菜单
Run the loop from `prove-the-failure` and confirm red goes green, then hold a cooldown before any further action is permitted. Use after `adversarial-review` returns PASS, before `immunise`.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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/CleanExpo/skills-library --skill verify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。
Dispatch the eight operational roles against a task, each grounded in the vault before it reads any code, with a live view of what is running. Use for work that wants several independent lenses at once — map, plan, build, verify, review, security, CI, release — rather than one agent doing everything.
Thin wrapper around the Pi-CEO Hour-1 provisioner. Given a client slug + email, invokes `swarm/inbox/provisioner.py` to spin up the full Linear project + Supabase portal_content + ContextBot Telegram intake bot + welcome email + Phill ping. Use when the user says "onboard {client}", "spin up portal for", "new client setup", or right after `stripe-milestone-invoice` confirms a deposit cleared.
PM-Core — the first autonomous execution agent. Claims the highest-priority agent-ready ticket from the Unite-Group Linear board, clones the repo, implements the fix on a feature branch, runs CI, and opens a PR. Model: claude-sonnet-5 (senior agent tier). Trigger: /pm-core [ticket-id] or autonomous via Hermes cron.
正在显示 SKILL.md
| name | verify |
| description | Run the loop from `prove-the-failure` and confirm red goes green, then hold a cooldown before any further action is permitted. Use after `adversarial-review` returns PASS, before `immunise`. |
Green is a measurement, not a belief. The only evidence that counts is the same loop that proved the failure, run again, now passing — and passing because of the fix, not alongside it.
Before trusting green, confirm the loop can still go red. Revert the fix, run the loop, watch it fail.
A loop that is green with and without the fix proves nothing. This is the positive control, and skipping it is how a fix that changed nothing gets marked verified.
No further action for the cooldown window. Not immunise, not another repair, not a related cleanup.
Default: 30 minutes, or one full cycle of the affected scheduled job, whichever is longer.
The cooldown exists because a fix that breaks something adjacent usually reveals it on the next scheduled run, not immediately. Chaining straight into the next action means the breakage lands inside the next repair and the two become indistinguishable.
During cooldown: watch. Do not act.
Paste both outputs. A claim of green without its output is not a verification, and this chain treats it as an unverified claim regardless of how confident the wording is.
If the loop is still red, the fix failed. Return to propose-fix, consuming one attempt from its bound of three. Do not adjust the loop to make it pass — modifying the loop to accommodate the fix inverts the entire chain.
If a new failure appears during cooldown, stop. Do not repair it inside this chain. Start a new chain at prove-the-failure, and record the adjacency — two failures in one window is a circuit-breaker signal.
immunise — stop it recurring.
