setup ignores --help and silently performs a destructive non-idempotent reinstall

[lcatlett]

Summary

claude-smart setup does not recognize --help. Instead of printing usage or erroring on an
unknown flag, it falls through to its interactive prompts. When run non-interactively the
prompts immediately hit EOF, the defaults are silently accepted, and a full non-idempotent
reinstall executes — rewriting config, restoring hook manifests over user edits, and
restarting services.

This makes the most destructive command in the tool the one most likely to be invoked by
accident. --help is the flag people reach for precisely because they don't know what a
command does yet.

Impact

A single claude-smart setup --help — from a script, CI, a wrapper, or an AI coding agent
inspecting the CLI — performs the following without any confirmation:

• Rewrites ~/.reflexio/.env (announced as "removed managed Reflexio settings" then
  "configured local Reflexio defaults"). User-set values here are replaced with defaults.
• Rewrites ~/.claude-smart/.env
• Calls _restore_publish_hooks_from_source(), which copies the full hook manifest back over
  the installed one — silently discarding any user customization of which hooks are enabled
• Rewrites ~/.claude-smart/install-complete
• Re-runs the marketplace install, uv sync, and the bundled Reflexio source install
• Restarts the backend and dashboard services

The hook-manifest restore is the most damaging: a user who had pruned hook events finds them
all silently re-enabled, with no message saying so.

Reproduction

claude-smart setup --help

Non-interactively (no tty), e.g. under timeout, a script, or a tool call.

Observed

Host (1=Claude Code, 2=Codex, 3=all, 4=OpenCode) [claude-code]: Setup mode (1=local, 2=managed Reflexio) [local]: [claude-smart setup] removed managed Reflexio settings from ~/.reflexio/.env
[claude-smart setup] configured local Reflexio defaults in ~/.reflexio/.env
Adding marketplace…✔ Marketplace 'reflexioai' already on disk — declared in user settings
Installing plugin "claude-smart@reflexioai"...✔ Plugin "claude-smart@reflexioai" is already installed (scope: user)
[claude-smart] running uv sync...
[claude-smart] installing bundled Reflexio source from ...
[claude-smart] starting dashboard build in background (~1-2 min on first install)
[claude-smart] starting backend service in background
[claude-smart] install complete. Backend started; dashboard auto-starts on session start.
Prepared claude-smart runtime at ~/.claude/plugins/cache/reflexioai/claude-smart/0.2.50.
Started claude-smart backend service.
Refreshed claude-smart dashboard service.

Both prompts print and are answered by EOF in the same line of output — the defaults are taken
with no input.

Expected

Any of:

1. --help prints usage and exits 0 without side effects.
2. An unrecognized flag errors and exits non-zero without side effects.
3. When stdin is not a tty, setup refuses to run rather than accepting defaults — or requires
   an explicit --yes / --non-interactive flag to proceed.

At minimum, EOF on a prompt should abort rather than silently select the default for a
destructive operation.

Relevant source

• src/claude_smart/cli.py — _restore_publish_hooks_from_source() copies the full manifest
  from the package source over the installed one on every setup/install run, before the
  read-only pruning is conditionally applied. Any user edit to the installed manifest is lost.

Additional note on hook configuration

Related, and possibly worth a separate issue: CLAUDE_SMART_READ_ONLY=1 is currently the only
supported way to change which hooks are registered, and _command_is_publish_hook() matches
only:

hook_entry\.sh\b[\s"']+(?:codex|claude-code)[\s"']+(?:stop|session-end)\b

So read-only mode prunes Stop and SessionEnd while leaving UserPromptSubmit,
PreToolUse, and PostToolUse registered. There is no supported way to do the inverse — keep
session storage/analysis while disabling the per-prompt and per-tool-call hooks. On one machine
those three accounted for 14,358 firings against 592 stop firings in ~/.claude-smart/hook.log
(~24:1), and 339 firings in a single session.

Environment

• claude-smart 0.2.50
• macOS, arm64
• Installed via the mise npm backend (npm:claude-smart), marketplace reflexioai