Dev.to AI 🤖 Ai 👁 0 📖 2 min read

cmc: stop interrupting Claude Code. Hold your messages, hand tasks to subagents

You give Claude Code a task. Thirty seconds later you remember something: "also add tests for the parser". Now you have two bad options: Press Enter, and your message lands in the middle of the agent's work. Wait, an

cmc: stop interrupting Claude Code. Hold your messages, hand tasks to subagents

You give Claude Code a task. Thirty seconds later you remember something: "also add tests for the parser". Now you have two bad options:

  • Press Enter, and your message lands in the middle of the agent's work.
  • Wait, and hope you still remember it when the turn ends.

I wanted a third option, so I built cmc — Claude Message Control: a Claude Code plugin that holds what you type while the agent works, and hands it over when the turn ends, without interrupting it.

Hold, don't interrupt

Prefix a message with ;; while the agent is working:

;; also add tests for the date parser
;; then update the changelog

Each message is held. You get a one-line notice, the prompt box is cleared, and the agent keeps going. When its turn ends, cmc delivers everything you held, in order, as one message from you:

Held messages delivered when the turn ends

No command to remember: the messages simply arrive when the agent is free. Before then you can still change your mind. /held lets you send, edit, reorder or delete a held message, and /send sends it right away.

The cmc hub

/cmc opens a panel beside the chat. It uses Claude Code's own plugin engine, so it works with the mouse:

The cmc hub with two held messages while the agent works

From there you can:

  • See the agent's state and every held message.
  • Send now a held message. If the agent is busy, cmc puts the text in the prompt box and Ctrl+Enter sends it into the running turn.
  • Edit, reorder or delete a message, or hand it to a subagent.
  • Read Help, a short page of docs inside the hub that explains everything.
  • Switch modes under Settings: deliver at the end of the turn or only on demand, send ;; messages to the main agent or to a subagent, and pick the subagent model and effort.

Subagents in their own worktree

Start a subagent with >>:

>> -m haiku -e low write a two-line haiku about message queues

The flags go first: -m model, -e effort, -w the whole conversation as context, -s a short summary instead, -a its result comes back by itself.

The subagent runs as its own Claude Code process:

  • Context: by default it gets the whole conversation, or a short summary if you prefer.
  • Isolation: it works in its own git worktree, so its edits never touch your working tree.
  • Model: it uses the model and effort you choose, or inherits yours.

Its conversation shows up in the hub as a chat, with your messages on one side and its replies on the other:

A subagent's conversation in the hub

You can keep writing to it, stop it, or send its result to the main agent with one click.

Install

You need Claude Code 2.1.330+ and Node.js 22.18+.

/plugin marketplace add petersonsenadevs/cmc
/plugin install cmc@r3v3rs3

The plugin is TypeScript that Node runs directly: there is no build step and there are no runtime dependencies. It runs on Windows, macOS and Linux. Subagents get acceptEdits at most and never bypassPermissions.

Why

The agent is fast, but my thoughts don't arrive on its schedule. cmc keeps them in a queue I control, and I decide what reaches the agent and when.

It's MIT-licensed and open source. Issues, ideas and PRs are welcome on cmc.

— r3v3rs3

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.