AGENT SKILL

auto-metrics

Self-evaluation scoring. Trigger via !score or !metrics — not automatic.

License 许可
Apache-2.0
最近更新
2026-11-10
Stars
★ 0
Forks
0

01 前言 Front Matter

name
auto-metrics
triggers
!score, !metrics, explicit score/metric request, session end via !close
changelog
docs/ciclos/cycle28-20260815.md
description
Self-evaluation scoring. Trigger via !score or !metrics — not automatic.
token_budget
2800

03 SKILL.md 正文

When to Use

Only on explicit request (!score, !metrics, !close). NOT automatic after every task.

Migration: !metrics here = self-scoring — use !metricas for before/after deltas.

PRE-FLIGHT (HARD GATE)

If .learnings/bias-calibration.json exists with samples >= 2: check bitácora for [audit] {today}. None → FAIL — run !audit first. Never score without audit if biases exist.

CORRECTION (if audit available)

  1. Subtract each dim's avg offset from self-score BEFORE thresholds 2. Log Bias corrected: {dim}={offset} 3. Thresholds (<7→immune, ≥9→mem_save) 4. Append today's self/audit pair

7 Dimensions (1-10)

Dim 1-3 4-6 7-9 10
Correctness Wrong Partial/buggy Minor gap 0 rework
Tokens Verbose Some bloat Lean Minimal
ErrPrev Repeated Partial fix No repeat +Immune pattern
Skill Wrong None loaded Correct +anti-pattern
Speed >5 wasted Back-forth Efficient 0 redundant
Breadth 1 dim 2-3 dims All relevant +unexpected
SkillEval No baseline <10% delta >=10% >=20% +mem_save

ACTION

≥8 Maintain · 6-7.9 Light review · 4-5.9 Improve+immune+dream · <4 Full stop+root cause

STORAGE

mem_save(type="learning", title="auto-score:{task}", content="Correctness:X|Tokens:X|...|Avg:X.X|Pattern:{what}")

TREND (every 10 or session end)

mem_search(query="auto-score:", limit=20) → per-dim means, prev(5) vs recent(5). Drop >0.5 → immune-system. Avg<6 → gap analysis.

docs/skills/auto-metrics/reference.md

Anti-Rationalization

Rationalization Red Flag Verification
"score sin audit cuando hay bias" bias-calibration.json samples≥2 sin [audit] PRE-FLIGHT: [audit] hoy o FAIL→!audit primero
"ignorar bias calibration" offset no restado antes de thresholds CORRECTION: restar offset por dim + log bias
"auto-score sin verificación externa" self-score sin contraste auditor gaps>1.5→immune/full-stop per external-auditor

Red Flags

  • Doing work without checking output format → STOP, re-read skill
  • Second occurrence of same rationalization → force RED zone

Verification

  • Output matches skill ## Output contract + file:line citaton
  • cross-ref-check.ps1 → SKILL.md OK

Refs

Cross-Refs: performance-tracker | metricas | quality-gate | security-scanner | external-auditor | immune-system

04 同仓库更多 Skills