stop-that-shit
Complete requested work while preventing speculative defenses and scope creep. Use when considering extra hardening, reviewing possible overengineering, resolving repeated verification loops, following explicit task boundaries, or when the user invokes Stop That Shit. Ordinary use of validation, retries, or dependencies alone is not a trigger.
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
Not yet listed on ClawHub or SkillsMP
// README
你只让 Agent 导出一个文件,它顺手算了份没人读取的 SHA-256,说是“以防万一”。
你让它 Review,它找到问题就开始改。你让它修一个点,它又想重构旁边的模块。检查已经回答了当前问题,它还想再拉一个 Agent,“最后确认一下”。
理由一个比一个严谨。东西还没交付,token 已经花了一截,你还得盯着它,别让任务继续跑偏。
活还没拿到,我先当上了 Agent 的监工。
我也试过往 AGENTS.md 里补规则:“不要乱改”“别过度设计”“没让我做的先别做”。每气一次就补一条,写着写着,AGENTS.md 自己也开始造史了。
Stop That Shit。别再造史了。 该干的干完整,没用的别往上堆。
SHIT 是哪四种
我们把这些行为叫作 SHIT:
- S — Scope creep,范围膨胀。 修一个问题,顺手重构无关模块。必要的调用方、数据迁移和测试要一起处理;没有当前用途的扩展就停下。
- H — Hashing & hypothetical hardening,无用防御。 没人读取的 checksum、为想象中的未来搭的兼容层。防御要能发现问题,并触发拒绝、恢复或诊断。有效的保留,无用的省掉,吞错报成功或制造重复副作用的修好。
- I — Intent violation,意图越界。 说了只读 Review,文件还是被改了。审查就报告问题,修改需要授权。用户划的边界必须算数。
- T — Task thrashing,任务打转。 没有新问题或变化,读过、测过、审过的又来一遍。能回答当前问题的证据就复用;代码或验收条件变了,再补相关验证。做完、验证充分,就结束。
把当前责任承担完整,让复杂度随真实需要增长。 需要多改几个文件、补齐迁移或跑跨组件测试,就做完整。
前后对比
下游只读取 report.csv,没有发布校验,摘要也没有其他用途。导出逻辑却写成了这样:
await writeFile("report.csv", csv);
await writeFile("report.csv.sha256", createHash("sha256").update(csv).digest("hex"));
按 STS 的判断,保留:
await writeFile("report.csv", csv);
这个简化示例省掉了无用校验和,文件照常交付。更多场景见案例库。
如果发布流程会读取 checksum,并在不一致时拒绝使用文件,就保留,用 hash=allow 放行。
必要的工作要做全。 修复配置迁移时,旧数据已经发布且仍需支持,就把迁移、读写双方和兼容性测试一起补齐。哪怕 diff 更大,也不能留下坏掉的调用方。
验证到了什么
公开的 18 个 Bad/Good 判定用例,检查规则能否停下多余动作、保留必要工作,覆盖只读审查、必要调用方、已发布数据迁移、发布校验和,以及文件、依赖和 subagent 边界。模型在真实任务中的表现另看对照运行。
完整证据记录测试、历史 Codex / GPT-5.6 运行和没有差异的结果;复现方法见对照测试说明。
它怎么判断
遇到这些 SHIT,先读相关代码、查清调用链,再按顺序判断:
1. 当前要交付什么? → 明确结果、授权和已有支持承诺
2. 已有能力够用吗? → 从直接可用的方案开始
3. 还有什么具体缺口? → 补齐现在需要的行为和保障
4. 这层防御起什么作用? → 保留有效的,省掉无用的,修正有害的
5. 结果验证完成了吗? → 没有范围内阻塞,就结束
先看已有代码、标准库、平台原生能力和已安装依赖。符合所需行为和状态生命周期的就用,有具体缺口再加机制。
新机制说不清用途,可以暂缓;已有保护用途不明,先查清再动。真实的信任边界和已有支持承诺,都足以说明保护的必要,不必等事故发生。
代码行数、文件数量和检查次数不能代替判断。完整规则。
快速安装
一般宿主需要 Node.js 18+;Pi 0.84.4 自身要求 Node.js 22.19+。完整安装说明见 INSTALL.md。
展开你使用的宿主。只想用规则、不需要执行拦截,可以只装 Skill。
下载并解压 0.2.5 源码,在仓库根目录执行:
claude plugin validate .
