Shunt Muse — Development skill for Claude Code

shunt-muse

Keep big files out of your coding agent's context. When Claude Code or Codex tries to read a large file in full, a hook stops it and points it at a skill that hands the files to a cheaper model (Muse Spark via `muse exec`). Only the short answer comes back into the session.

This is a fork of Spotify's [shunt](https://github.com/spotify/portal-ai-plugins/tree/main/plugins/shunt). The original delegates to AiKA modes on a Spotify Portal instance. If you don't have Portal, this version does the same thing with the Muse CLI you already have installed.

What it does

• Hooks (PreToolUse) block a full Read, or a bare cat/head/tail/less/more, of any file over 350 lines. Reads with an offset or limit, piped commands and small files go through untouched.
• bulk-read sends one or more files plus a question to Muse and prints the answer.
• code-write sends a spec plus a reference file to Muse and writes the generated file (tests, config, stubs).
• Skills (bulk-reader, code-writer) tell the agent when to call the scripts.

Every call is one shot. A follow-up question sends the files again, which costs nothing in your agent's context because they never enter it.

Numbers

One run on a 617-line test file: about 4,800 tokens went to Muse and roughly 300 came back, in 24 seconds at `low` effort. At the default `high`, a 1,055-line file took about 20 seconds and every cited line number was exact (`xhigh` gave the same answer in about 32 seconds). Spotify reports 82-94% savings with the original plugin on a 162K-line Java monorepo; I haven't reproduced that here.

Requirements

• Muse Code (muse) with muse exec working headless
• jq and perl (both ship with macOS)
• Claude Code with plugin support, and/or Codex CLI with hooks enabled

Install in Claude Code

claude plugin marketplace add alexanderradahl/shunt-muse
claude plugin install shunt-muse@muse-shunt

Start a new session. The skills show up as `