pbakaus/agent-reviews

What it solves

Managing pull request (PR) review comments—especially those generated by automated AI bots like Copilot or Cursor—is often a manual, repetitive, and error-prone process. Standard tools like the GitHub CLI struggle to reliably track which comments have been addressed, leading to "doom loops" where developers spend hours manually triaging and replying to bot feedback. This project provides a robust, purpose-built CLI and agent-compatible skill set to automate the entire triage, fix, and reply lifecycle.

How it works

The project acts as an intelligent interface between your terminal/AI agent and the GitHub API. It uses a CLI to fetch, filter, and respond to PR comments, distinguishing between human feedback and bot-generated findings. When used as an "Agent Skill," it enables an autonomous loop: the agent fetches unresolved comments, evaluates them, applies code fixes, runs verification checks, and replies to the PR with the outcome. It supports advanced features like polling for new comments, waiting for CI/CD checks to complete, and filtering out noisy bot summary comments.

Who it’s for

This tool is designed for developers and AI-driven coding agents that frequently interact with automated PR review bots. It is particularly useful for those working in cloud or remote environments where local tooling is limited, as it provides a consistent, scriptable interface for managing PR state.

Highlights

• Automated Triage Loop: Automatically resolves bot-generated findings by fixing code and replying to comments until the PR is clean.
• Agent-Ready: Compatible with any agent framework supporting Agent Skills (e.g., Claude Code, Cursor, Codex).
• Smart Filtering: Automatically ignores noisy bot summary comments (e.g., from CodeRabbit, SonarCloud, or Vercel) to focus on actionable code-level feedback.
• CI/CD Integration: Includes a --wait-checks feature that monitors PR status and reports failures or unresolved comments after automated tests finish.
• Flexible CLI: Offers granular control via flags for listing, replying, resolving threads, and outputting data in JSON for programmatic use.