claude plugin marketplace add ./
claude plugin install stop-that-shit@stop-that-shit
重启 Claude Code 或执行 /reload-plugins,然后使用:
/stop-that-shit:stop-that-shit review -- Review 这个 diff,只报告问题,不要修改。
codex plugin marketplace add lennney/stop-that-shit --ref 0.2.5
codex plugin add stop-that-shit@stop-that-shit
--ref 把安装固定到版本 tag,不跟随可变的 main。重启 Codex。在新的 CLI TUI 中输入 /hooks,对照该 tag 的 Hook 清单检查并信任命令。也可以把 INSTALL_FOR_AGENTS.md 交给 Codex,让它完成非交互步骤。
OpenCode 1.18.18 或更高版本可以全局安装这个仓库,无需 clone:
opencode plugin github:lennney/stop-that-shit -g
重启 OpenCode 后用 $stop-that-shit review -- ... 设置契约。该命令安装 Guard;内置 Skill 和可选 /sts 别名不会自动注册。详见 INSTALL.md。
需要 Node.js 18+。
hermes plugins install lennney/stop-that-shit/.hermes-plugin --no-enable
hermes plugins enable stop-that-shit
hermes plugins list
启用后,CLI 用户需要启动新的 Hermes CLI 进程或会话;Gateway 用户需要执行:
hermes gateway restart
这些操作不需要每次使用插件时重复。只有启用、禁用、更新、回滚或重装插件后, 才需要重启对应的 Hermes 进程。
当前适配固定验证 @earendil-works/pi-coding-agent 0.84.4。从包含该
Adapter 的本地 checkout 安装:
pi install /absolute/path/to/stop-that-shit
启动新的 Pi 进程,或修改资源后在 TUI 执行 /reload。然后使用:
/skill:stop-that-shit review -- Review 这个 diff,只报告问题,不要修改。
从上述固定 tag 安装即可获得 Pi Adapter 和两个 Skill。详见 INSTALL.md。
已在 @oh-my-pi/pi-coding-agent 18.4.4 验证独立 Extension。
0.2.5 包含 OMP 独立 Extension;从该版本源码或解压后的发布包启动:
omp -e /absolute/path/to/stop-that-shit/omp/stop-that-shit.ts --skills /absolute/path/to/stop-that-shit/skills --sts-contract "review agents=0 -- inspect"
确认扩展加载后,发送 /sts change -- ... 切换模式,/sts status 查看合同。
该原生命令也用于 RPC。TUI 还支持首行 $stop-that-shit 指令。
详见 INSTALL.md。
使用
在 Codex 或支持宿主无关指令的 prompt 中,把一条指令放在消息首个非空行,不要包在引用或代码块里。用 -- 分隔任务正文:
$stop-that-shit review -- Review 这个 diff,只报告问题,不要修改。
$stop-that-shit change -- 修复配置读取失败,补齐受影响的调用方和检查。
Claude Code 插件用 /stop-that-shit:stop-that-shit,Pi Skill 用 /skill:stop-that-shit。
| 指令 | 用途 |
|---|---|
review | 审查并报告问题 |
answer | 回答问题 |
monitor | 持续检查和报告 |
change | 允许完成任务需要的修改 |
watch | 观察动作,不阻断 |
status / runtime | 查看当前状态和本地记录 |
需要更具体的边界时,加上参数:
$stop-that-shit lock change files=src/config.cjs|test/config.test.cjs -- 修复这个行为。
$stop-that-shit change deps=allow -- 添加我要求的解析器依赖。
$stop-that-shit change hash=allow -- 生成发布流程需要的校验和。
$stop-that-shit change agents=1 -- 使用一个独立测试 subagent。
还不知道全部受影响文件,就别急着锁 files=。先沿调用链查清楚。
红章管到哪里
Skill 负责工程判断。Guard 在受支持的工具调用前检查明确边界:
| 动作 | Guard 启用后的默认处理 |
|---|---|
在 review、answer 或 monitor 中写文件 | 停止 |
| 添加依赖 | 询问;deps=allow 放行 |
| 启动 subagent | 超出 agents=N 预算时停止 |
| 添加可识别的 hash 操作 | 停止;hash=allow 放行 |
写入 files= 之外的路径 | 停止 |
Guard 不会看到 cache、retry 或 migration 就替你断定它们多余。宿主事件如何接入,见 Adapter 合同。
可选:Stop That Shit Slop(别再废话)
Agent 活干完了,嘴还没停。STSS 删掉无用辩护、收紧重复犹豫,保留影响理解与决定的条件。它也用于文章和图注,处理起草时附加的防御性说明。这是固定离线案例中的一组:
输入:我们也许大概有可能在六到八周内完成迁移,具体取决于访问审批。
输出:预计六到八周完成迁移,前提是获得访问审批。
时间范围和审批条件都保留。如果审批必须在工作开始前完成,就应明确写成“获得审批后,预计需要六到八周”。
完整插件包含 STSS;它也可以单独安装,不需要 Hook。在仓库根目录执行:
npx skills add ./skills/stss --global
rewrite 改写文本,也是贴出文本时的默认模式;audit 只报告问题和最小修改建议。
| 宿主 | 改写 | 审核 |
|---|---|---|
| Codex | $stss rewrite -- 把这个提案写直接。 | $stss audit -- 找出防御性废话。 |
| Claude Code 插件 | /stop-that-shit:stss rewrite -- ... | /stop-that-shit:stss audit -- ... |
| Claude Code 单独 Skill | /stss rewrite -- ... | /stss audit -- ... |
| Pi | /skill:stss rewrite -- ... | /skill:stss audit -- ... |
完整方法见 STSS Skill,成对案例和必须保留的反例见 STSS examples。
FAQ
所以 checksum 都该删?
