vibe-kanban_skill

Overview

This skill integrates multiple AI coding agents into a visual Kanban board and automates the full task lifecycle from sandboxed worktrees to pull requests. It embeds a Conductor pattern to run agent tasks in parallel using git worktree sandboxes and supports planview integration for task review and approval. The UI and CLI modes let teams choose visual tracking or scripted automation.

How this skill works

When a task card moves to In Progress, the system creates a dedicated git worktree branch, invokes the selected AI agent (Claude, Codex, Gemini, etc.) in that sandbox, and streams logs to the UI. Multiple cards produce parallel worktrees and concurrent agent runs (Conductor pattern). After review, diffs can be inspected, agents retried or swapped, and approved changes are pushed as PRs to GitHub.

When to use it

• Manage multiple AI coding agents from a single team Kanban board with visual progress tracking.
• Break an epic into independent tasks and run them in parallel across isolated git worktrees.
• Connect To Do → In Progress → Review → Done with your GitHub PR workflow.
• Use planview to review and approve task specs before agents execute work.
• Run headless CI/CD or scripted pipelines via CLI for automation and reproducibility.

Best practices

• Define clear task cards: inputs, outputs, constraints, and acceptance criteria.
• Assign the most suitable agent per task (design tasks to Claude, code to Codex, docs to Gemini).
• Use planview approval to gate agent execution for high-risk or complex changes.
• Keep worktrees lightweight and clean up orphaned worktrees unless debugging.
• Review agent logs and diffs in Review column before creating PRs; prefer retries over blind merges.

Example use cases

• Parallelize an epic: split frontend, backend, and tests into separate cards and run agents concurrently.
• Automate a CI pipeline: run scripts/pipeline.sh to perform check → conductor → PR without UI.
• Experiment locally: npx vibe-kanban for quick local UI with automatic browser launch.
• Agent swap during review: replace an unsatisfactory agent run with a different agent and re-execute.
• Use planview to annotate diffs and request changes, then re-run agents until approved.