HN 표시: VS Code Agent Kanban: AI 지원 개발자를 위한 작업 관리

VS Code Agent Kanban is available on VS Code Marketplace | GitHub Note: VS Code Agent Kanban has had a major update since this post to add an optional Git worktree workflow (detailed here), automating the creation of worktrees from the task board or @kanban GitHub Copilot participant command. Updated documentation is available at github.com/appsoftwareltd/vscode-agent-kanban Main Features Agent Kanban has 4 main features: - GitOps & team friendly kanban board integration inside VS Code - Structured plan /todo /implement via@kanban commands - Leverages your existing agent harness rather than trying to bundle a built in one .md task format provides a permanent (editable) source of truth including considerations, decisions and actions, that is resistant to context rot The VS Code Agent Kanban Board, Markdown file and GitHub Copilot Chat: Youtube (Quick Demo) (Note that the image below shows the updated UI, the video needs an update but illustrates the workflow) Background There's a problem that anyone who regularly uses AI coding agents will recognise immediately: context rot. You start a conversation with Copilot about a feature. You go deep - exploring approaches, debating trade-offs, building a shared understanding. Then a week later you come back to it, or you hit the context limit and clear chat, and that entire history is gone. You're starting over. Or worse, the next session drifts because the agent has no memory of the decisions already made. Modern AI coding agents are powerful, but they operate without memory. Every chat session is a blank slate. This creates a real workflow tension: - Context bloat - long-running tasks accumulate enormous context windows, slowing the LLM and increasing the chance of the agent losing track of earlier constraints. - No persistent task history - decisions, plans, and rationale disappear when you clear chat or close VS Code. - No structure - tasks live only in your head or in separate todo apps, disconnected from the code and the agent workflow. - Team coordination - there's no shared view of what the AI is working on, what's been decided, or what's in flight. The traditional fix is to paste context back in manually, write notes elsewhere, or use a project management tool that lives entirely outside the IDE. None of these feel right for an agent-first workflow. Agent Kanban is a VS Code extension that solves these problems via a markdown formatted task record, and a clear plan / todo / implement flow. Markdown Files as the Source of Truth Agent Kanban takes a deliberately simple approach: every task is a Markdown file. Each task lives in a .agentkanban/tasks/ folder as a .md file with YAML frontmatter. The frontmatter tracks the task title, which Kanban lane it's in, and timestamps. The body of the file is a structured conversation log, using [user] and [agent] markers. --- title: Implement OAuth2 lane: doing created: 2026-03-08T10:00:00.000Z updated: 2026-03-08T14:30:00.000Z description: OAuth2 integration for the API --- ## Conversation [user] Let's plan the OAuth2 implementation. We need to support both device code and client credentials flows. (User types 'plan' in the chat window) [agent] Here's my analysis of the two OAuth2 approaches for your API... [user] Make changes here, here and here (User types 'plan' in the chat window) [agent] Here's my updated plan ... (User types 'todo' in the chat window, agent creates todos, User types 'implement' when ready for the agent to start work) This is intentionally boring and readable. No proprietary formats, no databases, no opaque state. Just text files you can open, edit, search, and commit. By having the user confirm readiness in GitHub copilot chat, the robust built in agent harness can start work, with all of the features and capabilities that come with it. This is a key design decision - early testing with a custom harness proved to have limitations while working within the context of a VS Code integrated extension. GitOps Friendly by Design The entire .agentkanban/ folder is designed to be committed to version control (if you choose to). This gives you a few things for free: Permanent task history. Every decision, plan, and implementation conversation is preserved in your repository. Future developers (and future you) can see not just what was built but why, in the actual words of the planning conversation. Diffable, mergeable state. One file per task, board config in YAML - these are standard text files. They diff and merge naturally. No merge conflicts from opaque binary state. No sync issues between team members. Shared team context. When you commit task files, your whole team gets visibility into what the AI is working on, what's been decided, and what the current state of any feature is. Pull the latest and open the board - you're immediately up to date. Audit trail. In regulated or enterprise environments, having a git-tracked record of the planning and decision-making process for AI-assisted development is inc