What is openwar?

OpenWar is a set of rules, and an optional companion program, designed to change how an AI agent behaves rather than how smart it is. The problem it tries to fix is that AI agents often jump straight into a task, skip asking clarifying questions, and try risky actions without pausing to check first. OpenWar replaces those habits with the pattern of a careful, experienced colleague who confirms the plan before acting. The simplest way to use it is a single markdown file that gets pasted into any agent's system prompt, taking effect right away with no installation needed. For stronger enforcement, there is also a runtime, written in TypeScript, that watches what the agent outputs at each step and catches the moments when it tries to skip a required rule. Work happens in phases. Before anything starts, the agent has to write a confirmation summary covering the objective, what it will deliver, constraints, tools it needs, and anything unclear, and execution cannot begin until a person accepts that summary. While working, the agent proceeds step by step, and if it plans to do something destructive or outside what was agreed, it has to stop and explicitly ask for permission first. If the agent gets stuck, the session pauses and saves its state so the work can pick back up later. At the end, it produces a short completion report. OpenWar is meant to sit on top of tools people already use, such as Claude Code, Codex CLI, Gemini CLI, aider, or a cloud AI service accessed with your own API key, rather than replacing any of them. It does not make the underlying model smarter, it just adds guardrails around how that model is allowed to behave. It can be tried immediately with no API key at all, either as a plain system prompt or by validating a task brief with no model call involved.