Set Up Agent Guard Shell

Install or refresh Agent Guard shell command masking without making the user
locate the versioned plugin-cache binary.

Workflow

1. Explain that this changes the user's shell rc and is separate from plugin
   hook setup.

2. Resolve ../../bin/agent-guard relative to this skill directory and confirm
   that it is executable. Use that plugin-local binary even if another
   agent-guard is on PATH; do not install or update a standalone copy.

3. Tell the user that setup-shell will update the managed block in their shell
   rc with command wrapping enabled by default. Obtain host approval for the
   home-directory write before running:

   "<plugin-local-agent-guard>" setup-shell

4. On success, report every rc path from the command output and tell the user to
   restart the shell and any agent sessions launched from that shell. A fish
   login shell gets two rc files (the integration is POSIX shell code that fish
   cannot load, and agent shell snapshots run bash or zsh); relay the fish notice
   from the output as-is, including that no automatic agx is installed at a
   fish prompt, a bare agent-guard may be unavailable for plugin-only installs,
   and the command prints an executable path that works without PATH.

5. If the approved write is blocked by the host sandbox, relay the exact error.
   Do not retry the same blocked write, choose another rc, or edit the rc by
   hand. Show the exact plugin-local command for the user to run directly in a
   terminal, wait for confirmation, and then report the required restarts.

Use setup-shell --no-command-wrapping only when the user explicitly requests
the persistent opt-out. AGENT_GUARD_COMMAND_WRAPPING=off is the runtime-only
opt-out.