---
title: "我總是搞不清我的程式設計代理實際上做了什麼，所以我打造了 RoadR"
url: "https://claudemods.dev/zh-tw/builds/i-kept-losing-track-of-what-my-coding-agents-had-actually-do-3f01de"
source: "https://x.com/i/status/2108953165303353795"
source_type: "x"
author: "Mateusz Woźniak"
author_handle: "Shuffzord"
stars: 0
category: "多代理與協作"
categories: ["agents"]
tags: ["agent-tracking", "mcp", "planning", "roadmap", "skills"]
permission_level: null
scan_complete: false
added: "2026-10-16T16:48:24.941Z"
updated: "2026-10-10T16:09:54.000Z"
locale: "zh-tw"
content_locale: "zh-tw"
---

# 我總是搞不清我的程式設計代理實際上做了什麼，所以我打造了 RoadR

我總是搞不清我的程式設計代理實際上做了什麼，所以我打造了 RoadRaven：一個本機路線圖樹，Claude Code 在工作時會更新它。計畫保持真實，而不是過時。

隨附一個 MCP 伺服器、技能和一個 Claude Code 外掛。

## Description

https://t.co/HSluYZuJYC https://t.co/MeB8qdp01Q

## Original README

I kept losing track of what my coding agents had actually done, so I built RoadRaven: a local roadmap tree that Claude Code updates while it works. The plan stays true instead of going stale.

Ships with an MCP server, skills and a Claude Code mod.

https://github.com/Shuffzord/RoadRaven

The plan is a plain roadmap.json in your repo. Agents read and write it through an MCP server or a local WebSocket, and nodes change colour as the work happens. No cloud, no accounts.

Two Claude Code skills come with it: one plans a request as phases and runs each with a sub-agent, re-running your tests before anything counts as done. The other is how a single agent claims a task and reports back.

New: a Claude Code mod, a pane inside the terminal. It shows which agent is on which node and the acceptance checks waiting for me. I pass or fail a batch and the orchestrating agent picks it up.

Windows and Linux, alpha, source-available. In Claude Code:

/plugin marketplace add Shuffzord/RoadRaven
/plugin install roadraven@roadraven