Dispatch

The idea

Managing a backlog of GitHub Issues or Azure DevOps work items is tedious — especially when you have AI coding agents that could handle them automatically if only they knew what to do. Dispatch closes that gap.

Point it at your issue tracker and it reads through your open work items, thinks through each one, and hands them off to an AI agent with a clear brief. The agent does the work, opens a pull request linked back to the original issue, and moves on to the next item. You review PRs instead of managing context.

What it does

Dispatch pulls work items from GitHub Issues, Azure DevOps Work Items, or a folder of local markdown files. Before handing off to an agent, it runs a spec generation step — an AI pass that explores your codebase and produces a structured task list scoped to the actual change needed. That spec becomes the brief the coding agent works from, which dramatically improves the quality of the output compared to forwarding the raw issue text.

Each task runs in an isolated agent session. You can run them in parallel to chew through a backlog quickly, chain them serially if order matters, or keep them fully isolated from each other so one failure doesn't cascade. When a session finishes, Dispatch handles the branch name, writes a conventional commit message, pushes, and opens a pull request that auto-closes the originating issue.

There's also a two-phase planner-executor pipeline for more complex work — one agent pass to plan, a second to execute — and a fix-tests mode that runs your test suite and feeds failures back to the agent until they pass.

Supported agents

Dispatch works with OpenCode (the default), GitHub Copilot, Claude Code, and OpenAI Codex. OpenCode is the most battle-tested path; Claude Code and Codex support is newer and may have rough edges.

Installation

npm install -g @pruddiman/dispatch

License

MIT