howdeploy/Z.A.E.B.A.L - 仓库详情

AI 仓库情报摘要

FR-AI / ANALYSIS

为什么值得关注

它利用用户的沮丧信号来解决编程代理因未修正假设而重复错误的常见故障模式，具有显式恢复和故障开放安全设计，并且不锁定特定工具，能与多种代理平台集成。

适合谁使用

使用 Claude Code、Codex CLI、Kimi CLI 或 OpenCode 的开发者
在生产环境中部署 AI 编程代理的团队
研究 AI 代理安全性和故障恢复的研究人员
寻求内置审计机制的编程代理工具开发者

典型使用场景

当用户表达不满时自动停止代理重复执行相同的错误操作
将多次失败升级至独立审计员进行深入分析
通过强制重新检查假设并要求用户显式确认来从代理死胡同中恢复
在多个支持的代理主机中集成一致的自审计行为

项目优势

零外部依赖，仅使用 Python 标准库
多语言（俄语、英语、中文）冒犯性语言检测，并带有意图分类
故障开放设计确保格式错误或审计员故障不会中断主机会话
支持四种主流代理主机，每个主机有专用适配器，共享同一核心

使用前须知

检测为启发式，可能因讽刺或非常规语境产生误报或漏报
状态锁定使用 POSIX fcntl，在 Windows 系统上可能丢失更新
在配置级别上的外部审计是同步的，用户必须等待审计员响应或超时

README 快速开始

Zaebal? · Audit · Errors · Break · Analyze · Leave no assumption

Read this in other languages 🇺🇸 English · 🇷🇺 Русский · 🇨🇳 简体中文

Profanity-triggered self-audit for coding agents. Z.A.E.B.A.L. treats user frustration as an operational signal: stop, re-check the agent's assumptions, and escalate repeated failures to an independent auditor.

Capabilities · How it works · Install · Configuration · Architecture

Why it exists

When a coding agent gets stuck, it often repeats the same action with small variations because one underlying belief about the task or codebase is wrong. The agent still treats that belief as a fact, so another self-check can reproduce the same mistake.

Z.A.E.B.A.L. adds a feedback loop to the user-message boundary:

profanity and direct complaints become an audit signal;
positive profanity such as “fucking great” does not add to the streak and closes an active incident as an acknowledgment;
repeated signals escalate from a local protocol to a full stop;
at level 3, an external agent reads the transcript and repository evidence;
work resumes only after an explicit user acknowledgment.

There is intentionally no technical tool lock. The protocol changes the agent's instructions and asks it to stop; the human always retains the final control.

Capability map

Capability	What it does	Implementation
Multilingual detection	Detects Russian, English, and Chinese profanity, including punctuation-separated and common leetspeak forms.	`core/wordlists/{ru,en,zh}.txt` + NFKC normalization
Intent classification	Separates praise, directed complaints, and ambiguous frustration before changing the streak.	`classify()`; weights `0`, `1.0`, and `0.5`
Session escalation	Tracks each session in a 30-minute sliding window and selects L1, L2, or L3.	Atomic JSON state + POSIX `fcntl` lock
Three audit protocols	Injects increasingly strict instructions: independent checks, assumption inventory, and full stop.	`core/protocol/L1.md` → `L3.md`
External auditor	Runs the same or a cross-vendor CLI against the transcript tail and repository evidence.	Claude, Codex, Kimi, or OpenCode
Four host adapters	Hooks Claude Code, Codex CLI, Kimi CLI, and OpenCode at user-message submission.	JSON hooks, TOML hook, or TypeScript plugin
Explicit recovery	Resets

项目描述

Profanity-triggered self-audit protocol for coding agents (Claude Code, Codex CLI, Kimi CLI, OpenCode). Python stdlib only.

Z.A.E.B.A.L