Get it for macOS

Runs BB threads on any of Command Code's 70 coding models, so a bug has to survive reviewers on Claude, GPT, GLM, Kimi or DeepSeek without leaving BB. Shell commands, file reads and edits arrive as BB timeline cards with streamed output and diffs, and reasoning stays separate from the reply. Needs Command Code installed and signed in, plus the adapter this repository ships. Command Code's headless mode cannot prompt per tool call, so writes are one setting rather than a per-call approval.


Overview

View source

Adds Command Code to BB's provider picker, so threads can run on any of the 70 models it exposes — Claude, GPT, GLM, Kimi, DeepSeek, MiniMax and the rest — without leaving BB.

What you get

Command Code is a terminal coding agent that learns your code style. It has no Agent Client Protocol server of its own, so this plugin ships one: a small adapter that drives cmd in headless mode and translates its event stream into ACP. BB's own ACP bridge does the rest.

The translation is what makes threads read like every other BB thread:

  • Tool calls become BB timeline cards. A shell command renders as a command execution with its output streaming in. A file read links the file it touched. A write carries a diff.
  • Reasoning is separated from the answer. Command Code's thinking stream lands in BB's reasoning item, not inline with the reply.
  • Every model is selectable. The adapter reformats cmd --list-models into the shape BB's picker reads, with Command Code's own default model first.
  • Context survives the turn. cmd -p is one-shot, so each prompt resumes the same Command Code session underneath. A restart does not lose the thread.

What it costs you

Three installs, all public:

npm i -g command-code                                # the agent
npm i -g github:aqidd/bb-plugin-command-code         # the ACP adapter
bb plugin install git:https://github.com/aqidd/bb-plugin-command-code

Then cmd login. Command Code accounts are separate from BB; bring-your-own-key also works through cmd's own config.

The provider stays hidden in BB's picker until command-code-acp is on the machine's PATH, so a partial install shows nothing rather than failing at run time.

The permission tradeoff

Command Code's headless mode has no interactive permission channel. Verified against 1.53.0: without --yolo every write and shell tool is refused outright, and with it they all run unsupervised. There is no middle setting to map BB's per-tool approval onto.

So this plugin does not pretend to prompt per call. It exposes the choice once, in Settings → Plugins → Command Code:

  • Allow file writes and shell commands: on — cmd runs with --yolo.
  • off — the provider can read and search, nothing else.

If you want Command Code for review and exploration but not for edits, turn it off. Per-call approval through Command Code's PreToolUse hook is tracked in the repository issues.

Known gaps

  • No per-tool permission prompts (above).
  • BB's context meter shows how full the model's context is; token and cost totals are not reported.
  • No thread fork or edit-past-message: Command Code cannot clone a session.
  • Images, including image URLs, reach vision-capable models as a file path the model opens itself, costing one extra tool call.

MIT licensed. Issues and pull requests at aqidd/bb-plugin-command-code.

More in Agents & Providers 40

Add, choose, configure, route, or coordinate who runs a thread.

View all