用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/mitchellfyi/dex --skill dxloop命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
Orchestrate the full Dex ticket lifecycle from planning through PR completion.
Run independent full-scope Dex review waves until the selected risk tier's consecutive clean-pass gate succeeds.
Execute the approved implementation plan with TDD discipline and completeness verification.
正在显示 SKILL.md
| name | dxloop |
| description | Run a prompt in a loop until the requested work is fully implemented and verified. |
Use the terminal dxloop command for an audited plan-and-implementation loop.
The command launches a dedicated Claude session and binds every completion
receipt to that launch's session, purpose, phase, and generation.
Do not activate loop files from the current Claude session. A skill command does not receive the Claude hook session ID, so a file-only activation cannot prove that the current session owns it. Another Claude session in the same checkout could otherwise claim the loop before this session's first Stop hook.
Give the user a runnable terminal command containing their original task:
dxloop '<prompt>'
Quote the prompt safely for their shell. State plainly that this skill has not started a loop; the terminal command starts it. Then stop normally.
The dedicated wrapper owns activation, saves the original prompt for compaction recovery, launches planning, rotates to a fresh implementation generation, and cleans up on completion, timeout, interruption, or direct human pause. Do not reproduce those state transitions in this skill.
If the surrounding prompt says this is already a dxloop planning or
implementation session, follow that prompt instead of handing off again:
Never discover a current completion generation at signal time. Never write a
bare .complete, readiness, pause, or human-control marker yourself.