A bouncer for AI agents that generate SHA-256s nobody asked for

What it does

Stop That Shit is a cross-platform guard layer for AI coding agents. It sits between tools like Codex, Claude Code, OpenCode, and Hermes Agent CLI and your codebase, intercepting actions that violate the task you actually assigned—like writing files during a review-only request, adding dependencies you never mentioned, or generating checksums with no downstream consumer. When it catches an overstep, it returns a blunt red-stamp denial instead of letting the agent proceed.

The interesting bit

The project treats agent over-engineering as a supply-chain problem: it uses a shared Skill + Hook + Adapter architecture so the same boundary rules apply across four different CLI hosts. It even has an acronym for the misbehavior—SHIT (Scope creep, Hashing, Intent violation, Task thrashing)—and judges necessity by asking whether any reachable code or deployment state actually requires the action, not whether the action looks plausible.

Key highlights

• Supports Codex, Claude Code, OpenCode, and Hermes Agent CLI through a unified adapter layer.
• Blocks specific covered actions by default: file writes in review mode, unrecognized hash operations, out-of-bounds file edits, unapproved dependency adds, and excessive subagent spawning.
• Uses explicit task modes (review, change, answer, monitor) to arm the guard; defaults to passive OBSERVING so it does not guess intent.
• Local-first runtime stores only metadata, distinguishing checked actions, denials, and unobserved host effects.
• Includes a paired eval and evidence log (including GPT-5.6 pilot runs) rather than claiming perfection without proof.

Caveats

• The guard only covers supported Hook paths; hosts must provide before-action events, and Stop That Shit explicitly treats the final execution outcome as unobserved.
• It blocks recognizable hash operations and file writes by default, but will not reject actions simply because they contain words like cache or migration—it needs explicit evidence of necessity or user authorization.
• Semantic judgment relies on a four-question Skill heuristic; when evidence is insufficient, the agent is advised to report or pause rather than proceed.

Verdict

Worth trying if you are tired of AI agents adding SHA-256 files, guard clauses, and bonus refactors to every task. If your agent workflow is already tightly scripted and rarely drifts, this is probably overkill for you.