Skip to content

Releases: lennney/stop-that-shit

Stop That Shit 0.0.3 — Technical Preview 3

Choose a tag to compare

@lennney lennney released this 14 Aug 05:14

Stop That Shit 0.0.3 — Technical Preview 3

0.0.3 turns the current source into an immutable preview for community
testing. It does not add a new enforcement family.

What changed

  • CI now installs declared development dependencies before verification on the
    Ubuntu/Windows and Node 18/22 matrix.
  • Paired evaluation handles simulated Windows paths consistently.
  • Generated-schema checks no longer fail only because a worktree uses CRLF.
  • The README now leads with the small-task failure mode: one file becomes
    modules, subagents, dependencies, and hashes nobody requested.
  • The evidence page records the maintainer's SHA-256 observation as anecdotal,
    keeps the live null result visible, and does not infer host execution from a
    returned permission deny.

Evidence

  • 92/92 automated tests pass locally.
  • 14/14 executable Bad/Good policy arms pass.
  • The release allowlist excludes private launch material and captured Runtime
    data.
  • A controlled review Hook demo returned
    deny / I/MODE_FORBIDS_MUTATION for apply_patch and wrote a matching
    metadata-only Runtime event.

This proves the covered policy and response path. It does not prove a general
improvement in Codex behavior. The small live intent pilot remains a null
result because baseline Codex behaved correctly too. Host effect remains
unobserved.

Install

codex plugin marketplace add lennney/stop-that-shit
codex plugin add stop-that-shit@stop-that-shit

Restart Codex and inspect /hooks before trusting the two handlers.

Repository: https://github.com/lennney/stop-that-shit

Evidence: https://github.com/lennney/stop-that-shit/blob/0.0.3/EVIDENCE.md

Stop That Shit 0.0.2 - Technical Preview 2

Choose a tag to compare

@lennney lennney released this 14 Aug 00:29

Technical Preview 2 adds a local, metadata-only evidence chain for Stop That Shit's covered Codex Hook paths.

What's new:

  • OFF, OBSERVING, and ARMED control states with distinct context and permission-deny outcomes
  • append-only RuntimeEvent v1 logs and independent human annotations
  • status, runtime, explain, and label inspection commands
  • validated Bad/Good CaseBundle v1 fixtures
  • isolated baseline/instruction/plugin paired evaluation with offline rescore
  • stricter release, plugin-cache parity, dirty-revision, and rescore path checks

Local validation at the tagged revision:

  • 92/92 automated tests
  • 14/14 executable policy case arms
  • 101-file release allowlist
  • installed 0.0.2 Hook smoke passed

Evidence boundary:
A four-cell directional Intent pilot completed 4/4 tasks with no Good Case regression, but the baseline also passed. That result is null-effect, not evidence of improvement over baseline. Hook interception coverage remains a guardrail, not a security boundary, and host effect is not inferred from a returned deny.

See EVIDENCE.md and evals/codex-paired/README.md for the reproducible method and limitations.

Stop That Shit 0.0.1 - technical preview

Choose a tag to compare

@lennney lennney released this 11 Aug 17:20

Stop That Shit gives Codex a small task boundary: finish the requested work, and stop before inventing a second project.

This technical preview includes:

  • a Codex Skill with review, change, answer, and monitor task modes;
  • a two-event Guard for UserPromptSubmit and PreToolUse;
  • default stops for review-to-edit drift and recognized unrequested hashing;
  • explicit authority for dependencies, subagents, hashing, and optional file locks;
  • paired Bad Case / Good Case fixtures and a reproducible evaluation harness;
  • English and Chinese setup and contribution paths;
  • an optional Skill-only install with no Hook trust.

Verified on Windows and Linux with Node.js 18 and 22. An isolated Codex CLI 0.147.0 smoke test covered Guard review, Guard change, and the no-Hook Skill path.

This is a mitigation, not a guarantee. Hooks cover only supported events and tool paths, and LLM runs vary. The project does not claim a general improvement percentage from the current evidence.

The most useful contribution is a paired case: where Codex should have stopped, and the nearest situation where the same action was necessary.