onboard

name	onboard
description	First-run setup for style-switcher. Creates the user-level data directory and writes config.json. Default apply target is a delimited fragment inside ~/.claude/CLAUDE.md, since the plugin layers multiple snippets together rather than swapping single output-styles.

Onboard

Run on first use, or when the user wants to reconfigure.

Data location

DATA_ROOT="${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/style-switcher"
mkdir -p "$DATA_ROOT/snippets" "$DATA_ROOT/recipes" "$DATA_ROOT/state" "$DATA_ROOT/backups"

Config goes in $DATA_ROOT/config.json. Never write config to ~/.claude/.

Steps

1. Resolve $DATA_ROOT as above. Create snippets/, state/, backups/.

2. Detect existing config. If $DATA_ROOT/config.json exists, ask whether to (a) edit a single field, (b) reset entirely (back up to config.json.bak-<timestamp> first), or (c) abort.

3. Pick apply target. Three modes — pick one and store it in config:

   • claude-md-fragment (default) — manages a <!-- style-switcher:start --> … <!-- style-switcher:end --> block inside ~/.claude/CLAUDE.md. Layered on top of the user's normal config.
   • custom-path — user supplies an absolute path; same delimited-block protocol.
   • repo-sandbox — playground mode. Renames ~/.claude/CLAUDE.md → ~/.claude/CLAUDE_HELD.md to take it out of the harness, then writes the full recipe content as <cwd>/CLAUDE.md (not a managed block — the entire file is the persona). This is the safest way to test a persona in isolation without your serious config bleeding through. /go-away reverses both moves.

   Note on repo-sandbox: while a recipe is active in this mode, every Claude Code session on the machine sees the held state (no user-level CLAUDE.md is loaded). If the user mostly runs one Claude Code session at a time this is fine — if they run several concurrently, warn them.

4. Write config. See shape below.

5. Confirm. Print $DATA_ROOT, the apply target path, and remind the user the snippet library is empty — point them at add-snippet.

The plugin ships with no built-in snippets and no fixed category list. The user defines both as they go.

config.json shape

{
  "data_root": "/home/<user>/.local/share/claude-plugins/style-switcher",
  "apply_target": {
    "type": "claude-md-fragment",
    "path": "/home/<user>/.claude/CLAUDE.md"
  },
  // or, for repo-sandbox mode:
  // "apply_target": { "type": "repo-sandbox" },
  "active_layers": {},
  "active_recipe": null,
  "applied_at": null,
  "configured_at": "2026-04-29T00:00:00Z"
}

active_layers maps category → snippet-name. active_recipe holds the id of the currently-applied recipe (or null when layers were composed manually via apply-layers). Both are empty/null until the user applies something.

Layout

$DATA_ROOT/snippets/
├── <category-1>/
│   ├── <snippet-a>.md
│   └── <snippet-b>.md
├── <category-2>/
│   └── ...

A "category" is just a subdirectory. Suggested dimensions: verbosity, initiative, tone, formatting, hedging, emoji. The user defines whichever they want.

Composition rule

Within a single category, at most one snippet can be active. Across categories, snippets compose — they all get written into the delimited block in target file order.