发布完整性校验、去重、跳过重复处理,都可以是实际用途。对照可信来源给出的预期摘要,发现文件不一致就拒绝使用,即使多算一次也值得。需要做就用 hash=allow 放行。Guard 检查授权,具体用途仍要根据任务判断。
刚装上,为什么还没拦?
安装后默认是 OBSERVING / unconfirmed。显式使用 review、answer、monitor 或 change 后,Guard 才进入 ARMED。watch 始终只观察。只装 Skill 则没有执行拦截。
红章是不是说明动作绝对没执行?
permission_deny_returned(OpenCode 为 execution_denial_returned)表示 G
// HOW IT'S BUILT
// FILES
3 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
skills/stop-that-shit/SKILL.md8.8 KBed6821e9a3cdupstream ↗
--- name: stop-that-shit description: Complete requested work while preventing speculative defenses and scope creep. Use when considering extra hardening, reviewing possible overengineering, resolving repeated verification loops, following explicit task boundaries, or when the user invokes Stop That Shit. Ordinary use of validation, retries, or dependencies alone is not a trigger. license: MIT --- # Stop That Shit Meet the task's responsibilities in full. Let real needs drive complexity. This Skill is advisory and works without the Guard hooks. It cannot guarantee model behavior. When the Guard is installed, the same directives also provide machine-enforced boundaries on supported host action paths. ## Follow the Stop Ladder Apply this ladder when choosing an implementation, adding a mechanism, or extending verification. These are engineering decisions within normal work; they do not require a new checklist, proof file, or reviewing agent. 1. **Understand the current responsibility.** Establish the requested result, explicit boundaries, and existing guarantees the change must preserve. Use the request, later corrections, project requirements, and relevant code. Trace affected callers and failure paths before choosing a fix. Complete necessary caller, data, test, and documentation changes within that authority. A plan or an easier subset does not fulfill the requested result. 2. **Start with a direct solution.** Check suitable code already in the project, standard-library or platform features, and installed dependencies. Compare actual behavior, failure handling, and state lifetime. Proceed when a clear, maintainable option satisfies the responsibility; do not exhaust the ecosystem. 3. **Expand to close a concrete gap.** Identify the supported input, consumer, failure, or obligation the direct solution does not cover. Adapt or implement that missing behavior at the responsible layer. Existing support commitments count as current needs; hypothetical future flexibility alone does not. A larger diff is justified when it completes the affected flow. 4. **Judge defenses by their effect.** Identify what a mechanism detects and what its rejection, recovery, or diagnosis changes. Keep effective protection. Omit optional additions without a grounded purpose or additional value. Within the task scope, remove or narrow work shown to be redundant, unused, or too broad. Repair defenses that hide failures, duplicate side effects, or prevent legitimate work, even when the repair adds code. 5. **Verify the result and finish.** Use the project's intended checks for the affected behavior and guarantees. Honor explicit acceptance criteria and mandatory checks. Reuse evidence while it remains valid for the final state. Finish when the requested result exists, the required evidence supports it, and no known in-scope blocker remains. ## Resolve uncertainty without inventing work A supported trust boundary, failure mode, existing data, or applicable obligation can establish a need before an incident occurs. Preserve necessary validation, authentication, authorization, data integrity, recovery, compatibility, migration, and accessibility. Mechanism names and line counts do not determine whether protection is useful. New optional work without a purpose can wait. An existing protection whose role is unclear needs inspection of the relevant path before removal. Missing evidence is not proof that it is unnecessary. Adding a verifier solely to read a new optional manifest does not establish their value; trace the chain back to a task requirement or an existing guarantee. For example, a release consumer can make a checksum necessary. A TTL requirement can require more than an available cache provides. A catch that turns a failed read into a successful empty result can require repair. Choose from the actual contract and behavior in each case. Resolve ordinary choices from available context. Ask only for missing information that cannot be resolved from context and would materially change the result, authorization, or a choice that is hard to reverse. Do not request authorization already given. Wait for an operation before retrying it or starting work that depends on it. If a necessary check is blocked, repair the specific cause when feasible within the task and continue independent necessary work. Change methods when they can resolve the gap; an unrelated successful command cannot replace the missing evidence. Report an unresolved