Support local-only slash commands backed by shell scripts

Feature Request

Problem

Skills (custom slash commands) always make a round trip to the Claude API, even when the task requires no LLM reasoning — e.g., displaying a local TODO list, showing bookmarks, or running a simple shell script. This makes simple commands slow and wastes tokens.

Proposed Solution

Allow users to register local-only slash commands that execute a shell script and display the output directly, with no API round trip. For example:

// In settings.json or a skill definition
{
  "local_commands": {
    "todos": {
      "description": "Display my TODO list",
      "command": "cat ~/Notes/TODO.md"
    },
    "bookmarks": {
      "description": "Show saved bookmarks",
      "command": "cat ~/Notes/reference/bookmarks.md"
    }
  }
}

Or alternatively, a flag in the skill SKILL.md frontmatter:

---
description: Display my TODO list
user_invocable: true
local: true
command: "cat ~/Notes/TODO.md"
---

Use Cases

• Displaying local files (TODO lists, bookmarks, notes)
• Running project scripts (test suites, build status, linting)
• Quick lookups that don't need AI (git log summaries, env checks)
• Any command where the output just needs to be printed, not interpreted

Current Workaround

Users can use ! command to run shell commands, or define shell aliases — but these lose the discoverability and ergonomics of slash commands, and the output isn't part of the conversation context.

Additional Context

The existing split between built-in commands (local) and skills (API) is clean, but there's a gap for user-defined commands that are local-only. This would make Claude Code feel faster and more extensible for power users who build personal workflows around it.