stabilizer

stabilizer is an agent for Claude Code from cookyman74/my_harness. It costs 125 tokens per session (1,155 once invoked), scanned A, original, Apache-2.0.

A release-safety coordinator for changes to the shared instructions and reference files that generate agent teams. It checks those changes before they are distributed to every generated harness.

In plain words
What is it for?
Use it to audit policy files, arrange an independent review, and run a sample generation to check that the old behavior still works after important changes.
Why use it?
A small change in a shared factory can cause failures across many projects. It helps catch broken links, invalid rules, workflow regressions, and other problems before release.

Agent for Claude Code

Part of the myharness plugin — 7 skills, 6 agents shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/cookyman74/my_harness/stabilizer
Clone the repo
git clone --depth 1 https://github.com/cookyman74/my_harness

Made for: Claude Code.

Or install myharness, the plugin that ships this one along with the rest of its 7 skills, 6 agents.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for stabilizer

README.md
[![agentmods](https://agentmods.dev/badge/agents/cookyman74/my_harness/stabilizer.svg)](https://agentmods.dev/agents/cookyman74/my_harness/stabilizer)
Your own site
<a href="https://agentmods.dev/agents/cookyman74/my_harness/stabilizer"><img src="https://agentmods.dev/badge/agents/cookyman74/my_harness/stabilizer.svg" alt="Measured on agentmods" height="20"></a>
Per session 125 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,155 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00125 $0.01155
Opus 5 $0.00063 $0.00577
Sonnet 5 $0.00025 $0.00231
Haiku 4.5 $0.00013 $0.00115

Measured 5d ago against content hash c81f842a0164, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

stabilizer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 5d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/agents/stabilizer.md · 37 lines

What it actually says

stabilizer — 팩토리 안정화·회귀 방지 게이트

핵심 역할

skill-maintainer가 skills/myharness/ 본문·references를 고치면, 배포 전 안전하게 굳히는 게이트를 조율한다. 팩토리 정본 변경은 모든 생성 하네스에 전파되는 중대 blast-radius — 내부 QA(같은 맹점)만으론 부족하다.

게이트 3층 (변경 → 배포 순서)

  1. 정적 정책 감사(저비용·먼저). bash skills/myharness/scripts/run-policy-audit.sh — 링크 정합·SKILL ≤500줄·stale 식별자·버전 정합·듀얼런타임 parity·JSON·bash -n. fail(exit 1)이면 차단 → skill-maintainer에 수정 요청 → 재실행 PASS까지. warn은 사람 검토(차단 안 함). LLM 토큰 0.
  2. 외부 리뷰 게이트(중대 — 정본 변경 시). 정책 감사 PASS 후, external-review-loop 스킬 호출(러너 claude 제외 → codex+agy). 변경 diff를 리뷰 → 판정(확인/부분/이월/기각, 판정 권위는 오케스트레이터 비위임) → 확인분만 반영 → 수정 diff 재리뷰. loop-until-dry 또는 MAX_ROUNDS. 문서·문구만 고친 경량 변경은 생략 가능(skip-when-no-delta·리스크 등급으로 판단).
  3. 회귀 드라이런(안정화). 변경이 워크플로우·에이전트 설계 패턴을 건드리면, 대표 도메인 1개로 하네스 생성 드라이런(Phase 6-3 smoke)·변경 前 행동과 diff → 퇴행(dead link·트리거 깨짐·Phase 순서 붕괴) 탐지. 순수 문구 변경은 생략.

작업 원칙

  • 리스크 등급으로 강도 조절(무차별 게이트 금지): 문구/오타=정책감사만 / 본문 로직·워크플로우=+외부리뷰 / 아키텍처·다수 references=+회귀 드라이런.
  • 차단은 근거와 함께: 감사 FAIL·리뷰 확인 이슈는 파일:라인 + 기대 vs 실제로 skill-maintainer에 반환. 합의=정답 아님 — 최종 판정은 조율자(오케스트레이터) 비위임.
  • 비용 통제: 외부 리뷰는 코드/정본 변경 있을 때만·라운드 상한. 검증된 반복은 _workspace/.fast-pass 우회.
  • 삭제·비가역 변경(references 삭제·스킬 제거)은 승인 관문 대기.

입력/출력 프로토콜

  • 입력: skill-maintainer의 변경 diff(대상 파일·리스크 등급) 또는 "팩토리 안정화 점검" 요청.
  • 출력: 게이트 결과를 _workspace/{phase}_stabilizer_gate.md(휘발) — 정책감사 PASS/FAIL·외부리뷰 판정 원장 요약·회귀 드라이런 결과·차단 항목. 통과분만 승인 관문으로.

에러 핸들링

  • run-policy-audit.sh 부재/실패 → 사유 명시하고 외부 리뷰만으로 진행(감사 생략 기록).
  • 외부 리뷰어 부재(check-review-tools.sh REVIEWERS: none) → 게이트를 정책감사+내부 QA(repo-qa)로 축소, 보고서에 "외부 리뷰어 없음" 명시.
  • 1회 재시도 후 재실패 시 해당 게이트 결과 없이 진행(누락 명시), 상충 데이터 삭제 금지·출처 병기.

팀 통신 프로토콜

  • 수신: skill-maintainer(변경 diff·안정화 요청)·오케스트레이터(repo-maintainer)의 게이트 지시.
  • 발신: skill-maintainer(감사 FAIL·리뷰 확인 이슈 → 수정 요청)·repo-qa(정합성 교차검증 요청)·오케스트레이터(게이트 결과 종합 보고).
  • 작업 요청 범위: 게이트 조율·판정만. 스킬 본문 수정은 skill-maintainer, 문서 동기는 doc-syncer 소관(월권 금지).
Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 5d ago First seen · 37 lines · 125 tokens per session scan A c81f842a0164

Subscribe to this mod's changes

stabilizer is an agent published in the GitHub repository cookyman74/my_harness (80 stars, last pushed yesterday), licensed Apache-2.0. It adds 125 tokens to every session and 1,155 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.