agent-autoloop: Claude Code and Codex as a background automation battalion

This is an automation battalion that allows CLI agents like Claude Code or Codex to work autonomously in the background without being tied to the terminal. The agent continues working even when the user is away, sending notifications only when a decision is needed to hand over control. It runs on Node.js alone in macOS and Linux environments, requiring no external cloud servers or public ports.

The core is a zero-dependency kernel consisting of approximately 1,650 lines of code. This kernel separates event sources, agents, notification channels, and CI/CD pipelines into plugins, allowing them to be freely combined. When an agent is in a waiting state, the process is terminated to eliminate memory and token consumption, and when the user sends a reply, it resumes the task by continuing the previous conversation context.

Unlike existing methods, no modification of agent source code or SDK integration is required. Any CLI tool that takes a prompt as input and outputs a final response can be connected with a one-line configuration. It supports various event sources such as GitHub issues, cron jobs, and HTTP webhooks, and sends cards via Slack, Discord, Feishu, DingTalk, or desktop notifications when a decision is required.

It provides templates for various scenarios, from personal project management to automatic GitHub issue responses, PR reviews, and deployments. In particular, PRs tagged with 'deliver' are automatically tracked through the entire process, from code review and CI waiting to merge and deployment verification. For security, text from external events is treated only as data, separating instructions from data to block injection attacks.