AgentSkillsCN

follow-test-runbook

按照项目测试手册的步骤逐一执行远程测试（通过 SSH 远程连接），并在不修改代码的前提下报告测试结果。

SKILL.md
--- frontmatter
name: follow-test-runbook
description: Execute a project's test runbook step-by-step (remote via ssh), reporting results without making code changes.
argument-hint: "[path/to/runbook.md]"
disable-model-invocation: false
user-invocable: true
allowed-tools: Read, Grep, Glob, Bash

Purpose

Follow a project's testing runbook and execute its commands safely.

This skill is designed to offload repetitive testing steps (often remote via ssh) to a cheaper model while keeping the main session focused on design and implementation.

Runbook discovery

Find the runbook in this priority order:

  1. •If invoked with an argument, treat it as the runbook path.
  2. •Otherwise, read CLAUDE.md in the project root and look for a line: Test runbook: <path>
  3. •If still unclear, ask the user for the runbook path.

What counts as an executable step

Only execute commands found in fenced code blocks.

  • •Prefer blocks fenced as bash (or sh).
  • •Do not execute prose steps.
  • •Do not execute non-shell code blocks.

If a runbook includes "recommended" commands that require adaptation (placeholders like <host>, $VAR, ..., or comments like "edit as needed"), ask a clarifying question before running them.

Safety rules (non-negotiable)

  • •Read-only posture: do not edit project files as part of this skill.
  • •Do not run destructive commands that could render the host/VM unusable. If a runbook command looks destructive or high-risk, stop and ask.
  • •Only run commands explicitly present in executable fenced blocks.
  • •If a runbook requires secrets, ask for them; do not guess or fabricate.

Examples of commands to treat as destructive/high-risk:

  • •disk/filesystem: dd, mkfs, newfs, fsck -y, disklabel, gpt destroy, zpool destroy, wipefs
  • •system power: shutdown, reboot, halt, poweroff, init 0, init 6
  • •mass deletion: rm -rf /, or rm -rf with ambiguous paths

Execution workflow

  1. •Identify the runbook and summarize what you will run.
  2. •Extract executable command blocks in order.
  3. •Execute commands step-by-step.
  4. •Stop at the first failure unless the runbook explicitly labels a step as non-fatal.
  5. •Report results in a detailed per-step format.

Output format

For each executed step, report:

  • •Step name (if present) and where it came from (runbook section heading).
  • •The exact command executed.
  • •Exit code.
  • •Key output (error lines and a small amount of surrounding context).

At the end, provide:

  • •Overall status: pass/fail
  • •First failing step (if any)
  • •Recommended next diagnostic command(s) (only if safe)