Case file
Stop That Shit
Keep AI coding agents inside the current task boundary
Problem
Coding agents can present unused SHA-256 digests, compatibility layers, full-suite tests, and extra process as caution while a small task keeps expanding.
Product judgement
Do not judge overengineering by line count or keywords. Check user authority, current consumers, reachable evidence, and whether omission would fail current acceptance.
Prototype and system
The Skill applies the four-question Stop Ladder, while the Guard checks review/change, dependency, hash, file, and subagent boundaries on covered Host action paths.
Current evidence
The public 0.1.0 release provides adapters for Codex, Claude Code, OpenCode, and Hermes Agent CLI, with paired Bad/Good Cases that preserve necessary work.
Next step
Keep turning sanitized reports into paired Bad/Good Cases that change one decisive fact, while evaluating routing, Hook decisions, and Host effects separately.
Related writing