setup - JB Kaizen, configured by the model instead of the terminal wizard

install.sh normally asks seven questions itself when run interactively ([ -t 0 ]). A Bash call made by Claude from inside a skill has no TTY: forcing the interactive wizard (--wizard) would make every question read EOF immediately, ask_roots would break out of its retry loop at the first "not found" ([ -t 0 ] || break), and the user would silently get defaults with no questions asked at all - the opposite of "one click, the wizard configures everything".

So this skill never shells into the interactive wizard. You ask the eight questions yourself with AskUserQuestion, then run install.sh exactly once, non-interactively, with every answer passed as a flag.

1. Ask the eight questions

One AskUserQuestion call, all eight questions together, defaults pre-selected. Ask in the language the user is writing in. The eight, with the meaning to show the user for each (mirrors extension/README.md's step table):

The install also adds the four short aliases /kz-setup, /kz-learn, /kz-audit, /kz-balance to ~/.claude/skills/. If the user says they do not want them, append --no-aliases to the install command below.

Pre-select sensible defaults (dependencies: yes if any missing; roots: ~/workspace if it exists and holds git repos; repositories: all studied; routing hook: on; nightly: off; private remote: none; VS Code: off; Hoply: skip) so a user who just wants it working can accept every default in one submission.

2. Resolve the plugin root

Same base every other plugin-mode path uses - CLAUDE_CONFIG_DIR, never a bare $HOME/.claude:

C="${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
ROOT="${CLAUDE_PLUGIN_ROOT:-$(ls -td "$C"/plugins/cache/*/kaizen/*/ 2>/dev/null | head -1)}"
[ -f "$ROOT/install.sh" ] || {
  echo "cannot find the kaizen plugin root - reinstall with /plugin install kaizen@jb-kaizen"
  exit 1; }

CLAUDE_PLUGIN_ROOT is set by the plugin runtime itself whenever this skill is actually invoked as a plugin skill; the ls -td ... | head -1 fallback covers the rare case where it is unset (e.g. this skill run standalone during development) by taking the newest cache checkout. The [ -f ... ] guard also catches an empty ROOT - without it a missing match would silently become bash "/install.sh".

3. Run exactly one non-interactive command

Build the flags from the answers and run once, output redirected to a log:

bash "$ROOT/install.sh" --plugin --yes ${NO_ALIASES} --workspace <roots> \
  [--schedule HH:MM | --no-schedule] [--store-remote <url>] [--vscode all|none|claude-launcher,graphify-watch] \
  [--hook|--no-hook] \
  > /tmp/kaizen-setup.log 2>&1
tail -20 /tmp/kaizen-setup.log

• <roots> is the comma-separated list from question 2.
• --schedule HH:MM when question 5 was yes, --no-schedule when it was no.
• --store-remote <url> only when question 6 gave a URL; omit it entirely when the user skipped or cleared it.
• --vscode takes the companion list from question 7: all, none, or a comma-separated subset of claude-launcher,graphify-watch. install.sh matches those tokens only - any other word (yes) means none.
• ${NO_ALIASES} is empty by default; set it to --no-aliases only when the user said no to the short aliases.
• Question 4 maps to --hook / --no-hook (in --plugin mode the hook is not installed either way, but the answer is what gets persisted to config.json). Question 3's repository toggles have no flag - they are saved as disabledRepos in config.json on the normal wizard_save_config path.
• Never pass --wizard from inside this skill - see the note above.
• Question 8 maps to --hoply-reminder (yes) or --no-hoply-reminder (skip).

3b. Hoply: sign-in in the browser, never a token in the chat

When question 8 was yes, connect it through the installer (no secret involved, safe to run from here: inside a Claude Code session it never starts an interactive sign-in, it only adds the server, records the connection in config.json and prints the sign-in line):

bash "$ROOT/install-entry.sh" --hoply > /tmp/kaizen-hoply.log 2>&1
tail -8 /tmp/kaizen-hoply.log

Then tell the user to sign in with their Hoply account: in Claude Code run /mcp, pick hoply, choose Authenticate (or claude mcp login hoply in their own terminal). Never ask for a token or a password in AskUserQuestion or in chat: anything typed in a session is stored in its transcript. A personal token is only for CI and headless machines (HOPLY_TOKEN=... npx -y jb-kaizen --yes).

4. Print the terminal fallback

Always mention the escape hatch for anyone who wants the real interactive TUI instead of answering through the model:

bash "$ROOT/install.sh" --plugin --wizard

pasted into their own terminal (has a TTY, so the seven questions work normally there).

5. What --plugin skips, and why

--plugin skips copying skills and the routing hook into ~/.claude/skills / ~/.claude/hooks - the plugin itself already delivers those (they load straight from CLAUDE_PLUGIN_ROOT, no copy needed). Everything else - dependencies, workspace roots, repository registry, nightly schedule, private-store remote, VS Code companions - is installed and saved exactly as it would be from a repo checkout.

Re-running this skill (or install.sh --plugin --yes) at any time is safe and idempotent: saved answers in config.json fill in anything not explicitly re-asked, flags always win over the saved config, and nothing destructive happens on a repeat run.

6. Where the answers live

Every answer is saved to ~/.claude/self-learn/config.json (path resolved through CLAUDE_CONFIG_DIR the same way as everywhere else). To change one thing later - the nightly time, the workspace roots, the private remote - re-run /kaizen:setup and answer only that question differently; every other answer is carried forward from the saved config.