Sando — AI skill for Claude Code

Sando

Sando is a local plugin for Claude Code and Codex. It keeps repeated and oversized tool output under control, preserves complete redacted artifacts when they fit the admission limit, and provides bounded local tool surfaces. It runs locally and makes no LLM calls.

Release notes: [Sando 0.4.2](docs/changelogs/0.4.2.md).

Install the plugin

Install Sando from the host marketplace. Each marketplace source includes its hooks and bundles, so installation needs no build step.

Requirements

• Claude Code or Codex
• Node.js >=22.22.0 <23, with node available in PATH

Check the Node version with:

node --version

Claude Code

Run these commands in Claude Code:

/plugin marketplace add yuzushi-dev/yuzushi-plugins
/plugin install sando@yuzushi

Codex

Add the marketplace:

codex plugin marketplace add yuzushi-dev/yuzushi-plugins

Then open `/plugins`, select `sando`, and install/enable it. Start a new session if Codex was already open.

Each host wires only the hooks and MCP surfaces declared by its manifest. The remaining bundled CLI, audit, artifact recovery, proxy, statusline, metrics, and accounting launchers are manual entrypoints. The bundle is self-contained; the optional npm package is not required.

Provider accounting and paired controls

Sando records provider-reported input, cache-read, cache-write, output, reasoning, and turn counts at session stop. It also records mechanical context trimming separately. A weighted token estimate is diagnostic; provider cost and blended rates are shown only when the provider or host reports them.

Run an explicit control session with the plugin still installed:

SANDO_EXPERIMENT=read-heavy \
SANDO_EXPERIMENT_ARM=control \
codex

Treatment sessions use `SANDO_EXPERIMENT_ARM=apply` (the default). Use the same experiment and optional `SANDO_EXPERIMENT_WORKLOAD` fo