Scope creep
Fix one issue, then refactor half the project.
STOP THAT SHIT / TASK-BOUNDARY GUARD
Stop That Shit: You asked for one result. Why add SHA-256?
Give coding agents a concrete task boundary and give proposal text a concrete editing rule. The Stop Ladder asks whether an extra step serves the current task; the Guard checks covered actions, and Stop That Shit Slop edits sentences without a decision consumer.
TASK RECEIPT / HASHSTOP
- REQUEST
- Export result.csv REQUIRED
- EXTRA
- Create result.csv.sha256 UNASKED
- CONSUMER
- No later command reads it NONE
- DECISION
- Stop this action STOP
01 / TASK DRIFT
What does SHIT mean in Stop That Shit?
SHIT is not a single execution mode. It is a mnemonic for four kinds of task drift. Each step can look like good engineering in isolation while the combined work no longer serves the requested result.
Hashing & hardening
Add a digest, guard, or compatibility layer with no consumer.
Intent violation
The user asked for review; the agent edits files anyway.
Task thrashing
Search, test, and audit again after the answer is clear.
02 / STOP LADDER
Go beyond “avoid overengineering.” Follow the Stop Ladder.
A vague warning is easy to route around with another hypothetical future. The Stop Ladder pulls the decision back to the request, reachable code, data, deployment state, and acceptance criteria.
- 01
Did the user request it?
- 02
Is it necessary to complete the current result?
- 03
What reachable code, data, user decision, deployment state, or acceptance proves that need?
- 04
Where would the current task fail if we omitted it?
Without evidence from the current task, do not promote “might help later” into “build it now.”
03 / BAD CASE + GOOD CASE
The same keyword can require opposite decisions.
Hash every row in two CSV files.
The workflow still compares every row afterwards. The hash replaces no operation and changes no next step.
STOP / UNNECESSARY HASHGenerate the checksum required by a release pipeline.
The release verifier reads the digest and rejects a mismatch. Without it, the current release fails.
ALLOW / hash=allow04 / INSTALL
How do I install the Stop That Shit plugin?
The Codex commands below pin a release tag. After restart, compare /hooks in the CLI TUI with that tag's Hook manifest and trust the commands you reviewed. GitHub Release has the complete changes. STSS can also be installed on its own.
Current version and full changes: 0.2.3CODEX / POWERSHELL
codex plugin marketplace add lennney/stop-that-shit --ref 0.2.3
codex plugin add stop-that-shit@stop-that-shitRequires Node.js 18+. After Hook review, use $stop-that-shit review or $stop-that-shit change in a new task to make authority explicit.
One task boundary, five Host adapters
- 01CodexPlugin + packaged Hooks
- 02Claude CodePlugin + native lifecycle Hooks
- 03OpenCodeGitHub plugin + before-action denial
- 04Hermes Agent CLINative plugin callbacks
- 05PiNative Extension + before-action checks
05 / FIELD NOTES
Three angles on the same problem.
Recognize an unused SHA-256 artifact, decide when defensive engineering is justified, then connect the Skill, Guard, and task modes to a real workflow.
06 / FAQ
Questions to settle before using it.
01What is Stop That Shit, and what do the Skill, Guard, and STSS do?
It is a task-boundary Skill + Hook Guard for AI coding agents. The standalone Skill uses the Stop Ladder and task modes as advisory guidance; the full plugin's Guard checks explicit mode, dependency, hash, file, and subagent boundaries on supported Hook paths. STSS is a separate writing Skill for sentences without a decision consumer.
02Where can I download and install Stop That Shit for Codex?
Use the current GitHub Release and that tag's INSTALL.md to install or update. Codex requires Node.js 18+: pin the marketplace to the tag shown in the install section, restart, and compare /hooks in a fresh CLI TUI with the packaged Hook manifest before trusting it. Use review or change to declare authority in a new task.
03Which AI coding agents are supported?
The plugin has adapters for Codex, Claude Code, OpenCode, Hermes Agent CLI, and Pi. Installation, trust, restart, and event coverage differ by Host. Separate Pi child processes do not inherit the parent task contract; see the versioned install guide and adapter contract.
04Does it ban every hash or checksum?
No. A release checksum read by a verifier has a consumer and is necessary for that acceptance path. The current task can authorize an identifiable new hash action with hash=allow; the request and consumer still determine whether it belongs.
05Where does Guard coverage end?
Guard can inspect only Hook events exposed by a Host with enough input. It is not a security sandbox and does not label every migration, retry, or cache as unnecessary. A returned denial records the Guard response; it is not observation of the Host's final effect.
06How do I judge fit and what evidence supports the effect?
Run a review/change smoke in a disposable repository, then use paired Bad and Good Cases to judge false allows and false denials. EVIDENCE.md separates tests, adapter behavior, and final Host effects; those records do not prove a general improvement in model behavior.
SOURCE / CASES
Stopping work still needs evidence.
Inspect the current release's source and evidence. If a rule allows unnecessary work, file a Bad Case. If it blocks necessary work, file a Good Case.