AGENT SKILL

guardrails

Use this skill before executing a risky or hard-to-reverse action — migrations, deletions, deploy/config changes, force operations — to derive a risk level and an APPROVED/REJECTED call with a rollback plan. Don't use it for routine read-only work.

License 许可
Apache-2.0
最近更新
2026-11-11
Stars
★ 8
Forks
5

01 前言 Front Matter

docs
{ "status": "draft", "updated": 1788220800000 }
name
guardrails
color
red
owner
swarmery-core
version
1.0.0
description
Use this skill before executing a risky or hard-to-reverse action — migrations, deletions, deploy/config changes, force operations — to derive a risk level and an APPROVED/REJECTED call with a rollback plan. Don't use it for routine read-only work.

02 资源文件 1

03 SKILL.md 正文

Purpose

A risky action gets a structured go/no-go before it runs, derived from an Impact × Reversibility matrix — not ad-hoc vibes. The output contract is binary: APPROVED (possibly with conditions) or REJECTED (with the reason) — never ambiguous.

The check

  1. Short-circuit: read-only actions (viewing files, running tests) are APPROVED immediately — no ceremony.
  2. Risk level = Impact (Low / High / Critical) × Reversible (Yes / No): Low×Yes → Low; High×Yes → Medium; Low/High×No → High; Critical×anything → Critical. The matrix and per-action guidance: resources/risk-checks.md.
  3. Deterministic checks for the affected stack (typecheck/lint/tests for code, dry-run/lint for config, --dry-run for migrations where the tooling offers one). A failing deterministic check is an automatic REJECTED.
  4. Rollback plan with actionable commands ("DROP TABLE IF EXISTS x;", "git revert <sha>") — "revert the change" is not a plan.
  5. Critical never auto-approves. Anything Critical (prod data loss potential, credential changes, irreversible deletes) escalates to the user regardless of checks.

Verdict line: GUARDRAIL: {APPROVED|REJECTED} | Risk: {level} | Checks: {pass/fail/skip} | Rollback: {one line}.

How to use

What it does

Derives a risk level for a proposed action from an Impact × Reversibility matrix, runs the stack's deterministic checks, and emits an unambiguous APPROVED/REJECTED call with a concrete rollback plan — escalating anything Critical to the user instead of auto-approving.

When to use it

Before migrations, bulk deletes or renames, deploy/infra config changes, force operations, or anything an orchestrator flags as hard to reverse — especially inside autonomous runs where no human watches each step.

How to invoke

@core:tech-lead loads it via skills: and applies it before risky delegations; load it manually (/guardrails context or read resources/risk-checks.md) when about to run a risky command yourself.

Worked example

Proposed: migration adding a flight_logs table. Additive CREATE TABLE → Impact Low, Reversible Yes (DROP TABLE) → Risk Low; typecheck/lint/tests pass. GUARDRAIL: APPROVED | Risk: Low | Checks: 3 pass | Rollback: DROP TABLE IF EXISTS flight_logs;. The same flow on DROP COLUMN with data present would come back High and REJECTED pending an expand/contract plan.

04 同仓库更多 Skills