AgentSkillsCN

dry-run

在不修改文件的前提下，模拟命令的干运行执行过程。先设置 dry_run 标志，以只读模式执行命令，随后重置标志。此功能可用于安全测试工作流。触发关键词：干运行、模拟、测试命令、预览变更、安全模式、禁止写入。

SKILL.md
--- frontmatter
name: dry-run
description: Simulates command execution in dry-run mode without file modifications. Sets dry_run flag, executes command with read-only constraint, then resets flag. Useful for testing workflows safely. Triggers on keywords: dry run, simulate, test command, preview changes, safe mode, no write
project-agnostic: true
allowed-tools:
  - Bash
  - Write
  - Read

Dry Run Skill

Executes any command or skill in simulation mode, preventing all file modifications except session state.

Usage

code
/dry-run <any command or prompt>

Examples:

  • •/dry-run /po_spec path/to/spec.md - Preview full spec workflow
  • •/dry-run /spec IMPLEMENT path/to/spec.md - Test implementation without changes
  • •/dry-run why did you implement X this way? - Normal chat (no files affected)

Workflow

Step 1: Initialize Session Status

Find the Claude Code PID and create session directory:

bash
pgrep -x claude | head -1 || echo "shared"

Use the output as CLAUDE_PID. Then create the session directory:

bash
mkdir -p outputs/session/<CLAUDE_PID>

Check if outputs/session/<CLAUDE_PID>/status.yml exists. If not, create with initial schema:

yaml
dry_run: false

Step 2: Set Dry-Run Mode

Update outputs/session/<CLAUDE_PID>/status.yml:

yaml
dry_run: true

This signals to THIS session's file operations that writes are prohibited. Other Claude sessions (different PIDs) are not affected.

Step 3: Execute Delegated Command

Execute the provided command/prompt EXACTLY as given. All behavior remains normal EXCEPT:

CRITICAL CONSTRAINTS:

  • •NO file modifications allowed (Read, Grep, Glob, LSP, Bash read-only commands OK)
  • •ONLY exception: outputs/session/<CLAUDE_PID>/status.yml can be modified
  • •If command requires file writes, DESCRIBE what WOULD be changed instead
  • •For chat-only prompts (no file operations needed), respond normally

Step 4: Reset Dry-Run Mode

After execution completes (success or failure), reset state in outputs/session/<CLAUDE_PID>/status.yml:

yaml
dry_run: false

Step 5: Report Results

Provide summary:

  • •What was executed
  • •What file changes WOULD have occurred (if any)
  • •Verification that dry_run mode is reset

Implementation Notes

The skill acts as a wrapper:

  1. •It does NOT interpret or execute the delegated work itself
  2. •It sets the flag, then lets normal AI behavior handle the prompt
  3. •The dry_run flag is checked by pretooluse hook (hard enforcement at tool level)
  4. •After completion, it ensures cleanup

State Schema

outputs/session/<claude_pid>/status.yml:

yaml
dry_run: bool  # true = prevent file writes, false = normal mode

Session isolation by Claude PID ensures parallel agents don't interfere with each other. Future extensions may add additional session state fields.