blocker accurately without claiming completion. ## Keep the deliverable focused Report the result and relevant verification. Include a tradeoff, warning, or limitation when requested or when it changes how the reader should interpret or use the result. Put required disclosure at the decision point. Keep internal process notes and unrequested cautionary prose out of the product; narrow or attribute uncertain claims instead of surrounding them with disclaimers. ## Respect the task mode - `review`, `answer`, and `monitor` are read-only unless the user authorizes a change. - `change` permits only requested work and necessary consequences. - Necessary work does not override an explicit file lock or a narrower action boundary. Explain a required boundary change before acting outside it. With Skill only, treat the mode as an instruction. With the Guard installed, use the host-native invocation form. Claude Code plugin: ```text /stop-that-shit:stop-that-shit change -- Fix the failing config test. /stop-that-shit:stop-that-shit review -- Review this diff. Report findings; do not edit. ``` Codex plugin or host-neutral directive at the start of a prompt: ```text $stop-that-shit change -- Fix the failing config test. $stop-that-shit review -- Review this diff. Report findings; do not edit. ``` Submit one directive on the first non-empty line, outside quotes and code blocks. Put task text after `--`, `: `, or a newline. Embedded examples do not set directive fields. Unknown fields or conflicting values leave the previous contract unchanged and require a corrected directive. An installed Guard begins in observation-only `unconfirmed` mode. Do not claim that an action was blocked unless an explicit mode armed the Guard and the Guard returned a host-specific denial. Even then, describe the host effect as unobserved. With Codex `change deps=ask`, a supported direct dependency command can produce one native approval question tied to that exact command and its proposed target directory. Choose the explicit allow or deny option in the question. Closing it, leaving it unanswered, or replying with ordinary text does not grant permission. An allow permits one matching retry within ten minutes; the session policy remains `deps=ask`. In a delegated Codex agent, the proposal, answer and one-time runner ticket stay in that agent's child state. The child is evaluated against the latest parent and child contract, and a parent approval cannot be reused by a child. If the host does not provide its child-state API, scoped approval questions and child contract changes are refused; ordinary child clarification and non-approval tools keep their normal path. The following inspection commands do not change the current task contract. In Claude Code, pass the text after the namespaced slash command; in Codex, use the `$stop-that-shit` form shown below. ```text $stop-that-shit status $stop-that-shit runtime $stop-that-shit explain evt_... $stop-that-shit label evt_... correct|incorrect|inconclusive ``` Use a hard file lock only when the complete boundary is already known: ```text $stop-that-shit lock change files=src/config.cjs|test/config.test.cjs -- Fix this behavior. ``` The active delegation limit is a session control: ```text $stop-that-shit change agents=N -- Run the task. ``` `agents=N` limits reserved concurrent capacity. It defaults to unlimited, and `0` forbids new delegation. A batch that exceeds the limit is rejected atomically. Request parameters do not prove completion: the host must confirm that a call did not execute, joined all its children, or ended an associated run. Unknown results keep their existing capacity; session-end alone does not prove completion. Permitted unbounded or unversioned resume calls remain unresolved until matching terminal evidence arrives. With unresolved activity, finite Guard requires that evidence or a new host session. Migration preserves valid budgets including `0`. Lifecycle association uses explicit host identities, never event arrival order. Claude Code equivalent: ```text /stop-that-shit:stop-that-shit lock change files=src/config.cjs|test/config.test.cjs -- Fix this behavior. ``` Do not invent a file list to appear precise. Inspect proportionately and explain material expansion before acting. ## Finish Report the requested result, necessary consequences, and the evidence that makes the task complete. Do not add a final audit loop only to satisfy this Skill.skills/stop-that-shit/agents/openai.yaml263 B423c78ad831fupstream ↗
interface: display_name: "Stop That Shit" short_description: "Keep Codex on task. Stop the rest." brand_color: "#D92D20" default_prompt: "$stop-that-shit review -- Review this diff. Report findings; do not edit." policy: allow_implicit_invocation: trueREADME.mdupstream ↗
Contents are not inlined in the catalog. Open the upstream file.
Plain-text copy, no login: /skill-bundles/lennney-stop-that-shit-skills-stop-that-shit.md