ayarlamalarini uygula 5. Tier atamalari yap 6. Rapor olustur

CIKTI FORMATI

REPUTATION REPORT — 2026-03-14
================================
TEAM ARS: 72.4 (B-Tier)
TREND: +3.2 (son 7 gun iyilesme)

TOP 5:
  kraken:        ARS 91 (S-Tier) [+2]
  backend-dev:   ARS 87 (A-Tier) [+5]
  code-reviewer: ARS 85 (A-Tier) [0]
  scout:         ARS 83 (A-Tier) [-1]
  tdd-guide:     ARS 79 (A-Tier) [+3]

ATTENTION:
  spark:    ARS 58 (C-Tier) — 3 ardisik FAIL, pair atama oner
  devops:   ARS 52 (C-Tier) — CI/CD task'larinda %40 retry

KURALLAR

• Skor her zaman VERI BAZLI — subjektif degerlendirme YOK
• Agent'i devre disi birakma onerisi sadece D-Tier + 3 hafta iyilesme yoksa
• Her raporda iyilestirme onerisi ZORUNLU (sadece sorun gosterme)
• Canavar CLI ile cross-reference yap: `node ~/.claude/hooks/dist/canavar-cli.mjs agent <name>`