LB623/no-negative-echo

What it solves

It prevents AI agents from including "negative echoes" in their final deliverables. This happens when an agent corrects a plan during a conversation but accidentally includes references to the rejected ideas in the final output—for example, naming a dish "Tomato and Egg (without pork)" instead of simply "Tomato and Egg."

How it works

The project provides a set of guidelines and a technical "Skill" (following the Agent Skills specification) that forces the agent to regenerate delivery text based solely on the final, verified state of the work. It uses a three-question filter for every output surface (titles, commit messages, PR descriptions, etc.) to determine if a piece of information is actually necessary for a reader who hasn't seen the conversation history.

Users can implement this in two ways:

1. Installing the Skill: A formal installation that allows the agent to use specific scripts (like check_surface.py) for high-assurance checks before final delivery.
2. AGENTS.md: Adding a concise set of core behavioral rules to a project's AGENTS.md file to ensure the agent consistently follows these constraints throughout the session.

Who it’s for

Developers and users working with AI agents (such as Codex) who require professional, clean final deliverables like commit messages, pull request descriptions, and documentation without conversational residue.

Highlights

• Dual implementation: Supports both a formal installable Skill and a lightweight AGENTS.md configuration.
• Surface-specific auditing: Evaluates different delivery surfaces (filenames, metadata, hooks) independently.
• Strict filtering logic: Distinguishes between rejected conversational attempts (to be removed) and actual baseline changes or security/compatibility facts (to be kept).
• Verification tools: Includes a scanner for text and filenames and a testing suite with evaluation protocols.